Datasets:

GEM

/

SciDuet

like 2

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119 ], "paper_content_text": [ "Introduction Cognates are words having similarities in their spelling and meaning in two languages, either because the two languages are typologically related, e.g., maladie vs malattia ('disease'), or because they were both borrowed from the same source (informatique vs informatica).", "The advantage of their use in Statistical Machine Translation (SMT) is that the procedure can be based on comparable corpora, i.e., similar corpora which are not translations of each other (Sharoff et al., 2013) .", "Given that there are more sources of comparable corpora in comparison to parallel ones, the lexicon obtained from them is likely to be richer and more variable.", "Detection of cognates is a well-known task, which has been explored for a range of languages using different methods.", "The two main approaches applied to detection of the cognates are the generative and discriminative paradigms.", "The first one is based on detection of the edit distance between potential candidate pairs.", "The distance can be a simple Levenshtein distance, or a distance measure with the scores learned from an existing parallel set (Tiedemann, 1999; Mann and Yarowsky, 2001) .", "The discriminative paradigm uses standard approaches to machine learning, which are based on (1) extracting features, e.g., character n-grams, and (2) learning to predict the transformations of the source word needed to (Jiampojamarn et al., 2010; Frunza and Inkpen, 2009) .", "Given that SMT is usually based on a full-form lexicon, one of the possible issues in generation of cognates concerns the similarity of words in their root form vs the similarity in endings.", "For example, the Ukrainian wordform áëèaeíüîãî 'near gen ' is cognate to Russian áëèaeíåãî, the root is identical, while the ending is considerably different (üîãî vs åãî).", "Regular differences in the endings, which are shared across a large number of words, can be learned separately from the regular differences in the roots.", "One also needs to take into account the false friends among cognates.", "For example, diseñar means 'to design' in Spanish vs desenhar in Portuguese means 'to draw'.", "There are also often cases of partial cognates, when the words share the meaning in some contexts, but not in others, e.g., aeåíà in Russian means 'wife', while its Bulgarian cognate aeåíà has two meanings: 'wife' and 'woman'.", "Yet another complexity concerns a frequency mismatch.", "Two cognates might differ in their frequency.", "For example, dibujo in Spanish ('a drawing', rank 1779 in the Wikipedia frequency list) corresponds to a relatively rare cognate word debuxo in Portuguese (rank 104,514 in Wikipedia), while another Portuguese word desenho is more commonly used in this sense (rank 884 in the Portuguese Wikipedia).", "For MT tasks we need translations that are equally appropriate in the source and target language, therefore cognates useful for a high-quality dictionary for SMT need to have roughly the same frequency in comparable corpora and they need to be used in similar contexts.", "This study investigates the settings for extracting cognates for related languages in Romance and Slavonic language families for the task of reducing the number of unknown words for SMT.", "This in-cludes the effects of having constraints for the cognates to be similar in their roots and in the endings, to occur in distributionally similar contexts and to have similar frequency.", "Methodology The methodology for producing the list of cognates is based on the following steps: 1) Produce several lists of cognates using a family of distance measures, discussed in Section 2.1 from comparable corpora, 2) Prune the candidate lists by ranking items, this is done using a Machine Learning (ML) algorithm trained over parallel corpora for detecting the outliers, discussed in Section 2.2; The initial frequency lists for alignment are based Wikipedia dumps for the following languages: Romance (French, Italian, Spanish, Portuguese) and Slavonic (Bulgarian, Russian, Ukrainian), where the target languages are Spanish and Russian 1 .", "Cognate detection We extract possible lists of cognates from comparable corpora by using a family of similarity measures: L direct matching between the languages using Levenshtein distance (Levenshtein, 1966) ; L(w s , w t ) = 1 − ed(w s , w t ) L-R Levenshtein distance with weights computed separately for the roots and for the endings; LR(r s , r t , e s , e t ) = α×ed(rs,rt)+β×ed(es,et) α+β L-C Levenshtein distance over word with similar number of starting characters (i.e.", "prefix); LC(c s , c t ) = 1 − ed(c s , c t ), same prefix 0, otherwise where ed(., .)", "is the normalised Levenshtein distance in characters between the source word w s and the target word w t .", "The r s and r t are the stems produced by the Snowball stemmer 2 .", "Since the Snowball stemmer does not support Ukrainian and Bulgarian, we used the Russian model for making the stem/ending split.", "e s , e t are the characters at the end of a word form given a stem and c s , c t are the first n characters of a word.", "In this work, we set the weights α = 0.6 and β = 0.4 giving more importance to the roots.", "We set a higher weight to roots on the L-R, which is language dependent, and compare to the L-C metric, which is language independent.", "We transform the Levenshtein distances into similarity metrics by subtracting the normalised distance score from one.", "The produced lists contain for each source word the possible n-best target words accordingly to the maximum scores with one of the previous measures.", "The n-best list allows possible cognate translations to a given source word that share a part of the surface form.", "Different from (Mann and Yarowsky, 2001) , we produce n-best cognate lists scored by edit distance instead of 1-best.", "An important problem when comparing comparable corpora is the way of representing the search space, where an exhaustive method compares all the combinations of source and target words (Mann and Yarowsky, 2001) .", "We constraint the search space by comparing each source word against the target words that belong to a frequency window around the frequency of the source word.", "This constraint only applies for the L and L-R metrics.", "We use Wikipedia dumps for the source and target side processed in the form frequency lists.", "We order the target side list into bins of similar frequency and for the source side we filter words that appear only once.", "We use the window approach given that the frequency between the corpora under study can not be directly comparable.", "During testing we use a wide window of ±200 bins to minimise the loss of good candidate translations.", "The second search space constraint heuristic is the L-C metric.", "This metric only compares source words with the target words upto a given n prefix.", "For c s , c t in L-C , we use the first four characters to compare groups of words as suggested in (Kondrak et al., 2003) .", "Cognate Ranking Given that the n-best lists contain noise, we aim to prune them by an ML ranking model.", "However, there is a lack of resources to train a classification model for cognates (i.e.", "cognate vs. false friend), as mentioned in (Fišer and Ljubešić, 2013) .", "Available data that can be used to judge the cognate lists are the alignment pairs extracted from parallel data.", "We decide to use a ranking model to avoid data imbalance present in classification and to use the probability scores of the alignment pairs as ranks, as opposed to the classification model used by (Irvine and Callison-Burch, 2013) .", "Moreover, we also use a popular domain adaptation technique (Daumé et al., 2010) given that we have access to different domains of parallel training data that might be compatible with our comparable corpora.", "The training data are the alignments between pairs of words where we rank them accordingly to their correspondent alignment probability from the output of GIZA++ (Och and Ney, 2003) .", "We then use a heuristic to prune training data in order to simulate cognate words.", "Pairs of words scored below the Levenshtein similarity threshold of 0.5 are not considered as cognates given that they are likely to have a different surface form.", "We represent the training and test data with features extracted from different edit distance scores and distributional measures.", "The edit distances features are as follows: 1) Similarity measure L and 2) Number of times of each edit operation.", "Thus, the model assigns a different importance to each operation.", "The distributional feature is based on the cosine between the distributional vectors of a window of n words around the word currently under comparison.", "We train distributional similarity models with word2vec (Mikolov et al., 2013a) for the source and target side separately.", "We extract the continuous vector for each word in the window, concatenate it and then compute the cosine between the concatenated vectors of the source and the target.", "We suspect that the vectors will have similar behaviour between the source and the target given that they are trained under parallel Wikipedia articles.", "We develop two ML models: 1) Edit distance scores and 2) Edit distance scores and distributional similarity score.", "We use SVMlight (Joachims, 1998) Results and Discussion In this section we describe the data used to produce the n-best lists and train the cognate ranking models.", "We evaluate the ranking models with heldout data from each training domain.", "We also provide manual evaluation over the ranked n-best lists for error analysis.", "Data The n-best lists to detect cognates were extracted from the respective Wikipedias by using the method described in Section 2.1.", "The training data for the ranking model consists of different types of parallel corpora.", "The parallel corpora are as follows: 1) Wiki-titles we use the inter language links to create a parallel corpus from the tittles of the Wikipedia articles, with about 500K aligned links (i.e.", "'sentences') per language pair (about 200k for bg-ru), giving us about 200K training instances per language pair 3 , 2) Opensubs is an open source corpus of subtitles built by the fan community, with 1M sentences, 6M tokens, 100K words, giving about 90K training instances (Tiedemann, 2012) and 3) Zoo is a proprietary corpus of subtitles produced by professional translators, with 100K sentences, 700K tokens, 40K words and giving about 20K training instances per language pair.", "Our objective is to create MT dictionaries from the produced n-best lists and we use parallel data as a source of training to prune them.", "We are interested in the corpora of subtitles because the chosen domain of our SMT experiments is subtitling, while the proposed ranking method can be used in other application domains as well.", "We consider Zoo and Opensubs as two different domains given that they were built by different types of translators and they differ in size and quality.", "The heldout data consists of 2K instances for each corpus.", "We use Wikipedia documents and Opensusbs subtitles to train word2vec for the distributional similarity feature.", "We use the continuous bag-ofwords algorithm for word2vec and set the parameters for training to 200 dimensions and a window of 8 words.", "The Wikipedia documents with an average number of 70K documents for each language, and Opensubs subtitles with 1M sentences for each language.", "In practice we only use the Wikipedia data given that for Opensubs the model is able to find relatively few vectors, for example a vector is found only for 20% of the words in the pt-es pair.", "Evaluation of the Ranking Model We define two ranking models as: model E for edit distance features and model EC for both edit Table 1 shows the results of the ranking procedure.", "For the Romance family language pairs the model EC with context features consistently reduces the error compared to the solely use of edit distance metrics.", "The only exception is the it-es EC model with poor results for the domain of Wiki-titles.", "The models for the Slavonic family behave similarly to the Romance family, where the use of context features reduces the ranking error.", "The exception is the bg-ru model on the Opensubs domain.", "A possible reason for the poor results on the ites and bg-ru models is that the model often assigns a high similarity score to unrelated words.", "For example, in it-es, mortes 'deads' is treated as close to categoria 'category'.", "A possible solution is to map the vectors form the source side into the space of the target side via a learned transformation matrix (Mikolov et al., 2013b) .", "Preliminary Results on Comparable Corpora After we extracted the n-best lists for the Romance family comparable corpora, we applied one of the ranking models on these lists and we manually evaluated over a sample of 50 words 4 .", "We set n to 10 for the n-best lists.", "We use a frequency window of 200 for the n-best list search heuristic and the domain of the comparable corpora to Wiki-titles 4 The sample consists of words with a frequency between 2K and 5. for the domain adaptation technique.", "The purpose of manual evaluation is to decide whether the ML setup is sensible on the objective task.", "Each list is evaluated by accuracy at 1 and accuracy at 10.", "We also show success and failure examples of the ranking and the n-best lists.", "Table 2 shows the preliminary results of the ML model E on a sample of Wikipedia dumps.", "The L and L-R lists consistently show poor results.", "A possible reason is the amount of errors given the first step to extract the n-best lists.", "For example, in pt-es, for the word vivem 'live' the 10-best list only contain one word with a similar meaning viva 'living' but it can be also translated as 'cheers'.", "In the pt-es list for the word representação 'description' the correct translation representación is not among the 10-best in the L list.", "However, it is present in the 10-best for the L-C list and the ML model EC ranks it in the first place.", "The edit distance model E still makes mistakes like with the list L-C, the word vivem 'live' translates into viven 'living' and the correct translation is vivir.", "However, given a certain context/sense the previous translation can be correct.", "The ranking scores given by the SVM varies from each list version.", "For the L-C lists the scores are more uniform in increasing order and with a small variance.", "The L and L-R lists show the opposite behaviour.", "We add the produced Wikipedia n-best lists with the L metric into a SMT training dataset for the ptes pair.", "We use the Moses SMT toolkit (Koehn et al., 2007) to test the augmented datasets.", "We compare the augmented model with a baseline both trained by using the Zoo corpus of subtitles.", "We use a 1-best list consisting of 100K pairs.", "Te dataset used for pt-es baseline is: 80K training sentences, 1K sentences for tuning and 2K sen- Lang Pairs acc@1 acc@10 acc@1 acc@10 acc@1 acc@10 pt-es 20 60 22 59 32 70 it-es 16 53 18 45 44 66 fr-es 10 48 12 51 29 59 A possible reason for low improvement in terms of the BLEU scores is because MT evaluation metrics, such as BLEU, compare the MT output with a human reference.", "The human reference translations in our corpus have been done from English (e.g., En→Es), while the test translations come from a related language (En→Pt→Es), often resulting in different paraphrases of the same English source.", "While our OOV rate improved, the evaluation scores did not reflected this, because our MT output was still far from the reference even in cases it was otherwise acceptable.", "List L List L-R List L-C Conclusions and future Work We have presented work in progress for developing MT dictionaries extracted from comparable resources for related languages.", "The extraction heuristic show positive results on the n-best lists that group words with the same starting char-5 https://github.com/clab/fast_align 6 https://kheafield.com/code/kenlm/ 7 The p-value for the uk-ru pair is 0.06 we do not consider this result as statistically significant.", "acters, because the used comparable corpora consist of related languages that share a similar orthography.", "However, the lists based on the frequency window heuristic show poor results to include the correct translations during the extraction step.", "Our ML models based on similarity metrics over parallel corpora show rankings similar to heldout data.", "However, we created our training data using simple heuristics that simulate cognate words (i.e.", "pairs of words with a small surface difference).", "The ML models are able to rank similar words on the top of the list and they give a reliable score to discriminate wrong translations.", "Preliminary results on the addition of the n-best lists into an SMT system show modest improvements compare to the baseline.", "However, the OOV rate shows improvements around 10% reduction on word types, because of the wide variety of lexical choices introduced by the MT dictionaries.", "Future work involves the addition of unsupervised morphology features for the n-best list extraction, i.e.", "first step, given that the use of starting characters shows to be an effective heuristic to prune the search space and language independent.", "Finally, we will measure the contribution for all the produced cognate lists, where we can try different strategies to add the dictionaries into an SMT system (Irvine and Callison-Burch, 2014) ." ] }

{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", "Preliminary Results on Comparable Corpora", "Conclusions and future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119 ], "paper_content_text": [ "Introduction Cognates are words having similarities in their spelling and meaning in two languages, either because the two languages are typologically related, e.g., maladie vs malattia ('disease'), or because they were both borrowed from the same source (informatique vs informatica).", "The advantage of their use in Statistical Machine Translation (SMT) is that the procedure can be based on comparable corpora, i.e., similar corpora which are not translations of each other (Sharoff et al., 2013) .", "Given that there are more sources of comparable corpora in comparison to parallel ones, the lexicon obtained from them is likely to be richer and more variable.", "Detection of cognates is a well-known task, which has been explored for a range of languages using different methods.", "The two main approaches applied to detection of the cognates are the generative and discriminative paradigms.", "The first one is based on detection of the edit distance between potential candidate pairs.", "The distance can be a simple Levenshtein distance, or a distance measure with the scores learned from an existing parallel set (Tiedemann, 1999; Mann and Yarowsky, 2001) .", "The discriminative paradigm uses standard approaches to machine learning, which are based on (1) extracting features, e.g., character n-grams, and (2) learning to predict the transformations of the source word needed to (Jiampojamarn et al., 2010; Frunza and Inkpen, 2009) .", "Given that SMT is usually based on a full-form lexicon, one of the possible issues in generation of cognates concerns the similarity of words in their root form vs the similarity in endings.", "For example, the Ukrainian wordform áëèaeíüîãî 'near gen ' is cognate to Russian áëèaeíåãî, the root is identical, while the ending is considerably different (üîãî vs åãî).", "Regular differences in the endings, which are shared across a large number of words, can be learned separately from the regular differences in the roots.", "One also needs to take into account the false friends among cognates.", "For example, diseñar means 'to design' in Spanish vs desenhar in Portuguese means 'to draw'.", "There are also often cases of partial cognates, when the words share the meaning in some contexts, but not in others, e.g., aeåíà in Russian means 'wife', while its Bulgarian cognate aeåíà has two meanings: 'wife' and 'woman'.", "Yet another complexity concerns a frequency mismatch.", "Two cognates might differ in their frequency.", "For example, dibujo in Spanish ('a drawing', rank 1779 in the Wikipedia frequency list) corresponds to a relatively rare cognate word debuxo in Portuguese (rank 104,514 in Wikipedia), while another Portuguese word desenho is more commonly used in this sense (rank 884 in the Portuguese Wikipedia).", "For MT tasks we need translations that are equally appropriate in the source and target language, therefore cognates useful for a high-quality dictionary for SMT need to have roughly the same frequency in comparable corpora and they need to be used in similar contexts.", "This study investigates the settings for extracting cognates for related languages in Romance and Slavonic language families for the task of reducing the number of unknown words for SMT.", "This in-cludes the effects of having constraints for the cognates to be similar in their roots and in the endings, to occur in distributionally similar contexts and to have similar frequency.", "Methodology The methodology for producing the list of cognates is based on the following steps: 1) Produce several lists of cognates using a family of distance measures, discussed in Section 2.1 from comparable corpora, 2) Prune the candidate lists by ranking items, this is done using a Machine Learning (ML) algorithm trained over parallel corpora for detecting the outliers, discussed in Section 2.2; The initial frequency lists for alignment are based Wikipedia dumps for the following languages: Romance (French, Italian, Spanish, Portuguese) and Slavonic (Bulgarian, Russian, Ukrainian), where the target languages are Spanish and Russian 1 .", "Cognate detection We extract possible lists of cognates from comparable corpora by using a family of similarity measures: L direct matching between the languages using Levenshtein distance (Levenshtein, 1966) ; L(w s , w t ) = 1 − ed(w s , w t ) L-R Levenshtein distance with weights computed separately for the roots and for the endings; LR(r s , r t , e s , e t ) = α×ed(rs,rt)+β×ed(es,et) α+β L-C Levenshtein distance over word with similar number of starting characters (i.e.", "prefix); LC(c s , c t ) = 1 − ed(c s , c t ), same prefix 0, otherwise where ed(., .)", "is the normalised Levenshtein distance in characters between the source word w s and the target word w t .", "The r s and r t are the stems produced by the Snowball stemmer 2 .", "Since the Snowball stemmer does not support Ukrainian and Bulgarian, we used the Russian model for making the stem/ending split.", "e s , e t are the characters at the end of a word form given a stem and c s , c t are the first n characters of a word.", "In this work, we set the weights α = 0.6 and β = 0.4 giving more importance to the roots.", "We set a higher weight to roots on the L-R, which is language dependent, and compare to the L-C metric, which is language independent.", "We transform the Levenshtein distances into similarity metrics by subtracting the normalised distance score from one.", "The produced lists contain for each source word the possible n-best target words accordingly to the maximum scores with one of the previous measures.", "The n-best list allows possible cognate translations to a given source word that share a part of the surface form.", "Different from (Mann and Yarowsky, 2001) , we produce n-best cognate lists scored by edit distance instead of 1-best.", "An important problem when comparing comparable corpora is the way of representing the search space, where an exhaustive method compares all the combinations of source and target words (Mann and Yarowsky, 2001) .", "We constraint the search space by comparing each source word against the target words that belong to a frequency window around the frequency of the source word.", "This constraint only applies for the L and L-R metrics.", "We use Wikipedia dumps for the source and target side processed in the form frequency lists.", "We order the target side list into bins of similar frequency and for the source side we filter words that appear only once.", "We use the window approach given that the frequency between the corpora under study can not be directly comparable.", "During testing we use a wide window of ±200 bins to minimise the loss of good candidate translations.", "The second search space constraint heuristic is the L-C metric.", "This metric only compares source words with the target words upto a given n prefix.", "For c s , c t in L-C , we use the first four characters to compare groups of words as suggested in (Kondrak et al., 2003) .", "Cognate Ranking Given that the n-best lists contain noise, we aim to prune them by an ML ranking model.", "However, there is a lack of resources to train a classification model for cognates (i.e.", "cognate vs. false friend), as mentioned in (Fišer and Ljubešić, 2013) .", "Available data that can be used to judge the cognate lists are the alignment pairs extracted from parallel data.", "We decide to use a ranking model to avoid data imbalance present in classification and to use the probability scores of the alignment pairs as ranks, as opposed to the classification model used by (Irvine and Callison-Burch, 2013) .", "Moreover, we also use a popular domain adaptation technique (Daumé et al., 2010) given that we have access to different domains of parallel training data that might be compatible with our comparable corpora.", "The training data are the alignments between pairs of words where we rank them accordingly to their correspondent alignment probability from the output of GIZA++ (Och and Ney, 2003) .", "We then use a heuristic to prune training data in order to simulate cognate words.", "Pairs of words scored below the Levenshtein similarity threshold of 0.5 are not considered as cognates given that they are likely to have a different surface form.", "We represent the training and test data with features extracted from different edit distance scores and distributional measures.", "The edit distances features are as follows: 1) Similarity measure L and 2) Number of times of each edit operation.", "Thus, the model assigns a different importance to each operation.", "The distributional feature is based on the cosine between the distributional vectors of a window of n words around the word currently under comparison.", "We train distributional similarity models with word2vec (Mikolov et al., 2013a) for the source and target side separately.", "We extract the continuous vector for each word in the window, concatenate it and then compute the cosine between the concatenated vectors of the source and the target.", "We suspect that the vectors will have similar behaviour between the source and the target given that they are trained under parallel Wikipedia articles.", "We develop two ML models: 1) Edit distance scores and 2) Edit distance scores and distributional similarity score.", "We use SVMlight (Joachims, 1998) Results and Discussion In this section we describe the data used to produce the n-best lists and train the cognate ranking models.", "We evaluate the ranking models with heldout data from each training domain.", "We also provide manual evaluation over the ranked n-best lists for error analysis.", "Data The n-best lists to detect cognates were extracted from the respective Wikipedias by using the method described in Section 2.1.", "The training data for the ranking model consists of different types of parallel corpora.", "The parallel corpora are as follows: 1) Wiki-titles we use the inter language links to create a parallel corpus from the tittles of the Wikipedia articles, with about 500K aligned links (i.e.", "'sentences') per language pair (about 200k for bg-ru), giving us about 200K training instances per language pair 3 , 2) Opensubs is an open source corpus of subtitles built by the fan community, with 1M sentences, 6M tokens, 100K words, giving about 90K training instances (Tiedemann, 2012) and 3) Zoo is a proprietary corpus of subtitles produced by professional translators, with 100K sentences, 700K tokens, 40K words and giving about 20K training instances per language pair.", "Our objective is to create MT dictionaries from the produced n-best lists and we use parallel data as a source of training to prune them.", "We are interested in the corpora of subtitles because the chosen domain of our SMT experiments is subtitling, while the proposed ranking method can be used in other application domains as well.", "We consider Zoo and Opensubs as two different domains given that they were built by different types of translators and they differ in size and quality.", "The heldout data consists of 2K instances for each corpus.", "We use Wikipedia documents and Opensusbs subtitles to train word2vec for the distributional similarity feature.", "We use the continuous bag-ofwords algorithm for word2vec and set the parameters for training to 200 dimensions and a window of 8 words.", "The Wikipedia documents with an average number of 70K documents for each language, and Opensubs subtitles with 1M sentences for each language.", "In practice we only use the Wikipedia data given that for Opensubs the model is able to find relatively few vectors, for example a vector is found only for 20% of the words in the pt-es pair.", "Evaluation of the Ranking Model We define two ranking models as: model E for edit distance features and model EC for both edit Table 1 shows the results of the ranking procedure.", "For the Romance family language pairs the model EC with context features consistently reduces the error compared to the solely use of edit distance metrics.", "The only exception is the it-es EC model with poor results for the domain of Wiki-titles.", "The models for the Slavonic family behave similarly to the Romance family, where the use of context features reduces the ranking error.", "The exception is the bg-ru model on the Opensubs domain.", "A possible reason for the poor results on the ites and bg-ru models is that the model often assigns a high similarity score to unrelated words.", "For example, in it-es, mortes 'deads' is treated as close to categoria 'category'.", "A possible solution is to map the vectors form the source side into the space of the target side via a learned transformation matrix (Mikolov et al., 2013b) .", "Preliminary Results on Comparable Corpora After we extracted the n-best lists for the Romance family comparable corpora, we applied one of the ranking models on these lists and we manually evaluated over a sample of 50 words 4 .", "We set n to 10 for the n-best lists.", "We use a frequency window of 200 for the n-best list search heuristic and the domain of the comparable corpora to Wiki-titles 4 The sample consists of words with a frequency between 2K and 5. for the domain adaptation technique.", "The purpose of manual evaluation is to decide whether the ML setup is sensible on the objective task.", "Each list is evaluated by accuracy at 1 and accuracy at 10.", "We also show success and failure examples of the ranking and the n-best lists.", "Table 2 shows the preliminary results of the ML model E on a sample of Wikipedia dumps.", "The L and L-R lists consistently show poor results.", "A possible reason is the amount of errors given the first step to extract the n-best lists.", "For example, in pt-es, for the word vivem 'live' the 10-best list only contain one word with a similar meaning viva 'living' but it can be also translated as 'cheers'.", "In the pt-es list for the word representação 'description' the correct translation representación is not among the 10-best in the L list.", "However, it is present in the 10-best for the L-C list and the ML model EC ranks it in the first place.", "The edit distance model E still makes mistakes like with the list L-C, the word vivem 'live' translates into viven 'living' and the correct translation is vivir.", "However, given a certain context/sense the previous translation can be correct.", "The ranking scores given by the SVM varies from each list version.", "For the L-C lists the scores are more uniform in increasing order and with a small variance.", "The L and L-R lists show the opposite behaviour.", "We add the produced Wikipedia n-best lists with the L metric into a SMT training dataset for the ptes pair.", "We use the Moses SMT toolkit (Koehn et al., 2007) to test the augmented datasets.", "We compare the augmented model with a baseline both trained by using the Zoo corpus of subtitles.", "We use a 1-best list consisting of 100K pairs.", "Te dataset used for pt-es baseline is: 80K training sentences, 1K sentences for tuning and 2K sen- Lang Pairs acc@1 acc@10 acc@1 acc@10 acc@1 acc@10 pt-es 20 60 22 59 32 70 it-es 16 53 18 45 44 66 fr-es 10 48 12 51 29 59 A possible reason for low improvement in terms of the BLEU scores is because MT evaluation metrics, such as BLEU, compare the MT output with a human reference.", "The human reference translations in our corpus have been done from English (e.g., En→Es), while the test translations come from a related language (En→Pt→Es), often resulting in different paraphrases of the same English source.", "While our OOV rate improved, the evaluation scores did not reflected this, because our MT output was still far from the reference even in cases it was otherwise acceptable.", "List L List L-R List L-C Conclusions and future Work We have presented work in progress for developing MT dictionaries extracted from comparable resources for related languages.", "The extraction heuristic show positive results on the n-best lists that group words with the same starting char-5 https://github.com/clab/fast_align 6 https://kheafield.com/code/kenlm/ 7 The p-value for the uk-ru pair is 0.06 we do not consider this result as statistically significant.", "acters, because the used comparable corpora consist of related languages that share a similar orthography.", "However, the lists based on the frequency window heuristic show poor results to include the correct translations during the extraction step.", "Our ML models based on similarity metrics over parallel corpora show rankings similar to heldout data.", "However, we created our training data using simple heuristics that simulate cognate words (i.e.", "pairs of words with a small surface difference).", "The ML models are able to rank similar words on the top of the list and they give a reliable score to discriminate wrong translations.", "Preliminary results on the addition of the n-best lists into an SMT system show modest improvements compare to the baseline.", "However, the OOV rate shows improvements around 10% reduction on word types, because of the wide variety of lexical choices introduced by the MT dictionaries.", "Future work involves the addition of unsupervised morphology features for the n-best list extraction, i.e.", "first step, given that the use of starting characters shows to be an effective heuristic to prune the search space and language independent.", "Finally, we will measure the contribution for all the produced cognate lists, where we can try different strategies to add the dictionaries into an SMT system (Irvine and Callison-Burch, 2014) ." ] }

{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", "Preliminary Results on Comparable Corpora", "Conclusions and future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119 ], "paper_content_text": [ "Introduction Cognates are words having similarities in their spelling and meaning in two languages, either because the two languages are typologically related, e.g., maladie vs malattia ('disease'), or because they were both borrowed from the same source (informatique vs informatica).", "The advantage of their use in Statistical Machine Translation (SMT) is that the procedure can be based on comparable corpora, i.e., similar corpora which are not translations of each other (Sharoff et al., 2013) .", "Given that there are more sources of comparable corpora in comparison to parallel ones, the lexicon obtained from them is likely to be richer and more variable.", "Detection of cognates is a well-known task, which has been explored for a range of languages using different methods.", "The two main approaches applied to detection of the cognates are the generative and discriminative paradigms.", "The first one is based on detection of the edit distance between potential candidate pairs.", "The distance can be a simple Levenshtein distance, or a distance measure with the scores learned from an existing parallel set (Tiedemann, 1999; Mann and Yarowsky, 2001) .", "The discriminative paradigm uses standard approaches to machine learning, which are based on (1) extracting features, e.g., character n-grams, and (2) learning to predict the transformations of the source word needed to (Jiampojamarn et al., 2010; Frunza and Inkpen, 2009) .", "Given that SMT is usually based on a full-form lexicon, one of the possible issues in generation of cognates concerns the similarity of words in their root form vs the similarity in endings.", "For example, the Ukrainian wordform áëèaeíüîãî 'near gen ' is cognate to Russian áëèaeíåãî, the root is identical, while the ending is considerably different (üîãî vs åãî).", "Regular differences in the endings, which are shared across a large number of words, can be learned separately from the regular differences in the roots.", "One also needs to take into account the false friends among cognates.", "For example, diseñar means 'to design' in Spanish vs desenhar in Portuguese means 'to draw'.", "There are also often cases of partial cognates, when the words share the meaning in some contexts, but not in others, e.g., aeåíà in Russian means 'wife', while its Bulgarian cognate aeåíà has two meanings: 'wife' and 'woman'.", "Yet another complexity concerns a frequency mismatch.", "Two cognates might differ in their frequency.", "For example, dibujo in Spanish ('a drawing', rank 1779 in the Wikipedia frequency list) corresponds to a relatively rare cognate word debuxo in Portuguese (rank 104,514 in Wikipedia), while another Portuguese word desenho is more commonly used in this sense (rank 884 in the Portuguese Wikipedia).", "For MT tasks we need translations that are equally appropriate in the source and target language, therefore cognates useful for a high-quality dictionary for SMT need to have roughly the same frequency in comparable corpora and they need to be used in similar contexts.", "This study investigates the settings for extracting cognates for related languages in Romance and Slavonic language families for the task of reducing the number of unknown words for SMT.", "This in-cludes the effects of having constraints for the cognates to be similar in their roots and in the endings, to occur in distributionally similar contexts and to have similar frequency.", "Methodology The methodology for producing the list of cognates is based on the following steps: 1) Produce several lists of cognates using a family of distance measures, discussed in Section 2.1 from comparable corpora, 2) Prune the candidate lists by ranking items, this is done using a Machine Learning (ML) algorithm trained over parallel corpora for detecting the outliers, discussed in Section 2.2; The initial frequency lists for alignment are based Wikipedia dumps for the following languages: Romance (French, Italian, Spanish, Portuguese) and Slavonic (Bulgarian, Russian, Ukrainian), where the target languages are Spanish and Russian 1 .", "Cognate detection We extract possible lists of cognates from comparable corpora by using a family of similarity measures: L direct matching between the languages using Levenshtein distance (Levenshtein, 1966) ; L(w s , w t ) = 1 − ed(w s , w t ) L-R Levenshtein distance with weights computed separately for the roots and for the endings; LR(r s , r t , e s , e t ) = α×ed(rs,rt)+β×ed(es,et) α+β L-C Levenshtein distance over word with similar number of starting characters (i.e.", "prefix); LC(c s , c t ) = 1 − ed(c s , c t ), same prefix 0, otherwise where ed(., .)", "is the normalised Levenshtein distance in characters between the source word w s and the target word w t .", "The r s and r t are the stems produced by the Snowball stemmer 2 .", "Since the Snowball stemmer does not support Ukrainian and Bulgarian, we used the Russian model for making the stem/ending split.", "e s , e t are the characters at the end of a word form given a stem and c s , c t are the first n characters of a word.", "In this work, we set the weights α = 0.6 and β = 0.4 giving more importance to the roots.", "We set a higher weight to roots on the L-R, which is language dependent, and compare to the L-C metric, which is language independent.", "We transform the Levenshtein distances into similarity metrics by subtracting the normalised distance score from one.", "The produced lists contain for each source word the possible n-best target words accordingly to the maximum scores with one of the previous measures.", "The n-best list allows possible cognate translations to a given source word that share a part of the surface form.", "Different from (Mann and Yarowsky, 2001) , we produce n-best cognate lists scored by edit distance instead of 1-best.", "An important problem when comparing comparable corpora is the way of representing the search space, where an exhaustive method compares all the combinations of source and target words (Mann and Yarowsky, 2001) .", "We constraint the search space by comparing each source word against the target words that belong to a frequency window around the frequency of the source word.", "This constraint only applies for the L and L-R metrics.", "We use Wikipedia dumps for the source and target side processed in the form frequency lists.", "We order the target side list into bins of similar frequency and for the source side we filter words that appear only once.", "We use the window approach given that the frequency between the corpora under study can not be directly comparable.", "During testing we use a wide window of ±200 bins to minimise the loss of good candidate translations.", "The second search space constraint heuristic is the L-C metric.", "This metric only compares source words with the target words upto a given n prefix.", "For c s , c t in L-C , we use the first four characters to compare groups of words as suggested in (Kondrak et al., 2003) .", "Cognate Ranking Given that the n-best lists contain noise, we aim to prune them by an ML ranking model.", "However, there is a lack of resources to train a classification model for cognates (i.e.", "cognate vs. false friend), as mentioned in (Fišer and Ljubešić, 2013) .", "Available data that can be used to judge the cognate lists are the alignment pairs extracted from parallel data.", "We decide to use a ranking model to avoid data imbalance present in classification and to use the probability scores of the alignment pairs as ranks, as opposed to the classification model used by (Irvine and Callison-Burch, 2013) .", "Moreover, we also use a popular domain adaptation technique (Daumé et al., 2010) given that we have access to different domains of parallel training data that might be compatible with our comparable corpora.", "The training data are the alignments between pairs of words where we rank them accordingly to their correspondent alignment probability from the output of GIZA++ (Och and Ney, 2003) .", "We then use a heuristic to prune training data in order to simulate cognate words.", "Pairs of words scored below the Levenshtein similarity threshold of 0.5 are not considered as cognates given that they are likely to have a different surface form.", "We represent the training and test data with features extracted from different edit distance scores and distributional measures.", "The edit distances features are as follows: 1) Similarity measure L and 2) Number of times of each edit operation.", "Thus, the model assigns a different importance to each operation.", "The distributional feature is based on the cosine between the distributional vectors of a window of n words around the word currently under comparison.", "We train distributional similarity models with word2vec (Mikolov et al., 2013a) for the source and target side separately.", "We extract the continuous vector for each word in the window, concatenate it and then compute the cosine between the concatenated vectors of the source and the target.", "We suspect that the vectors will have similar behaviour between the source and the target given that they are trained under parallel Wikipedia articles.", "We develop two ML models: 1) Edit distance scores and 2) Edit distance scores and distributional similarity score.", "We use SVMlight (Joachims, 1998) Results and Discussion In this section we describe the data used to produce the n-best lists and train the cognate ranking models.", "We evaluate the ranking models with heldout data from each training domain.", "We also provide manual evaluation over the ranked n-best lists for error analysis.", "Data The n-best lists to detect cognates were extracted from the respective Wikipedias by using the method described in Section 2.1.", "The training data for the ranking model consists of different types of parallel corpora.", "The parallel corpora are as follows: 1) Wiki-titles we use the inter language links to create a parallel corpus from the tittles of the Wikipedia articles, with about 500K aligned links (i.e.", "'sentences') per language pair (about 200k for bg-ru), giving us about 200K training instances per language pair 3 , 2) Opensubs is an open source corpus of subtitles built by the fan community, with 1M sentences, 6M tokens, 100K words, giving about 90K training instances (Tiedemann, 2012) and 3) Zoo is a proprietary corpus of subtitles produced by professional translators, with 100K sentences, 700K tokens, 40K words and giving about 20K training instances per language pair.", "Our objective is to create MT dictionaries from the produced n-best lists and we use parallel data as a source of training to prune them.", "We are interested in the corpora of subtitles because the chosen domain of our SMT experiments is subtitling, while the proposed ranking method can be used in other application domains as well.", "We consider Zoo and Opensubs as two different domains given that they were built by different types of translators and they differ in size and quality.", "The heldout data consists of 2K instances for each corpus.", "We use Wikipedia documents and Opensusbs subtitles to train word2vec for the distributional similarity feature.", "We use the continuous bag-ofwords algorithm for word2vec and set the parameters for training to 200 dimensions and a window of 8 words.", "The Wikipedia documents with an average number of 70K documents for each language, and Opensubs subtitles with 1M sentences for each language.", "In practice we only use the Wikipedia data given that for Opensubs the model is able to find relatively few vectors, for example a vector is found only for 20% of the words in the pt-es pair.", "Evaluation of the Ranking Model We define two ranking models as: model E for edit distance features and model EC for both edit Table 1 shows the results of the ranking procedure.", "For the Romance family language pairs the model EC with context features consistently reduces the error compared to the solely use of edit distance metrics.", "The only exception is the it-es EC model with poor results for the domain of Wiki-titles.", "The models for the Slavonic family behave similarly to the Romance family, where the use of context features reduces the ranking error.", "The exception is the bg-ru model on the Opensubs domain.", "A possible reason for the poor results on the ites and bg-ru models is that the model often assigns a high similarity score to unrelated words.", "For example, in it-es, mortes 'deads' is treated as close to categoria 'category'.", "A possible solution is to map the vectors form the source side into the space of the target side via a learned transformation matrix (Mikolov et al., 2013b) .", "Preliminary Results on Comparable Corpora After we extracted the n-best lists for the Romance family comparable corpora, we applied one of the ranking models on these lists and we manually evaluated over a sample of 50 words 4 .", "We set n to 10 for the n-best lists.", "We use a frequency window of 200 for the n-best list search heuristic and the domain of the comparable corpora to Wiki-titles 4 The sample consists of words with a frequency between 2K and 5. for the domain adaptation technique.", "The purpose of manual evaluation is to decide whether the ML setup is sensible on the objective task.", "Each list is evaluated by accuracy at 1 and accuracy at 10.", "We also show success and failure examples of the ranking and the n-best lists.", "Table 2 shows the preliminary results of the ML model E on a sample of Wikipedia dumps.", "The L and L-R lists consistently show poor results.", "A possible reason is the amount of errors given the first step to extract the n-best lists.", "For example, in pt-es, for the word vivem 'live' the 10-best list only contain one word with a similar meaning viva 'living' but it can be also translated as 'cheers'.", "In the pt-es list for the word representação 'description' the correct translation representación is not among the 10-best in the L list.", "However, it is present in the 10-best for the L-C list and the ML model EC ranks it in the first place.", "The edit distance model E still makes mistakes like with the list L-C, the word vivem 'live' translates into viven 'living' and the correct translation is vivir.", "However, given a certain context/sense the previous translation can be correct.", "The ranking scores given by the SVM varies from each list version.", "For the L-C lists the scores are more uniform in increasing order and with a small variance.", "The L and L-R lists show the opposite behaviour.", "We add the produced Wikipedia n-best lists with the L metric into a SMT training dataset for the ptes pair.", "We use the Moses SMT toolkit (Koehn et al., 2007) to test the augmented datasets.", "We compare the augmented model with a baseline both trained by using the Zoo corpus of subtitles.", "We use a 1-best list consisting of 100K pairs.", "Te dataset used for pt-es baseline is: 80K training sentences, 1K sentences for tuning and 2K sen- Lang Pairs acc@1 acc@10 acc@1 acc@10 acc@1 acc@10 pt-es 20 60 22 59 32 70 it-es 16 53 18 45 44 66 fr-es 10 48 12 51 29 59 A possible reason for low improvement in terms of the BLEU scores is because MT evaluation metrics, such as BLEU, compare the MT output with a human reference.", "The human reference translations in our corpus have been done from English (e.g., En→Es), while the test translations come from a related language (En→Pt→Es), often resulting in different paraphrases of the same English source.", "While our OOV rate improved, the evaluation scores did not reflected this, because our MT output was still far from the reference even in cases it was otherwise acceptable.", "List L List L-R List L-C Conclusions and future Work We have presented work in progress for developing MT dictionaries extracted from comparable resources for related languages.", "The extraction heuristic show positive results on the n-best lists that group words with the same starting char-5 https://github.com/clab/fast_align 6 https://kheafield.com/code/kenlm/ 7 The p-value for the uk-ru pair is 0.06 we do not consider this result as statistically significant.", "acters, because the used comparable corpora consist of related languages that share a similar orthography.", "However, the lists based on the frequency window heuristic show poor results to include the correct translations during the extraction step.", "Our ML models based on similarity metrics over parallel corpora show rankings similar to heldout data.", "However, we created our training data using simple heuristics that simulate cognate words (i.e.", "pairs of words with a small surface difference).", "The ML models are able to rank similar words on the top of the list and they give a reliable score to discriminate wrong translations.", "Preliminary results on the addition of the n-best lists into an SMT system show modest improvements compare to the baseline.", "However, the OOV rate shows improvements around 10% reduction on word types, because of the wide variety of lexical choices introduced by the MT dictionaries.", "Future work involves the addition of unsupervised morphology features for the n-best list extraction, i.e.", "first step, given that the use of starting characters shows to be an effective heuristic to prune the search space and language independent.", "Finally, we will measure the contribution for all the produced cognate lists, where we can try different strategies to add the dictionaries into an SMT system (Irvine and Callison-Burch, 2014) ." ] }

{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", "Preliminary Results on Comparable Corpora", "Conclusions and future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119 ], "paper_content_text": [ "Introduction Cognates are words having similarities in their spelling and meaning in two languages, either because the two languages are typologically related, e.g., maladie vs malattia ('disease'), or because they were both borrowed from the same source (informatique vs informatica).", "The advantage of their use in Statistical Machine Translation (SMT) is that the procedure can be based on comparable corpora, i.e., similar corpora which are not translations of each other (Sharoff et al., 2013) .", "Given that there are more sources of comparable corpora in comparison to parallel ones, the lexicon obtained from them is likely to be richer and more variable.", "Detection of cognates is a well-known task, which has been explored for a range of languages using different methods.", "The two main approaches applied to detection of the cognates are the generative and discriminative paradigms.", "The first one is based on detection of the edit distance between potential candidate pairs.", "The distance can be a simple Levenshtein distance, or a distance measure with the scores learned from an existing parallel set (Tiedemann, 1999; Mann and Yarowsky, 2001) .", "The discriminative paradigm uses standard approaches to machine learning, which are based on (1) extracting features, e.g., character n-grams, and (2) learning to predict the transformations of the source word needed to (Jiampojamarn et al., 2010; Frunza and Inkpen, 2009) .", "Given that SMT is usually based on a full-form lexicon, one of the possible issues in generation of cognates concerns the similarity of words in their root form vs the similarity in endings.", "For example, the Ukrainian wordform áëèaeíüîãî 'near gen ' is cognate to Russian áëèaeíåãî, the root is identical, while the ending is considerably different (üîãî vs åãî).", "Regular differences in the endings, which are shared across a large number of words, can be learned separately from the regular differences in the roots.", "One also needs to take into account the false friends among cognates.", "For example, diseñar means 'to design' in Spanish vs desenhar in Portuguese means 'to draw'.", "There are also often cases of partial cognates, when the words share the meaning in some contexts, but not in others, e.g., aeåíà in Russian means 'wife', while its Bulgarian cognate aeåíà has two meanings: 'wife' and 'woman'.", "Yet another complexity concerns a frequency mismatch.", "Two cognates might differ in their frequency.", "For example, dibujo in Spanish ('a drawing', rank 1779 in the Wikipedia frequency list) corresponds to a relatively rare cognate word debuxo in Portuguese (rank 104,514 in Wikipedia), while another Portuguese word desenho is more commonly used in this sense (rank 884 in the Portuguese Wikipedia).", "For MT tasks we need translations that are equally appropriate in the source and target language, therefore cognates useful for a high-quality dictionary for SMT need to have roughly the same frequency in comparable corpora and they need to be used in similar contexts.", "This study investigates the settings for extracting cognates for related languages in Romance and Slavonic language families for the task of reducing the number of unknown words for SMT.", "This in-cludes the effects of having constraints for the cognates to be similar in their roots and in the endings, to occur in distributionally similar contexts and to have similar frequency.", "Methodology The methodology for producing the list of cognates is based on the following steps: 1) Produce several lists of cognates using a family of distance measures, discussed in Section 2.1 from comparable corpora, 2) Prune the candidate lists by ranking items, this is done using a Machine Learning (ML) algorithm trained over parallel corpora for detecting the outliers, discussed in Section 2.2; The initial frequency lists for alignment are based Wikipedia dumps for the following languages: Romance (French, Italian, Spanish, Portuguese) and Slavonic (Bulgarian, Russian, Ukrainian), where the target languages are Spanish and Russian 1 .", "Cognate detection We extract possible lists of cognates from comparable corpora by using a family of similarity measures: L direct matching between the languages using Levenshtein distance (Levenshtein, 1966) ; L(w s , w t ) = 1 − ed(w s , w t ) L-R Levenshtein distance with weights computed separately for the roots and for the endings; LR(r s , r t , e s , e t ) = α×ed(rs,rt)+β×ed(es,et) α+β L-C Levenshtein distance over word with similar number of starting characters (i.e.", "prefix); LC(c s , c t ) = 1 − ed(c s , c t ), same prefix 0, otherwise where ed(., .)", "is the normalised Levenshtein distance in characters between the source word w s and the target word w t .", "The r s and r t are the stems produced by the Snowball stemmer 2 .", "Since the Snowball stemmer does not support Ukrainian and Bulgarian, we used the Russian model for making the stem/ending split.", "e s , e t are the characters at the end of a word form given a stem and c s , c t are the first n characters of a word.", "In this work, we set the weights α = 0.6 and β = 0.4 giving more importance to the roots.", "We set a higher weight to roots on the L-R, which is language dependent, and compare to the L-C metric, which is language independent.", "We transform the Levenshtein distances into similarity metrics by subtracting the normalised distance score from one.", "The produced lists contain for each source word the possible n-best target words accordingly to the maximum scores with one of the previous measures.", "The n-best list allows possible cognate translations to a given source word that share a part of the surface form.", "Different from (Mann and Yarowsky, 2001) , we produce n-best cognate lists scored by edit distance instead of 1-best.", "An important problem when comparing comparable corpora is the way of representing the search space, where an exhaustive method compares all the combinations of source and target words (Mann and Yarowsky, 2001) .", "We constraint the search space by comparing each source word against the target words that belong to a frequency window around the frequency of the source word.", "This constraint only applies for the L and L-R metrics.", "We use Wikipedia dumps for the source and target side processed in the form frequency lists.", "We order the target side list into bins of similar frequency and for the source side we filter words that appear only once.", "We use the window approach given that the frequency between the corpora under study can not be directly comparable.", "During testing we use a wide window of ±200 bins to minimise the loss of good candidate translations.", "The second search space constraint heuristic is the L-C metric.", "This metric only compares source words with the target words upto a given n prefix.", "For c s , c t in L-C , we use the first four characters to compare groups of words as suggested in (Kondrak et al., 2003) .", "Cognate Ranking Given that the n-best lists contain noise, we aim to prune them by an ML ranking model.", "However, there is a lack of resources to train a classification model for cognates (i.e.", "cognate vs. false friend), as mentioned in (Fišer and Ljubešić, 2013) .", "Available data that can be used to judge the cognate lists are the alignment pairs extracted from parallel data.", "We decide to use a ranking model to avoid data imbalance present in classification and to use the probability scores of the alignment pairs as ranks, as opposed to the classification model used by (Irvine and Callison-Burch, 2013) .", "Moreover, we also use a popular domain adaptation technique (Daumé et al., 2010) given that we have access to different domains of parallel training data that might be compatible with our comparable corpora.", "The training data are the alignments between pairs of words where we rank them accordingly to their correspondent alignment probability from the output of GIZA++ (Och and Ney, 2003) .", "We then use a heuristic to prune training data in order to simulate cognate words.", "Pairs of words scored below the Levenshtein similarity threshold of 0.5 are not considered as cognates given that they are likely to have a different surface form.", "We represent the training and test data with features extracted from different edit distance scores and distributional measures.", "The edit distances features are as follows: 1) Similarity measure L and 2) Number of times of each edit operation.", "Thus, the model assigns a different importance to each operation.", "The distributional feature is based on the cosine between the distributional vectors of a window of n words around the word currently under comparison.", "We train distributional similarity models with word2vec (Mikolov et al., 2013a) for the source and target side separately.", "We extract the continuous vector for each word in the window, concatenate it and then compute the cosine between the concatenated vectors of the source and the target.", "We suspect that the vectors will have similar behaviour between the source and the target given that they are trained under parallel Wikipedia articles.", "We develop two ML models: 1) Edit distance scores and 2) Edit distance scores and distributional similarity score.", "We use SVMlight (Joachims, 1998) Results and Discussion In this section we describe the data used to produce the n-best lists and train the cognate ranking models.", "We evaluate the ranking models with heldout data from each training domain.", "We also provide manual evaluation over the ranked n-best lists for error analysis.", "Data The n-best lists to detect cognates were extracted from the respective Wikipedias by using the method described in Section 2.1.", "The training data for the ranking model consists of different types of parallel corpora.", "The parallel corpora are as follows: 1) Wiki-titles we use the inter language links to create a parallel corpus from the tittles of the Wikipedia articles, with about 500K aligned links (i.e.", "'sentences') per language pair (about 200k for bg-ru), giving us about 200K training instances per language pair 3 , 2) Opensubs is an open source corpus of subtitles built by the fan community, with 1M sentences, 6M tokens, 100K words, giving about 90K training instances (Tiedemann, 2012) and 3) Zoo is a proprietary corpus of subtitles produced by professional translators, with 100K sentences, 700K tokens, 40K words and giving about 20K training instances per language pair.", "Our objective is to create MT dictionaries from the produced n-best lists and we use parallel data as a source of training to prune them.", "We are interested in the corpora of subtitles because the chosen domain of our SMT experiments is subtitling, while the proposed ranking method can be used in other application domains as well.", "We consider Zoo and Opensubs as two different domains given that they were built by different types of translators and they differ in size and quality.", "The heldout data consists of 2K instances for each corpus.", "We use Wikipedia documents and Opensusbs subtitles to train word2vec for the distributional similarity feature.", "We use the continuous bag-ofwords algorithm for word2vec and set the parameters for training to 200 dimensions and a window of 8 words.", "The Wikipedia documents with an average number of 70K documents for each language, and Opensubs subtitles with 1M sentences for each language.", "In practice we only use the Wikipedia data given that for Opensubs the model is able to find relatively few vectors, for example a vector is found only for 20% of the words in the pt-es pair.", "Evaluation of the Ranking Model We define two ranking models as: model E for edit distance features and model EC for both edit Table 1 shows the results of the ranking procedure.", "For the Romance family language pairs the model EC with context features consistently reduces the error compared to the solely use of edit distance metrics.", "The only exception is the it-es EC model with poor results for the domain of Wiki-titles.", "The models for the Slavonic family behave similarly to the Romance family, where the use of context features reduces the ranking error.", "The exception is the bg-ru model on the Opensubs domain.", "A possible reason for the poor results on the ites and bg-ru models is that the model often assigns a high similarity score to unrelated words.", "For example, in it-es, mortes 'deads' is treated as close to categoria 'category'.", "A possible solution is to map the vectors form the source side into the space of the target side via a learned transformation matrix (Mikolov et al., 2013b) .", "Preliminary Results on Comparable Corpora After we extracted the n-best lists for the Romance family comparable corpora, we applied one of the ranking models on these lists and we manually evaluated over a sample of 50 words 4 .", "We set n to 10 for the n-best lists.", "We use a frequency window of 200 for the n-best list search heuristic and the domain of the comparable corpora to Wiki-titles 4 The sample consists of words with a frequency between 2K and 5. for the domain adaptation technique.", "The purpose of manual evaluation is to decide whether the ML setup is sensible on the objective task.", "Each list is evaluated by accuracy at 1 and accuracy at 10.", "We also show success and failure examples of the ranking and the n-best lists.", "Table 2 shows the preliminary results of the ML model E on a sample of Wikipedia dumps.", "The L and L-R lists consistently show poor results.", "A possible reason is the amount of errors given the first step to extract the n-best lists.", "For example, in pt-es, for the word vivem 'live' the 10-best list only contain one word with a similar meaning viva 'living' but it can be also translated as 'cheers'.", "In the pt-es list for the word representação 'description' the correct translation representación is not among the 10-best in the L list.", "However, it is present in the 10-best for the L-C list and the ML model EC ranks it in the first place.", "The edit distance model E still makes mistakes like with the list L-C, the word vivem 'live' translates into viven 'living' and the correct translation is vivir.", "However, given a certain context/sense the previous translation can be correct.", "The ranking scores given by the SVM varies from each list version.", "For the L-C lists the scores are more uniform in increasing order and with a small variance.", "The L and L-R lists show the opposite behaviour.", "We add the produced Wikipedia n-best lists with the L metric into a SMT training dataset for the ptes pair.", "We use the Moses SMT toolkit (Koehn et al., 2007) to test the augmented datasets.", "We compare the augmented model with a baseline both trained by using the Zoo corpus of subtitles.", "We use a 1-best list consisting of 100K pairs.", "Te dataset used for pt-es baseline is: 80K training sentences, 1K sentences for tuning and 2K sen- Lang Pairs acc@1 acc@10 acc@1 acc@10 acc@1 acc@10 pt-es 20 60 22 59 32 70 it-es 16 53 18 45 44 66 fr-es 10 48 12 51 29 59 A possible reason for low improvement in terms of the BLEU scores is because MT evaluation metrics, such as BLEU, compare the MT output with a human reference.", "The human reference translations in our corpus have been done from English (e.g., En→Es), while the test translations come from a related language (En→Pt→Es), often resulting in different paraphrases of the same English source.", "While our OOV rate improved, the evaluation scores did not reflected this, because our MT output was still far from the reference even in cases it was otherwise acceptable.", "List L List L-R List L-C Conclusions and future Work We have presented work in progress for developing MT dictionaries extracted from comparable resources for related languages.", "The extraction heuristic show positive results on the n-best lists that group words with the same starting char-5 https://github.com/clab/fast_align 6 https://kheafield.com/code/kenlm/ 7 The p-value for the uk-ru pair is 0.06 we do not consider this result as statistically significant.", "acters, because the used comparable corpora consist of related languages that share a similar orthography.", "However, the lists based on the frequency window heuristic show poor results to include the correct translations during the extraction step.", "Our ML models based on similarity metrics over parallel corpora show rankings similar to heldout data.", "However, we created our training data using simple heuristics that simulate cognate words (i.e.", "pairs of words with a small surface difference).", "The ML models are able to rank similar words on the top of the list and they give a reliable score to discriminate wrong translations.", "Preliminary results on the addition of the n-best lists into an SMT system show modest improvements compare to the baseline.", "However, the OOV rate shows improvements around 10% reduction on word types, because of the wide variety of lexical choices introduced by the MT dictionaries.", "Future work involves the addition of unsupervised morphology features for the n-best list extraction, i.e.", "first step, given that the use of starting characters shows to be an effective heuristic to prune the search space and language independent.", "Finally, we will measure the contribution for all the produced cognate lists, where we can try different strategies to add the dictionaries into an SMT system (Irvine and Callison-Burch, 2014) ." ] }

{ "paper_header_number": [ "1", "2", "2.1", "2.2", "3", "3.1", "3.2", "3.3", "4" ], "paper_header_content": [ "Introduction", "Methodology", "Cognate detection", "Cognate Ranking", "Results and Discussion", "Data", "Evaluation of the Ranking Model", "Preliminary Results on Comparable Corpora", "Conclusions and future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193 ], "paper_content_text": [ "Introduction Topic models comprise a family of methods for uncovering latent structure in text corpora, and are widely used tools in the digital humanities, political science, and other related fields (Boyd-Graber et al., 2017) .", "Latent Dirichlet allocation (LDA; Blei et al., 2003) is often used when there is no prior knowledge about a corpus.", "In the real world, however, most documents have non-textual attributes such as author (Rosen-Zvi et al., 2004) , timestamp , rating (McAuliffe and Blei, 2008) , or ideology (Eisenstein et al., 2011; Nguyen et al., 2015b) , which we refer to as metadata.", "Many customizations of LDA have been developed to incorporate document metadata.", "Two models of note are supervised LDA (SLDA; McAuliffe and Blei, 2008) , which jointly models words and labels (e.g., ratings) as being generated from a latent representation, and sparse additive generative models (SAGE; Eisenstein et al., 2011) , which assumes that observed covariates (e.g., author ideology) have a sparse effect on the relative probabilities of words given topics.", "The structural topic model (STM; Roberts et al., 2014) , which adds correlations between topics to SAGE, is also widely used, but like SAGE it is limited in the types of metadata it can efficiently make use of, and how that metadata is used.", "Note that in this work we will distinguish labels (metadata that are generated jointly with words from latent topic representations) from covariates (observed metadata that influence the distribution of labels and words).", "The ability to create variations of LDA such as those listed above has been limited by the expertise needed to develop custom inference algorithms for each model.", "As a result, it is rare to see such variations being widely used in practice.", "In this work, we take advantage of recent advances in variational methods (Kingma and Welling, 2014; Rezende et al., 2014; Miao et al., 2016; Srivastava and Sutton, 2017) to facilitate approximate Bayesian inference without requiring model-specific derivations, and propose a general neural framework for topic models with metadata, SCHOLAR.", "1 SCHOLAR combines the abilities of SAGE and SLDA, and allows for easy exploration of the following options for customization: 1.", "Covariates: as in SAGE and STM, we incorporate explicit deviations for observed covariates, as well as effects for interactions with topics.", "2.", "Supervision: as in SLDA, we can use metadata as labels to help infer topics that are relevant in predicting those labels.", "3.", "Rich encoder network: we use the encoding network of a variational autoencoder (VAE) to incorporate additional prior knowledge in the form of word embeddings, and/or to provide interpretable embeddings of covariates.", "4.", "Sparsity: as in SAGE, a sparsity-inducing prior can be used to encourage more interpretable topics, represented as sparse deviations from a background log-frequency.", "We begin with the necessary background and motivation ( §2), and then describe our basic framework and its extensions ( §3), followed by a series of experiments ( §4).", "In an unsupervised setting, we can customize the model to trade off between perplexity, coherence, and sparsity, with improved coherence through the introduction of word vectors.", "Alternatively, by incorporating metadata we can either learn topics that are more predictive of labels than SLDA, or learn explicit deviations for particular parts of the metadata.", "Finally, by combining all parts of our model we can meaningfully incorporate metadata in multiple ways, which we demonstrate through an exploration of a corpus of news articles about US immigration.", "In presenting this particular model, we emphasize not only its ability to adapt to the characteristics of the data, but the extent to which the VAE approach to inference provides a powerful framework for latent variable modeling that suggests the possibility of many further extensions.", "Our implementation is available at https://github.", "com/dallascard/scholar.", "Background and Motivation LDA can be understood as a non-negative Bayesian matrix factorization model: the observed document-word frequency matrix, X ∈ Z D×V (D is the number of documents, V is the vocabulary size) is factored into two low-rank matrices, Θ D×K and B K×V , where each row of Θ, θ i ∈ ∆ K is a latent variable representing a distribution over topics in document i, and each row of B, β k ∈ ∆ V , represents a single topic, i.e., a distribution over words in the vocabulary.", "2 While it is possible to factor the count data into unconstrained 2 Z denotes nonnegative integers, and ∆ K denotes the set of K-length nonnegative vectors that sum to one.", "For a proper probabilistic interpretation, the matrix to be factored is actually the matrix of latent mean parameters of the assumed data generating process, Xij ∼ Poisson(Λij).", "See Cemgil (2009) or Paisley et al.", "(2014) for details.", "matrices, the particular priors assumed by LDA are important for interpretability (Wallach et al., 2009) .", "For example, the neural variational document model (NVDM; Miao et al., 2016) allows θ i ∈ R K and achieves normalization by taking the softmax of θ i B.", "However, the experiments in Srivastava and Sutton (2017) found the performance of the NVDM to be slightly worse than LDA in terms of perplexity, and dramatically worse in terms of topic coherence.", "The topics discovered by LDA tend to be parsimonious and coherent groupings of words which are readily identifiable to humans as being related to each other (Chang et al., 2009) , and the resulting mode of the matrix Θ provides a representation of each document which can be treated as a measurement for downstream tasks, such as classification or answering social scientific questions (Wallach, 2016) .", "LDA does not require -and cannot make use of -additional prior knowledge.", "As such, the topics that are discovered may bear little connection to metadata of a corpus that is of interest to a researcher, such as sentiment, ideology, or time.", "In this paper, we take inspiration from two models which have sought to alleviate this problem.", "The first, supervised LDA (SLDA; McAuliffe and Blei, 2008) , assumes that documents have labels y which are generated conditional on the corresponding latent representation, i.e., y i ∼ p(y | θ i ).", "3 By incorporating labels into the model, it is forced to learn topics which allow documents to be represented in a way that is useful for the classification task.", "Such models can be used inductively as text classifiers (Balasubramanyan et al., 2012) .", "SAGE (Eisenstein et al., 2011) , by contrast, is an exponential-family model, where the key innovation was to replace topics with sparse deviations from the background log-frequency of words (d), i.e., p(word | softmax(d + θ i B)).", "SAGE can also incorporate deviations for observed covariates, as well as interactions between topics and covariates, by including additional terms inside the softmax.", "In principle, this allows for inferring, for example, the effect on an author's ideology on their choice of words, as well as ideological variations on each underlying topic.", "Unlike the NVDM, SAGE still constrains θ i to lie on the simplex, as in LDA.", "SLDA and SAGE provide two different ways that users might wish to incorporate prior knowl-edge as a way of guiding the discovery of topics in a corpus: SLDA incorporates labels through a distribution conditional on topics; SAGE includes explicit sparse deviations for each unique value of a covariate, in addition to topics.", "4 Because of the Dirichlet-multinomial conjugacy in the original model, efficient inference algorithms exist for LDA.", "Each variation of LDA, however, has required the derivation of a custom inference algorithm, which is a time-consuming and errorprone process.", "In SLDA, for example, each type of distribution we might assume for p(y | θ) would require a modification of the inference algorithm.", "SAGE breaks conjugacy, and as such, the authors adopted L-BFGS for optimizing the variational bound.", "Moreover, in order to maintain computational efficiency, it assumed that covariates were limited to a single categorical label.", "More recently, the variational autoencoder (VAE) was introduced as a way to perform approximate posterior inference on models with otherwise intractable posteriors (Kingma and Welling, 2014; Rezende et al., 2014) .", "This approach has previously been applied to models of text by Miao et al.", "(2016) and Srivastava and Sutton (2017) .", "We build on their work and show how this framework can be adapted to seamlessly incorporate the ideas of both SAGE and SLDA, while allowing for greater flexibility in the use of metadata.", "Moreover, by exploiting automatic differentiation, we allow for modification of the model without requiring any change to the inference procedure.", "The result is not only a highly adaptable family of models with scalable inference and efficient prediction; it also points the way to incorporation of many ideas found in the literature, such as a gradual evolution of topics , and hierarchical models (Blei et al., 2010; Nguyen et al., 2013 Nguyen et al., , 2015b ).", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity We begin by presenting the generative story for our model, and explain how it generalizes both SLDA and SAGE ( §3.1).", "We then provide a general explanation of inference using VAEs and how it applies to our model ( §3.2), as well as how to infer docu-4 A third way of incorporating metadata is the approach used by various \"upstream\" models, such as Dirichletmultinomial regression (Mimno and McCallum, 2008) , which uses observed metadata to inform the document prior.", "We hypothesize that this approach could be productively combined with our framework, but we leave this as future work.", "ment representations and predict labels at test time ( §3.3).", "Finally, we discuss how we can incorporate additional prior knowledge ( §3.4).", "Generative Story Consider a corpus of D documents, where document i is a list of N i words, w i , with V words in the vocabulary.", "For each document, we may have observed covariates c i (e.g., year of publication), and/or one or more labels, y i (e.g., sentiment).", "Our model builds on the generative story of LDA, but optionally incorporates labels and covariates, and replaces the matrix product of Θ and B with a more flexible generative network, f g , followed by a softmax transform.", "Instead of using a Dirichlet prior as in LDA, we employ a logistic normal prior on θ as in Srivastava and Sutton (2017) to facilitate inference ( §3.2): we draw a latent variable, r, 5 from a multivariate normal, and transform it to lie on the simplex using a softmax transform.", "6 The generative story is shown in Figure 1a and described in equations below: For each document i of length N i : # Draw a latent representation on the simplex from a logistic normal prior: r i ∼ N (r | µ 0 (α), diag(σ 2 0 (α))) θ i = softmax(r i ) # Generate words, incorporating covariates: η i = f g (θ i , c i ) For each word j in document i: w ij ∼ p(w | softmax(η i )) # Similarly generate labels: y i ∼ p(y | f y (θ i , c i )), where p(w | softmax(η i )) is a multinomial distribution and p(y | f y (θ i , c i )) is a distribution appropriate to the data (e.g., multinomial for categorical labels).", "f g is a model-specific combination of latent variables and covariates, f y is a multi-layer neural network, and µ 0 (α) and σ 2 0 (α) are the mean and diagonal covariance terms of a multivariate normal prior.", "To approximate a symmetric Dirichlet prior with hyperparameter α, these are given by the Laplace approximation (Hennig et al., 2012) to be µ 0,k (α) = 0 and σ 2 0,k = (K − 1)/(αK).", "If we were to ignore covariates, place a Dirichlet prior on B, and let η = θ i B, this model is equivalent to SLDA with a logistic normal prior.", "Similarly, we can recover a model that is like SAGE, but lacks sparsity, if we ignore labels, and let (1) where d is the V -dimensional background term (representing the log of the overall word frequency), θ i ⊗ c i is a vector of interactions between topics and covariates, and B cov and B int are additional weight (deviation) matrices.", "The background is included to account for common words with approximately the same frequency across documents, meaning that the B * weights now represent both positive and negative deviations from this background.", "This is the form of f g which we will use in our experiments.", "η i = d + θ i B + c i B cov + (θ i ⊗ c i ) B int , To recover the full SAGE model, we can place a sparsity-inducing prior on each B * .", "As in Eisenstein et al.", "(2011) , we make use of the compound normal-exponential prior for each element of the weight matrices, B * m,n , with hyperparameter γ, 7 τ m,n ∼ Exponential(γ), (2) B * m,n ∼ N (0, τ m,n ).", "(3) We can choose to ignore various parts of this model, if, for example, we don't have any labels or observed covariates, or we don't wish to use interactions or sparsity.", "8 Other generator networks could also be considered, with additional layers to represent more complex interactions, although this might involve some loss of interpretability.", "In the absence of metadata, and without sparsity, our model is equivalent to the ProdLDA model of Srivastava and Sutton (2017) with an explicit background term, and ProdLDA is, in turn, a 7 To avoid having to tune γ, we employ an improper Jeffery's prior, p(τm,n) ∝ 1/τm,n, as in SAGE.", "Although this causes difficulties in posterior inference for the variance terms, τ , in practice, we resort to a variational EM approach, with MAP-estimation for the weights, B, and thus alternate between computing expectations of the τ parameters, and updating all other parameters using some variant of stochastic gradient descent.", "For this, we only require the expectation of each τmn for each E-step, which is given by 1/B 2 m,n .", "We refer the reader to Eisenstein et al.", "(2011) for additional details.", "8 We could also ignore latent topics, in which case we would get a naïve Bayes-like model of text with deviations for each covariate p(wij | ci) ∝ exp(d + c i B cov ).", "Figure 1a presents the generative story of our model.", "Figure 1b illustrates the inference network using the reparametrization trick to perform variational inference on our model.", "Shaded nodes are observed; double circles indicate deterministic transformations of parent nodes.", "special case of SAGE, without background logfrequencies, sparsity, covariates, or labels.", "In the next section we generalize the inference method used for ProdLDA; in our experiments we validate its performance and explore the effects of regularization and word-vector initialization ( §3.4).", "The NVDM (Miao et al., 2016) uses the same approach to inference, but does not not restrict document representations to the simplex.", "Learning and Inference As in past work, each document i is assumed to have a latent representation r i , which can be interpreted as its relative membership in each topic (after exponentiating and normalizing).", "In order to infer an approximate posterior distribution over r i , we adopt the sampling-based VAE framework developed in previous work (Kingma and Welling, 2014; Rezende et al., 2014) .", "As in conventional variational inference, we assume a variational approximation to the posterior, q Φ (r i | w i , c i , y i ), and seek to minimize the KL divergence between it and the true posterior, p(r i | w i , c i , y i ), where Φ is the set of variational parameters to be defined below.", "After some manipulations (given in supplementary materials), we obtain the evidence lower bound (ELBO) for a sin-gle document, L(w i ) = E q Φ (r i |w i ,c i ,y i )   N i j=1 log p(w ij | r i , c i )   + E q Φ (r i |w i ,c i ,y i ) [log p(y i | r i , c i )] − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(4) As in the original VAE, we will encode the parameters of our variational distributions using a shared multi-layer neural network.", "Because we have assumed a diagonal normal prior on r, this will take the form of a network which outputs a mean vector, µ i = f µ (w i , c i , y i ) and diagonal of a covariance matrix, σ 2 i = f σ (w i , c i , y i ), such that q Φ (r i | w i , c i , y i ) = N (µ i , σ 2 i ) .", "Incorporating labels and covariates to the inference network used by Miao et al.", "(2016) and Srivastava and Sutton (2017) , we use: π i = f e ([W x x i ; W c c i ; W y y i ]), (5) µ i = W µ π i + b µ , (6) log σ 2 i = W σ π i + b σ , (7) where x i is a V -dimensional vector representing the counts of words in w i , and f e is a multilayer perceptron.", "The full set of encoder parameters, Φ, thus includes the parameters of f e and all weight matrices and bias vectors in Equations 5-7 (see Figure 1b ).", "This approach means that the expectations in Equation 4 are intractable, but we can approximate them using sampling.", "In order to maintain differentiability with respect to Φ, even after sampling, we make use of the reparameterization trick (Kingma and Welling, 2014), 9 which allows us to reparameterize samples from q Φ (r | w i , c i , y i ) in terms of samples from an independent source of noise, i.e., (s) ∼ N (0, I), r (s) i = g Φ (w i , c i , y i , (s) ) = µ i + σ i · (s) .", "We thus replace the bound in Equation 4 with a Monte Carlo approximation using a single sam- 9 The Dirichlet distribution cannot be directly reparameterized in this way, which is why we use the logistic normal prior on θ to approximate the Dirichlet prior used in LDA.", "ple 10 of (and thereby of r): L(w i ) ≈ N i j=1 log p(w ij | r (s) i , c i ) + log p(y i | r (s) i , c i ) − D KL [q Φ (r i | w i , c i , y i ) || p(r i | α)] .", "(8) We can now optimize this sampling-based approximation of the variational bound with respect to Φ, B * , and all parameters of f g and f y using stochastic gradient descent.", "Moreover, because of this stochastic approach to inference, we are not restricted to covariates with a small number of unique values, which was a limitation of SAGE.", "Finally, the KL divergence term in Equation 8 can be computed in closed form (see supplementary materials).", "Prediction on Held-out Data In addition to inferring latent topics, our model can both infer latent representations for new documents and predict their labels, the latter of which was the motivation for SLDA.", "In traditional variational inference, inference at test time requires fixing global parameters (topics), and optimizing the per-document variational parameters for the test set.", "With the VAE framework, by contrast, the encoder network (Equations 5-7) can be used to directly estimate the posterior distribution for each test document, using only a forward pass (no iterative optimization or sampling).", "If not using labels, we can use this approach directly, passing the word counts of new documents through the encoder to get a posterior q Φ (r i | w i , c i ).", "When we also include labels to be predicted, we can first train a fully-observed model, as above, then fix the decoder, and retrain the encoder without labels.", "In practice, however, if we train the encoder network using one-hot encodings of document labels, it is sufficient to provide a vector of all zeros for the labels of test documents; this is what we adopt for our experiments ( §4.2), and we still obtain good predictive performance.", "The label network, f y , is a flexible component which can be used to predict a wide range of outcomes, from categorical labels (such as star ratings; McAuliffe and Blei, 2008) to real-valued outputs (such as number of citations or box-office returns; Yogatama et al., 2011) .", "For categorical labels, predictions are given bŷ y i = argmax y ∈ Y p(y | r i , c i ).", "(9) Alternatively, when dealing with a small set of categorical labels, it is also possible to treat them as observed categorical covariates during training.", "At test time, we can then consider all possible one-hot vectors, e, in place of c i , and predict the label that maximizes the probability of the words, i.e., y i = argmax y ∈ Y N i j=1 log p(w ij | r i , e y ).", "(10) This approach works well in practice (as we show in §4.2), but does not scale to large numbers of labels, or other types of prediction problems, such as multi-class classification or regression.", "The choice to include metadata as covariates, labels, or both, depends on the data.", "The key point is that we can incorporate metadata in two very different ways, depending on what we want from the model.", "Labels guide the model to infer topics that are relevant to those labels, whereas covariates induce explicit deviations, leaving the latent variables to account for the rest of the content.", "Additional Prior Information A final advantage of the VAE framework is that the encoder network provides a way to incorporate additional prior information in the form of word vectors.", "Although we can learn all parameters starting from a random initialization, it is also possible to initialize and fix the initial embeddings of words in the model, W x , in Equation 5.", "This leverages word similarities derived from large amounts of unlabeled data, and may promote greater coherence in inferred topics.", "The same could also be done for some covariates; for example, we could embed the source of a news article based on its place on the ideological spectrum.", "Conversely, if we choose to learn these parameters, the learned values (W y and W c ) may provide meaningful embeddings of these metadata (see section §4.3).", "Other variants on topic models have also proposed incorporating word vectors, both as a parallel part of the generative process (Nguyen et al., 2015a) , and as an alternative parameterization of topic distributions (Das et al., 2015) , but inference is not scalable in either of these models.", "Because of the generality of the VAE framework, we could also modify the generative story so that word embeddings are emitted (rather than tokens); we leave this for future work.", "Experiments and Results To evaluate and demonstrate the potential of this model, we present a series of experiments below.", "We first test SCHOLAR without observed metadata, and explore the effects of using regularization and/or word vector initialization, compared to LDA, SAGE, and NVDM ( §4.1).", "We then evaluate our model in terms of predictive performance, in comparison to SLDA and an l 2 -regularized logistic regression baseline ( §4.2).", "Finally, we demonstrate the ability to incorporate covariates and/or labels in an exploratory data analysis ( §4.3).", "The scores we report are generalization to heldout data, measured in terms of perplexity; coherence, measured in terms of non-negative point-wise mutual information (NPMI; Chang et al., 2009; Newman et al., 2010) , and classification accuracy on test data.", "For coherence we evaluate NPMI using the top 10 words of each topic, both internally (using test data), and externally, using a decade of articles from the English Gigaword dataset (Graff and Cieri, 2003) .", "Since our model employs variational methods, the reported perplexity is an upper bound based on the ELBO.", "As datasets we use the familiar 20 newsgroups, the IMDB corpus of 50,000 movie reviews (Maas et al., 2011) , and the UIUC Yahoo answers dataset with 150,000 documents in 15 categories (Chang et al., 2008) .", "For further exploration, we also make use of a corpus of approximately 4,000 timestamped news articles about US immigration, each annotated with pro-or anti-immigration tone (Card et al., 2015) .", "We use the original author-provided implementations of SAGE 11 and SLDA, 12 while for LDA we use Mallet.", "13 .", "Our implementation of SCHOLAR is in TensorFlow, but we have also provided a preliminary PyTorch implementation of the core of our model.", "14 For additional details about datasets and implementation, please refer to the supplementary material.", "It is challenging to fairly evaluate the relative computational efficiency of our approach compared to past work (due to the stochastic nature of our ap-11 github.com/jacobeisenstein/SAGE 12 github.com/blei-lab/class-slda 13 mallet.cs.umass.edu 14 github.com/dallascard/scholar proach to inference, choices about hyperparameters such as tolerance, and because of differences in implementation).", "Nevertheless, in practice, the performance of our approach is highly appealing.", "For all experiments in this paper, our implementation was much faster than SLDA or SAGE (implemented in C and Matlab, respectively), and competitive with Mallet.", "Unsupervised Evaluation Although the emphasis of this work is on incorporating observed labels and/or covariates, we briefly report on experiments in the unsupervised setting.", "Recall that, without metadata, SCHOLAR equates to ProdLDA, but with an explicit background term.", "15 We therefore use the same experimental setup as Srivastava and Sutton (2017) (learning rate, momentum, batch size, and number of epochs) and find the same general patterns as they reported (see Table 1 and supplementary material): our model returns more coherent topics than LDA, but at the cost of worse perplexity.", "SAGE, by contrast, attains very high levels of sparsity, but at the cost of worse perplexity and coherence than LDA.", "As expected, the NVDM produces relatively low perplexity, but very poor coherence, due to its lack of constraints on θ.", "Further experimentation revealed that the VAE framework involves a tradeoff among the scores; running for more epochs tends to result in better perplexity on held-out data, but at the cost of worse coherence.", "Adding regularization to encourage sparse topics has a similar effect as in SAGE, leading to worse perplexity and coherence, but it does create sparse topics.", "Interestingly, initializing the encoder with pretrained word2vec embeddings, and not updating them returned a model with the best internal coherence of any model we considered for IMDB and Yahoo answers, and the second-best for 20 newsgroups.", "The background term in our model does not have much effect on perplexity, but plays an important role in producing coherent topics; as in SAGE, the background can account for common words, so they are mostly absent among the most heavily weighted words in the topics.", "For instance, words like film and movie in the IMDB corpus are relatively unimportant in the topics learned by our Table 1 : Performance of our various models in an unsupervised setting (i.e., without labels or covariates) on the IMDB dataset using a 5,000-word vocabulary and 50 topics.", "The supplementary materials contain additional results for 20 newsgroups and Yahoo answers.", "model, but would be much more heavily weighted without the background term, as they are in topics learned by LDA.", "Text Classification We next consider the utility of our model in the context of categorical labels, and consider them alternately as observed covariates and as labels generated conditional on the latent representation.", "We use the same setup as above, but tune number of training epochs for our model using a random 20% of training data as a development set, and similarly tune regularization for logistic regression.", "Table 2 summarizes the accuracy of various models on three datasets, revealing that our model offers competitive performance, both as a joint model of words and labels (Eq.", "9), and a model which conditions on covariates (Eq.", "10).", "Although SCHOLAR is comparable to the logistic regression baseline, our purpose here is not to attain state-of-the-art performance on text classification.", "Rather, the high accuracies we obtain demonstrate that we are learning low-dimensional representations of documents that are relevant to the label of interest, outperforming SLDA, and have the same attractive properties as topic models.", "Further, any neural network that is successful for text classification could be incorporated into f y and trained end-to-end along with topic discovery.", "Exploratory Study We demonstrate how our model might be used to explore an annotated corpus of articles about immigration, and adapt to different assumptions about the data.", "We only use a small number of topics in this part (K = 8) for compact presentation.", "Tone as a label.", "We first consider using the annotations as a label, and train a joint model to infer topics relevant to the tone of the article (pro-or anti-immigration).", "Figure 2 shows a set of topics learned in this way, along with the predicted probability of an article being pro-immigration conditioned on the given topic.", "All topics are coherent, and the predicted probabilities have strong face validity, e.g., \"arrested charged charges agents operation\" is least associated with pro-immigration.", "Tone as a covariate.", "Next we consider using tone as a covariate, and build a model using both tone and tone-topic interactions.", "Table 3 shows a set of topics learned from the immigration data, along with the most highly-weighted words in the corresponding tone-topic interaction terms.", "As can be seen, these interaction terms tend to capture different frames (e.g., \"criminal\" vs. \"detainees\", and \"illegals\" vs. \"newcomers\", etc).", "Combined model with temporal metadata.", "Finally, we incorporate both the tone annotations and the year of publication of each article, treating the former as a label and the latter as a covariate.", "In this model, we also include an embedding matrix, W c , to project the one-hot year vectors down to a two-dimensional continuous space, with a learned deviation for each dimension.", "We omit the topics in the interest of space, but Figure 3 shows the learned embedding for each year, along with the top terms of the corresponding deviations.", "As can be seen, the model learns that adjacent years tend to produce similar deviations, even though we have not explicitly encoded this information.", "The leftright dimension roughly tracks a temporal trend with positive deviations shifting from the years of Clinton and INS on the left, to Obama and ICE on the right.", "16 Meanwhile, the events of 9/11 dominate the vertical direction, with the words sept, hijackers, and attacks increasing in probability as we move up in the space.", "If we wanted to look at each year individually, we could drop the embedding of years, and learn a sparse set of topic-year interactions, similar to tone in Table 3 .", "Additional Related Work The literature on topic models is vast; in addition to papers cited throughout, other efforts to incorporate metadata into topic models include Dirichletmultinomial regression (DMR; Mimno and McCallum, 2008) , Labeled LDA (Ramage et al., 2009) , and MedLDA (Zhu et al., 2009) .", "A recent paper also extended DMR by using deep neural networks to embed metadata into a richer document prior (Benton and Dredze, 2018) .", "A separate line of work has pursued parameterizing unsupervised models of documents using neural networks (Hinton and Salakhutdinov, Base topics (each row is a topic) Anti-immigration interactions Pro-immigration interactions ice customs agency enforcement homeland criminal customs arrested detainees detention center agency population born percent americans english jobs million illegals taxpayers english newcomers hispanic city judge case court guilty appeals attorney guilty charges man charged asylum court judge case appeals patrol border miles coast desert boat guard patrol border agents boat died authorities desert border bodies licenses drivers card visa cards applicants foreign sept visas system green citizenship card citizen apply island story chinese ellis international smuggling federal charges island school ellis english story guest worker workers bush labor bill bill border house senate workers tech skilled farm labor benefits bill welfare republican state senate republican california gov state law welfare students tuition Table 3 : Top words for topics (left) and the corresponding anti-immigration (middle) and pro-immigration (right) variations when treating tone as a covariate, with interactions.", "2009; Larochelle and Lauly, 2012) , including non-Bayesian approaches (Cao et al., 2015) .", "More recently, Lau et al.", "(2017) proposed a neural language model that incorporated topics, and He et al.", "(2017) developed a scalable alternative to the correlated topic model by simultaneously learning topic embeddings.", "Others have attempted to extend the reparameterization trick to the Dirichlet and Gamma distributions, either through transformations or a generalization of reparameterization (Ruiz et al., 2016) .", "Black-box and VAE-style inference have been implemented in at least two general purpose tools designed to allow rapid exploration and evaluation of models (Kucukelbir et al., 2015; .", "Conclusion We have presented a neural framework for generalized topic models to enable flexible incorporation of metadata with a variety of options.", "We take advantage of stochastic variational inference to develop a general algorithm for our framework such that variations do not require any model-specific algorithm derivations.", "Our model demonstrates the tradeoff between perplexity, coherence, and sparsity, and outperforms SLDA in predicting document labels.", "Furthermore, the flexibility of our model enables intriguing exploration of a text corpus on US immigration.", "We believe that our model and code will facilitate rapid exploration of document collections with metadata." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Background and Motivation", "SCHOLAR: A Neural Topic Model with Covariates, Supervision, and Sparsity", "Generative Story", "Learning and Inference", "Prediction on Held-out Data", "Additional Prior Information", "Experiments and Results", "Unsupervised Evaluation", "Text Classification", "Exploratory Study", "Additional Related Work", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208 ], "paper_content_text": [ "Introduction Recurrent neural networks (RNNs) have recently proven to be very effective sequence modeling tools, and are now state of the art for tasks such as machine translation , image captioning (Kiros et al., 2014; Anderson et al., 2017) and automatic speech recognition (Chorowski et al., 2015; Chiu et al., 2017) .", "The basic principle of RNNs is to iteratively compute a vectorial sequence representation, by applying at each time-step the same trainable func-tion to compute the new network state from the previous state and the last symbol in the sequence.", "These models are typically trained by maximizing the likelihood of the target sentence given an encoded source (text, image, speech) .", "Maximum likelihood estimation (MLE), however, has two main limitations.", "First, the training signal only differentiates the ground-truth target output from all other outputs.", "It treats all other output sequences as equally incorrect, regardless of their semantic proximity from the ground-truth target.", "While such a \"zero-one\" loss is probably acceptable for coarse grained classification of images, e.g.", "across a limited number of basic object categories (Everingham et al., 2010) it becomes problematic as the output space becomes larger and some of its elements become semantically similar to each other.", "This is in particular the case for tasks that involve natural language generation (captioning, translation, speech recognition) where the number of possible outputs is practically unbounded.", "For natural language generation tasks, evaluation measures typically do take into account structural similarity, e.g.", "based on n-grams, but such structural information is not reflected in the MLE criterion.", "The second limitation of MLE is that training is based on predicting the next token given the input and preceding ground-truth output tokens, while at test time the model predicts conditioned on the input and the so-far generated output sequence.", "Given the exponentially large output space of natural language sentences, it is not obvious that the learned RNNs generalize well beyond the relatively sparse distribution of ground-truth sequences used during MLE optimization.", "This phenomenon is known as \"exposure bias\" (Ranzato et al., 2016; .", "MLE minimizes the KL divergence between a target Dirac distribution on the ground-truth sentence(s) and the model's distribution.", "In this pa-per, we build upon the \"loss smoothing\" approach by Norouzi et al.", "(2016) , which smooths the Dirac target distribution over similar sentences, increasing the support of the training data in the output space.", "We make the following main contributions: • We propose a token-level loss smoothing approach, using word-embeddings, to achieve smoothing among semantically similar terms, and we introduce a special procedure to promote rare tokens.", "• For sequence-level smoothing, we propose to use restricted token replacement vocabularies, and a \"lazy evaluation\" method that significantly speeds up training.", "• We experimentally validate our approach on the MSCOCO image captioning task and the WMT'14 English to French machine translation task, showing that on both tasks combining token-level and sequence-level loss smoothing improves results significantly over maximum likelihood baselines.", "In the remainder of the paper, we review the existing methods to improve RNN training in Section 2.", "Then, we present our token-level and sequence-level approaches in Section 3.", "Experimental evaluation results based on image captioning and machine translation tasks are laid out in Section 4.", "Related work Previous work aiming to improve the generalization performance of RNNs can be roughly divided into three categories: those based on regularization, data augmentation, and alternatives to maximum likelihood estimation.", "Regularization techniques are used to increase the smoothness of the function learned by the network, e.g.", "by imposing an 2 penalty on the network weights, also known as \"weight decay\".", "More recent approaches mask network activations during training, as in dropout (Srivastava et al., 2014) and its variants adapted to recurrent models (Pham et al., 2014; Krueger et al., 2017) .", "Instead of masking, batch-normalization (Ioffe and Szegedy, 2015) rescales the network activations to avoid saturating the network's non-linearities.", "Instead of regularizing the network parameters or activations, it is also possible to directly regularize based on the entropy of the output distribution (Pereyra et al., 2017) .", "Data augmentation techniques improve the ro-bustness of the learned models by applying transformations that might be encountered at test time to the training data.", "In computer vision, this is common practice, and implemented by, e.g., scaling, cropping, and rotating training images (Le-Cun et al., 1998; Krizhevsky et al., 2012; Paulin et al., 2014) .", "In natural language processing, examples of data augmentation include input noising by randomly dropping some input tokens (Iyyer et al., 2015; Bowman et al., 2015; Kumar et al., 2016) , and randomly replacing words with substitutes sampled from the model .", "Xie et al.", "(2017) introduced data augmentation schemes for RNN language models that leverage n-gram statistics in order to mimic Kneser-Ney smoothing of n-grams models.", "In the context of machine translation, Fadaee et al.", "(2017) modify sentences by replacing words with rare ones when this is plausible according to a pretrained language model, and substitutes its equivalent in the target sentence using automatic word alignments.", "This approach, however, relies on the availability of additional monolingual data for language model training.", "The de facto standard way to train RNN language models is maximum likelihood estimation (MLE) .", "The sequential factorization of the sequence likelihood generates an additive structure in the loss, with one term corresponding to the prediction of each output token given the input and the preceding ground-truth output tokens.", "In order to directly optimize for sequence-level structured loss functions, such as measures based on n-grams like BLEU or CIDER, Ranzato et al.", "(2016) use reinforcement learning techniques that optimize the expectation of a sequence-level reward.", "In order to avoid early convergence to poor local optima, they pre-train the model using MLE.", "Leblond et al.", "(2018) build on the learning to search approach to structured prediction (Daumé III et al., 2009; Chang et al., 2015) and adapts it to RNN training.", "The model generates candidate sequences at each time-step using all possible tokens, and scores these at sequence-level to derive a training signal for each time step.", "This leads to an approach that is structurally close to MLE, but computationally expensive.", "Norouzi et al.", "(2016) introduce a reward augmented maximum likelihood (RAML) approach, that incorpo-rates a notion of sequence-level reward without facing the difficulties of reinforcement learning.", "They define a target distribution over output sentences using a soft-max over the reward over all possible outputs.", "Then, they minimize the KL divergence between the target distribution and the model's output distribution.", "Training with a general reward distribution is similar to MLE training, except that we use multiple sentences sampled from the target distribution instead of only the ground-truth sentences.", "In our work, we build upon the work of Norouzi et al.", "(2016) by proposing improvements to sequence-level smoothing, and extending it to token-level smoothing.", "Our token-level smoothing approach is related to the label smoothing approach of Szegedy et al.", "(2016) for image classification.", "Instead of maximizing the probability of the correct class, they train the model to predict the correct class with a large probability and all other classes with a small uniform probability.", "This regularizes the model by preventing overconfident predictions.", "In natural language generation with large vocabularies, preventing such \"narrow\" over-confident distributions is imperative, since for many tokens there are nearly interchangeable alternatives.", "Loss smoothing for RNN training We briefly recall standard recurrent neural network training, before presenting sequence-level and token-level loss smoothing below.", "Maximum likelihood RNN training We are interested in modeling the conditional probability of a sequence y = (y 1 , .", ".", ".", ", y T ) given a conditioning observation x, p θ (y|x) = T t=1 p θ (y t |x, y <t ), (1) where y <t = (y 1 , .", ".", ".", ", y t−1 ), the model parameters are given by θ, and x is a source sentence or an image in the contexts of machine translation and image captioning, respectively.", "In a recurrent neural network, the sequence y is predicted based on a sequence of states h t , p θ (y t |x, y <t ) = p θ (y t |h t ), (2) where the RNN state is computed recursively as h t = f θ (h t−1 , y t−1 , x) for t ∈ {1, ..T }, g θ (x) for t = 0.", "(3) The input is encoded by g θ and used to initialize the state sequence, and f θ is a non-linear function that updates the state given the previous state h t−1 , the last output token y t−1 , and possibly the input x.", "The state update function can take different forms, the ones including gating mechanisms such as LSTMs (Hochreiter and Schmidhuber, 1997) and GRUs (Chung et al., 2014) are particularly effective to model long sequences.", "In standard teacher-forced training, the hidden states will be computed by forwarding the ground truth sequence y * i.e.", "in Eq.", "(3) , the RNN has access to the true previous token y * t−1 .", "In this case we will note the hidden states h * t .", "Given a ground-truth target sequence y * , maximum likelihood estimation (MLE) of the network parameters θ amounts to minimizing the loss MLE (y * , x) = − ln p θ (y * |x) (4) = − T t=1 ln p θ (y * t |h * t ).", "(5) The loss can equivalently be expressed as the KLdivergence between a Dirac centered on the target output (with δ a (x) = 1 at x = a and 0 otherwise) and the model distribution, either at the sequencelevel or at the token-level: MLE (y * , x) = D KL δ y * ||p θ (y|x) (6) = T t=1 D KL δ y * t ||p θ (y t |h * t ) .", "(7) Loss smoothing approaches considered in this paper consist in replacing the Dirac on the groundtruth sequence with distributions with larger support.", "These distributions can be designed in such a manner that they reflect which deviations from ground-truth predictions are preferred over others.", "Sequence-level loss smoothing The reward augmented maximum likelihood approach of Norouzi et al.", "(2016) consists in replacing the sequence-level Dirac δ y * in Eq.", "(6) with a distribution r(y|y * ) ∝ exp r(y, y * )/τ, where r(y, y * ) is a \"reward\" function that measures the quality of sequence y w.r.t.", "y * , e.g.", "metrics used for evaluation of natural language processing tasks can be used, such as BLEU (Papineni et al., 2002) or CIDER (Vedantam et al., 2015) .", "The temperature parameter τ controls the concentration of the distribution around y * .", "When m > 1 ground-truth sequences are paired with the same input x, the reward function can be adapted to fit this setting and be defined as r(y, {y * (1) , .", ".", ".", ", y * (m) }).", "The sequence-level smoothed loss function is then given by Seq (y * , x) = D KL r(y|y * )||p θ (y|x) = H(r(y|y * )) − E r [ln p θ (y|x)] , (9) where the entropy term H(r(y|y * )) does not depend on the model parameters θ.", "In general, expectation in Eq.", "(9) is intractable due to the exponentially large output space, and replaced with a Monte-Carlo approximation: E r [− ln p θ (y|x)] ≈ − L l=1 ln p θ (y l |x).", "(10) Stratified sampling.", "Norouzi et al.", "(2016) show that when using the Hamming or edit distance as a reward, we can sample directly from r(y|y * ) using a stratified sampling approach.", "In this case sampling proceeds in three stages.", "(i) Sample a distance d from {0, .", ".", ".", ", T } from a prior distribution on d. (ii) Uniformly select d positions in the sequence to be modified.", "(iii) Sample the d substitutions uniformly from the token vocabulary.", "Details on the construction of the prior distribution on d for a reward based on the Hamming distance can be found in Appendix A.", "Importance sampling.", "For a reward based on BLEU or CIDER , we cannot directly sample from r(y|y * ) since the normalizing constant, or \"partition function\", of the distribution is intractable to compute.", "In this case we can resort to importance sampling.", "We first sample L sequences y l from a tractable proposal distribution q(y|y * ).", "We then compute the importance weights ω l ≈ r(y l |y * )/q(y l |y * ) L k=1 r(y k |y * )/q(y k |y * ) , (11) where r(y k |y * ) is the un-normalized reward distribution in Eq.", "(8).", "We finally approximate the expectation by reweighing the samples in the Monte Carlo approximation as E r [− ln p θ (y|x)] ≈ − L l=1 ω l ln p θ (y l |x).", "(12) In our experiments we use a proposal distribution based on the Hamming distance, which allows for tractable stratified sampling, and generates sentences that do not stray away from the ground truth.", "We propose two modifications to the sequencelevel loss smoothing of Norouzi et al.", "(2016) : sampling to a restricted vocabulary (described in the following paragraph) and lazy sequence-level smoothing (described in section 3.4).", "Restricted vocabulary sampling.", "In the stratified sampling method for Hamming and edit distance rewards, instead of drawing from the large vocabulary V, containing typically in the order of 10 4 words or more, we can restrict ourselves to a smaller subset V sub more adapted to our task.", "We considered three different possibilities for V sub .", "V : the full vocabulary from which we sample uniformly (default), or draw from our token-level smoothing distribution defined below in Eq.", "(13) .", "V ref s : uniformly sample from the set of tokens that appear in the ground-truth sentence(s) associated with the current input.", "V batch : uniformly sample from the tokens that appear in the ground-truth sentences across all inputs that appear in a given training mini-batch.", "Uniformly sampling from V batch has the effect of boosting the frequencies of words that appear in many reference sentences, and thus approximates to some extent sampling substitutions from the uni-gram statistics of the training set.", "Token-level loss smoothing While the sequence-level smoothing can be directly based on performance measures of interest such as BLEU or CIDEr, the support of the smoothed distribution is limited to the number of samples drawn during training.", "We propose smoothing the token-level Diracs δ y * t in Eq.", "(7) to increase its support to similar tokens.", "Since we apply smoothing to each of the tokens independently, this approach implicitly increases the support to an exponential number of sequences, unlike the sequence-level smoothing approach.", "This comes at the price, however, of a naive token-level independence assumption in the smoothing.", "We define the smoothed token-level distribution, similar as the sequence-level one, as a softmax over a token-level \"reward\" function, r(y t |y * where τ is again a temperature parameter.", "As a token-level reward r(y t , y * t ) we use the cosine similarity between y t and y * t in a semantic wordembedding space.", "In our experiments we use GloVe (Pennington et al., 2014) ; preliminary experiments with word2vec (Mikolov et al., 2013) yielded somewhat worse results.", "Promoting rare tokens.", "We can further improve the token-level smoothing by promoting rare tokens.", "To do so, we penalize frequent tokens when smoothing over the vocabulary, by subtracting β freq(y t ) from the reward, where freq(·) denotes the term frequency and β is a non-negative weight.", "This modification encourages frequent tokens into considering the rare ones.", "We experimentally found that it is also beneficial for rare tokens to boost frequent ones, as they tend to have mostly rare tokens as neighbors in the wordembedding space.", "With this in mind, we define a new token-level reward as: r freq (y t , y * t ) = r(y t , y * t ) (14) − β min freq(y t ) freq(y * t ) , freq(y * t ) freq(y t ) , where the penalty term is strongest if both tokens have similar frequencies.", "Combining losses In both loss smoothing methods presented above, the temperature parameter τ controls the concentration of the distribution.", "As τ gets smaller the distribution peaks around the ground-truth, while for large τ the uniform distribution is approached.", "We can, however, not separately control the spread of the distribution and the mass reserved for the ground-truth output.", "We therefore introduce a second parameter α ∈ [0, 1] to interpolate between the Dirac on the ground-truth and the smooth distribution.", "Usingᾱ = 1 − α, the sequence-level and token-level loss functions are then defined as To benefit from both sequence-level and tokenlevel loss smoothing, we also combine them by applying token-level smoothing to the different sequences sampled for the sequence-level smoothing.", "We introduce two mixing parameters α 1 and α 2 .", "The first controls to what extent sequencelevel smoothing is used, while the second controls to what extent token-level smoothing is used.", "The combined loss is defined as α 1 ,α 2 Seq, Tok (y * , x, r) = α 1 E r [ Tok (y, x)] +ᾱ 1 Tok (y * , x) = α 1 E r [α 2 Tok (y, x) +ᾱ 2 MLE (y, x)] +ᾱ 1 (α 2 Tok (y * , x) +ᾱ 2 MLE (y * , x)).", "(17) In our experiments, we use held out validation data to set mixing and temperature parameters.", "Algorithm 1 Sequence-level smoothing algorithm Input: x, y * Output: α seq (x, y * ) Encode x to initialize the RNN Forward y * in the RNN to compute the hidden states h * t Compute the MLE loss MLE(y * , x) for l ∈ {1, .", ".", ".", ", L} do Sample y l ∼ r(|y * ) if Lazy then Compute (y l , x) = − t log p θ (y l t |h * t ) else Forward y l in the RNN to get its hidden states h l t Compute (y l , x) = MLE(y l , x) end if end for α Seq (x, y * ) =ᾱ MLE(y * , x) + α L l (y l , x) Lazy sequence smoothing.", "Although sequencelevel smoothing is computationally efficient compared to reinforcement learning approaches (Ranzato et al., 2016; Rennie et al., 2017) , it is slower compared to MLE.", "In particular, we need to forward each of the samples y l through the RNN in teacher-forcing mode so as to compute its hidden states h l t , which are used to compute the sequence MLE loss as MLE (y l , x) = − T t=1 ln p θ (y l t |h l t ).", "(18) To speed up training, and since we already forward the ground truth sequence in the RNN to evaluate the MLE part of α Seq (y * , x), we propose to use the same hidden states h * t to compute both the MLE and the sequence-level smoothed loss.", "In this case: lazy (y l , x) = − T t=1 ln p θ (y l t |h * t ) (19) In this manner, we only have a single instead of L + 1 forwards-passes in the RNN.", "We provide the pseudo-code for training in Algorithm 1.", "Experimental evaluation In this section, we compare sequence prediction models trained with maximum likelihood (MLE) with our token and sequence-level loss smoothing on two different tasks: image captioning and machine translation.", "4.1 Image captioning 4.1.1 Experimental setup.", "We use the MS-COCO datatset (Lin et al., 2014) , which consists of 82k training images each annotated with five captions.", "We use the standard splits of Karpathy and Li (2015) , with 5k images for validation, and 5k for test.", "The test set results are generated via beam search (beam size 3) and are evaluated with the MS-COCO captioning evaluation tool.", "We report CIDER and BLEU scores on this internal test set.", "We also report results obtained on the official MS-COCO server that additionally measures METEOR (Denkowski and Lavie, 2014) and ROUGE-L (Lin, 2004) .", "We experiment with both non-attentive LSTMs and the ResNet baseline of the stateof-the-art top-down attention (Anderson et al., 2017) .", "The MS-COCO vocabulary consists of 9,800 words that occur at least 5 times in the training set.", "Additional details and hyperparameters can be found in Appendix B.1.", "Results and discussion Restricted vocabulary sampling In this section, we evaluate the impact of the vocabulary subset from which we sample the modified sentences for sequence-level smoothing.", "We experiment with two rewards: CIDER , which scores w.r.t.", "all five available reference sentences, and Hamming distance reward taking only a single reference into account.", "For each reward we train our (Seq) models with each of the three subsets detailed previously in Section 3.2, Restricted vocabulary sampling.", "From the results in Table 1 we note that for the inattentive models, sampling from V ref s or V batch has a better performance than sampling from the full vocabulary on all metrics.", "In fact, using these subsets introduces a useful bias to the model and improves performance.", "This improvement is most notable using the CIDER reward that scores candidate sequences w.r.t.", "to multiple references, which stabilizes the scoring of the candidates.", "With an attentive decoder, no matter the reward, re-sampling sentences with words from V ref rather than the full vocabulary V is better for both reward functions, and all metrics.", "Additional experimental results, presented in Appendix B.2, obtained with a BLEU-4 reward, in its single and (Xu et al., 2015) 70.", "Overall For reference, we include in Table 1 baseline results obtained using MLE, and our implementation of MLE with entropy regularization (MLE+γH) (Pereyra et al., 2017) , as well as the RAML approach of Norouzi et al.", "(2016) which corresponds to sequence-level smoothing based on the Hamming reward and sampling replacements from the full vocabulary (Seq, Hamming, V) We observe that entropy smoothing is not able to improve performance much over MLE for the model without attention, and even deteriorates for the attention model.", "We improve upon RAML by choosing an adequate subset of vocabulary for substitutions.", "We also report the performances of token-level smoothing, where the promotion of rare tokens boosted the scores in both attentive and nonattentive models.", "For sequence-level smoothing, choosing a taskrelevant reward with importance sampling yielded better results than plain Hamming distance.", "Moreover, we used the two smoothing schemes (Tok-Seq) and achieved the best results with CIDER as a reward for sequence-level smoothing combined with a token-level smoothing that promotes rare tokens improving CIDER from 93.59 (MLE) to 99.92 for the model without attention, and improving from 101.63 to 103.81 with attention.", "Qualitative results.", "In Figure 1 we showcase captions obtained with MLE and our three variants of smoothing i.e.", "token-level (Tok), sequencelevel (Seq) and the combination (Tok-Seq).", "We note that the sequence-level smoothing tend to generate lengthy captions overall, which is maintained in the combination.", "On the other hand, the token-level smoothing allows for a better recognition of objects in the image that stems from the robust training of the classifier e.g.", "the 'cement block' in the top right image or the carrots in the bottom right.", "More examples are available in Appendix B.4 Comparison to the state of the art.", "We compare our model to state-of-the-art systems on the MS-COCO evaluation server in Table 2 .", "We submitted a single model (Tok-Seq, CIDER , V ref s ) as well as an ensemble of five models with different initializations trained on the training set plus 35k images from the dev set (a total of 117k images) to the MS-COCO server.", "The three best results on the server (Rennie et al., 2017; Yao et al., 2017; Anderson et al., 2017) are trained in two stages where they first train using MLE, before switching to policy gradient methods based on CIDEr.", "Anderson et al.", "(2017) reported an increase of 5.8% of CIDER on the test split after the CIDER optimization.", "Moreover, Yao et al.", "(2017) uses additional information about image regions to train the attributes classifiers, while Anderson et al.", "(2017) pre-trains its bottom-up attention model on the Visual Genome dataset (Krishna et al., 2017) .", "Lu et al.", "(2017) ; Yao et al.", "(2017) use the same CNN encoder as ours (ResNet-152), Yang et al., 2016) use Inception-v3 (Szegedy et al., 2016) (2017) use Resnet-101, both of which have similar performances to ResNet-152 on ImageNet classification (Canziani et al., 2016) .", "Machine translation Experimental setup.", "For this task we validate the effectiveness of our approaches on two different datasets.", "The first is WMT'14 English to French, in its filtered version, with 12M sentence pairs obtained after dynamically selecting a \"clean\" subset of 348M words out of the original \"noisy\" 850M words .", "The second benchmark is IWSLT'14 German to English consisting of around 150k pairs for training.", "In all our experiments we use the attentive model of The hyperparameters of each of these models as well as any additional pre-processing can be found in Appendix C.1 To assess the translation quality we report the BLEU-4 metric.", "We present our results in Table 3 .", "On both benchmarks, we improve on both MLE and RAML approach of Norouzi et al.", "(2016) (Seq, Hamming, V): using the smaller batch-vocabulary for replacement improves results, and using importance sampling based on BLEU-4 further boosts results.", "In this case, unlike in the captioning experiment, token-level smoothing brings smaller improvements.", "The combination of both smoothing approaches gives best results, similar to what was observed for image captioning, improving the MLE BLEU-4 from 30.03 to 31.39 on WMT'14 and from 27.55 to 28.74 on IWSLT'14.", "The outputs of our best model are compared to the MLE in some examples showcased in Appendix C. Results and analysis Conclusion We investigated the use of loss smoothing approaches to improve over maximum likelihood estimation of RNN language models.", "We generalized the sequence-level smoothing RAML approach of Norouzi et al.", "(2016) to the tokenlevel by smoothing the ground-truth target across semantically similar tokens.", "For the sequencelevel, which is computationally expensive, we introduced an efficient \"lazy\" evaluation scheme, and introduced an improved re-sampling strategy.", "Experimental evaluation on image captioning and machine translation demonstrates the complementarity of sequence-level and token-level loss smoothing, improving over both the maximum likelihood and RAML." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4", "4.1.2", "4.2.1", "5" ], "paper_header_content": [ "Introduction", "Related work", "Loss smoothing for RNN training", "Maximum likelihood RNN training", "Sequence-level loss smoothing", "Token-level loss smoothing", "Combining losses", "Experimental evaluation", "Results and discussion", "Experimental setup.", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180 ], "paper_content_text": [ "Most previous supervised event extraction methods have relied on features derived from manual annotations, and thus cannot be applied to new event types without extra annotation effort.", "We take a fresh look at event extraction and model it as a generic grounding problem: mapping each event mention to a specific type in a target event ontology.", "We design a transferable architecture of structural and compositional neural networks to jointly represent and map event mentions and types into a shared semantic space.", "Based on this new framework, we can select, for each event mention, the event type which is semantically closest in this space as its type.", "By leveraging manual annotations available for a small set of existing event types, our framework can be applied to new unseen event types without additional manual annotations.", "When tested on 23 unseen event types, this zeroshot framework, without manual annotations, achieves performance comparable to a supervised model trained from 3,000 sentences annotated with 500 event mentions.", "1 Introduction The goal of event extraction is to identify event triggers and their arguments in unstructured text data, and then to assign an event type to each trigger and a semantic role to each argument.", "An example is shown in Figure 1 .", "Traditional supervised methods have typically modeled this task of event 1 The programs are publicly available for research purpose at: https://github.com/wilburOne/ZeroShotEvent extraction as a classification problem, by assigning event triggers to event types from a pre-defined fixed set.", "These methods rely heavily on manual annotations and features specific to each event type, and thus are not easily adapted to new event types without extra annotation effort.", "Handling new event types may even entail starting over, without being able to re-use annotations from previous event types.", "To make event extraction effective as new realworld scenarios emerge, we take a look at this task from the perspective of zero-shot learning, ZSL Norouzi et al., 2013; Socher et al., 2013a) .", "ZSL, as a type of transfer learning, makes use of separate, pre-existing classifiers to build a semantic, cross-concept space that maps between their respective classes.", "The resulting shared semantic space then allows for building a novel \"zero-shot\" classifier, i,e,, requiring no (zero) additional training examples, to handle unseen cases.", "We observe that each event mention has a structure consisting of a candidate trigger and arguments, with corresponding predefined name labels for the event type and argument roles.", "We propose to enrich the semantic representations of each event mention and event type with rich structures, and determine the type based on the semantic similarity between an event mention and each event type defined in a target ontology.", "Let's consider two example sentences: E1.", "The Government of China has ruled Tibet since 1951 after dispatching troops to the Himalayan region in 1950.", "E2.", "Iranian state television stated that the conflict between the Iranian police and the drug smugglers took place near the town of mirjaveh.", "In E1, as also diagrammed in Figure 1 , dis- patching is the trigger for the event mention of type Transport Person and in E2, conflict is the trigger for the event mention of type Attack.", "We make use of Abstract Meaning Representations (AMR) (Banarescu et al., 2013) to identify the candidate arguments and construct event mention structures as shown in Figure 2 (top).", "Figure 2 (bottom) also shows event type structures defined in the Automatic Content Extraction (ACE) guideline.", "2 We can see that a trigger and its event type name usually have some shared meaning.", "Furthermore, their structures also tend to be similar: a Transport Person event typically involves a Person as its patient role, while an Attack event involves a Person or Location as an Attacker.", "This observation matches the theory by Pustejovsky (1991) : \"the semantics of an event structure can be generalized and mapped to event mention structures in a systematic and predictable way\".", "Inspired by this theory, for the first time, we model event extraction as a generic grounding problem, by mapping each mention to its semantically closest event type.", "Given an event ontology, 2 https://en.wikipedia.org/wiki/Automatic content extraction where each event type structure is well-defined, we will refer to the event types for which we have annotated event mentions as seen types, while those without annotations as unseen types.", "Our goal is to learn a generic mapping function independent of event types, which can be trained from annotations for a limited number of seen event types and then used for any new unseen event types.", "We design a transferable neural architecture, which jointly learns and maps the structural representations of event mentions and types into a shared semantic space, by minimizing the distance between each event mention and its corresponding type.", "For event mentions with unseen types, their structures will be projected into the same semantic space using the same framework and assigned types with top-ranked similarity values.", "To summarize, to apply our new zero-shot transfer learning framework to any new unseen event types, we only need (1) a structured definition of the unseen event type (its type name along with role names for its arguments, from the event ontology); and (2) some annotations for one or a few seen event types.", "Without requiring any additional manual annotations for the new unseen types, our ZSL framework achieves performance comparable to supervised methods trained from a substantial amount of training data for the same types.", "Approach Overview Briefly here, we overview the phases involved in building our framework's shared semantic space that, in turn, is the basis for the ZSL framework.", "Given a sentence s, we start by identifying candidate triggers and arguments based on AMR parsing (Wang et al., 2015b) .", "For the example shown in Figure 1 , we identify dispatching as a trigger, and its candidate arguments: China, troops, Himalayan and 1950.", "The details will be described in Section 3.", "After this identification phase, we use our new neural architecture, as depicted in Figure 3 , to classify triggers into event types.", "(The classification of arguments into roles follows the same pipeline.)", "For each trigger t, e.g., dispatch-01, we determine its type by comparing its semantic representation with that of any event type in the event ontology.", "In order to incorporate the contexts into the semantic representation of t, we build a structure S t using AMR as shown in Figure 3.", "Each structure is composed of a set of tuples, e.g, dispatch-01, :ARG0, China .", "We use a matrix to represent each AMR relation, composing its semantics with two concepts for each tuple (in Section 4), and feed all tuple representations into a CNN to generate a dense vector representation V St for the event mention structure (in Section 5.1).", "Given a target event ontology, for each type y, e.g., Transport Person, we construct a type structure S y consisting of its predefined roles, and use a tensor to denote the implicit relation between any type and argument role.", "We compose the semantics of type and argument role with the tensor for each tuple, e.g., Transport Person, Destination (in Section 4).", "Then we generate the event type structure representation V Sy using the same CNN (in Section 5.1).", "By minimizing the semantic distance between dispatch-01 and Trans-port Person using their dense vectors, V St and V Sy respectively, we jointly map the representations of event mention and event types into a shared semantic space, where each mention is closest to its annotated type.", "After training that completes the construction of the semantic space, the compositional functions and CNNs are then used to project any new event mention (e.g., donate-01) into the semantic space and find its closest event type (e.g., Donation) (in Section 5.3).", "In the next sections we will elaborate each step in great detail.", "Trigger and Argument Identification Similar to , we identify candidate triggers and arguments based on AMR Parsing (Wang et al., 2015b ) and apply the same word sense disambiguation (WSD) tool (Zhong and Ng, 2010) to disambiguate word senses and link each sense to OntoNotes, as shown in Figure 1 .", "Given a sentence, we consider all noun and verb concepts that can be mapped to OntoNotes senses by WSD as candidate event triggers.", "In addition, the concepts that can be matched with verbs or nominal lexical units in FrameNet (Baker et al., 1998) are also considered as candidate triggers.", "For each candidate trigger, we consider any concepts that are involved in a subset of AMR rela-tions as candidate arguments 3 .", "We manually select this subset of AMR relations that are useful for capturing generic relations between event triggers and arguments, as shown in Table 1 .", "Categories Relations Core roles ARG0, ARG1, ARG2, ARG3, ARG4 Non-core roles mod, location, instrument, poss, manner, topic, medium, prep-X Temporal year, duration, decade, weekday, time Spatial destination, path, location Trigger and Type Structure Composition As Figure 3 shows, for each candidate trigger t, we construct its event mention structure S t based on its candidate arguments and AMR parsing.", "For each type y in the target event ontology, we construct a structure S y by including its pre-defined roles and using its type as the root.", "Each S t or S y is composed of a collection of tuples.", "For each event mention structure, a tuple consists of two AMR concepts and an AMR relation.", "For each event type structure, a tuple consists of a type name and an argument role name.", "Next we will describe how to compose semantic representations for event mention and event type respectively based on these structures.", "Event Mention Structure For each tuple u = w 1 , λ, w 2 in an event mention structure, we use a matrix to represent each AMR relation λ, and compose the semantics of λ between two concepts w 1 and w 2 as: V u = [V w 1 ; V w 2 ] = f ([V w 1 ; V w 2 ] · M λ ) where V w 1 , V w 2 ∈ R d are the vector representations of words w 1 and w 2 .", "d is the dimension size of each word vector.", "[ ; ] denotes the concatenation of two vectors.", "M λ ∈ R 2d×2d is the matrix representation for AMR relation λ. V u is the composition representation of tuple u, which consists of two updated vector representations V w 1 , V w 2 for w 1 and w 2 by incorporating the semantics of λ.", "Event Type Structure For each tuple u = y, r in an event type structure, where y denotes the event type and r denotes an argument role, following Socher et al.", "(2013b) , we assume an implicit relation exists between any pair of type and argument, and use a single and powerful tensor to represent the implicit relation: V u = [V y ; V r ] = f ([V y ; V r ] T · U [1:2d] · [V y ; V r ]) where V y and V r are vector representations for y and r. U [1:2d] ∈ R 2d×2d×2d is a 3-order tensor.", "V u is the composition representation of tuple u , which consists of two updated vector representations V y , V r for y and r by incorporating the semantics of their implicit relation U [1:2d] .", "Trigger and Argument Classification Trigger Classification for Seen Types Both event mention and event type structures are relatively simple and can be represented with a set of tuples.", "CNNs have been demonstrated effective at capturing sentence level information by aggregating compositional n-gram representations.", "In order to generate structure-level representations, we use CNN to learn to aggregate all edge and tuple representations.", "Input layer is a sequence of tuples, where the order of tuples is from top to bottom in the structure.", "Each tuple is represented by a d × 2 dimensional vector, thus each mention structure and each type structure are represented as a feature map of dimensionality d × 2h * and d × 2p * respectively, where h * and p * are the maximal number of tuples for event mention and type structures.", "We use zero-padding to the right to make the volume of all input structures consistent.", "Convolution layer Take S t with h * tuples: u 1 , u 2 , ..., u h * as an example.", "The input matrix of S t is a feature map of dimensionality d × 2h * .", "We make c i as the concatenated embeddings of n continuous columns from the feature map, where n is the filter width and 0 < i < 2h * + n. A convolution operation involves a filter W ∈ R nd , which is applied to each sliding window c i : c i = tanh(W · c i + b) where c i is the new feature representation, and b ∈ R d is a biased vector.", "We set filter width as 2 and stride as 2 to make the convolution function operate on each tuple with two input columns.", "Max-Pooling: All tuple representations c i are used to generate the representation of the input sequence by max-pooling.", "Learning: For each event mention t, we name the correct type as positive and all the other types in the target event ontology as negative.", "To train the composition functions and CNN, we first consider the following hinge ranking loss: L 1 (t, y) = j∈Y, j =y max{0, m − C t,y + C t,j } C t,y = cos([V t ; V St ], [V y ; V Sy ]) where y is the positive event type for t. Y is the type set of the target event ontology.", "[V t ; V St ] denotes the concatenation of representations of t and S t .", "j is a negative event type for t from Y .", "m is a margin.", "C t,y denotes the cosine similarity between t and y.", "The hinge loss is commonly used in zero-shot visual object classification task.", "However, it tends to overfit the seen types in our experiments.", "While clever data augmentation can help alleviate overfitting, we design two strategies: (1) we add \"negative\" event mentions into the training process.", "Here a \"negative\" event mention means that the mention has no positive event type among all seen types, namely it belongs to Other.", "(2) we design a new loss function as follows: where Y is the type set of the event ontology.", "Y is the seen type set.", "y is the annotated type.", "y is the type which ranks the highest among all event types for event mention t, while t belongs to Other.", "By minimizing L d 1 , we can learn the optimized model which can compose structure representations and map both event mention and types into a shared semantic space, where the positive type ranks the highest for each mention.", "Argument Classification for Seen Types For each mention, we map each candidate argument to a specific role based on the semantic similarity of the argument path.", "Take E1 as an example.", "China is matched to Agent based on the semantic similarity between dispatch-01→ :ARG0→ China and Transport-Person→Agent.", "Given a trigger t and a candidate argument a, we first extract a path S a = (u 1 , u 2 , ..., u p ), which connects t and a and consists of p tuples.", "Each predefined role r is also represented as a structure by incorporating the event type, S r = y, r .", "We apply the same framework to take the sequence of tuples contained in S a and S r into a weightsharing CNN to rank all possible roles for a. where R y and R Y are the set of argument roles which are predefined for trigger type y and all seen types Y .", "r is the annotated role and r is the argument role which ranks the highest for a when a or y is annotated as Other.", "In our experiments, we sample various size of \"negative\" training data for trigger and argument labeling respectively.", "In the following section, we describe how the negative training instances are generated.", "Zero-Shot Classification for Unseen Types During test, given a new event mention t , we compute its mention structure representation for S t and all event type structure representations for S Y = {S y 1 , S y 2 , ..., S yn } using the same parameters trained from seen types.", "Then we rank all event types based on their similarity scores with mention t .", "The top ranked prediction for t from the event type set, denoted as y(t , 1), is given by: y(t , 1) = arg max y∈Y cos([V t ; V S t ], [V y ; V Sy ]) Moreover, y(t , k) denotes the k th most probable event type predicted for t .", "We will investigate the event extraction performance based on the topk predicted event types.", "After determining the type y for mention t , for each candidate argument, we adopt the same ranking function to find the most appropriate role from the role set defined for y .", "Experiments Hyper-Parameters We used the English Wikipedia dump to learn trigger sense and argument embeddings based on the Continuous Skip-gram model .", "Table 2 We first used the ACE event schema 4 as our target event ontology and assumed the boundaries of triggers and arguments as given.", "Of the 33 ACE event types, we selected the top-N most popular event types from ACE05 data as \"seen\" types, and used 90% event annotations of these for training and 10% for development.", "We set N as 1, 3, 5, 10 respectively.", "We tested the zero-shot classification performance on the annotations for the remaining 23 unseen types.", "Table 3 shows the types that we selected for training in each experiment setting.", "The negative event mentions and arguments that belong to Other were sampled from the output of the system developed by based on ACE05 training sentences, which groups all candidate triggers and arguments into clusters based on semantic representations and assigns a type/role name to each cluster.", "We sampled the negative event mentions from the clusters (e.g., Build, Threaten) which do not map to ACE event types.", "We sampled the negative arguments from the arguments of negative event mentions.", "Table 4 shows the statistics of the training, development and testing data sets.", "To show the effectiveness of structural similarity in our approach, we designed a baseline, WSD-4 ACE event schema specification is at: https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/englishevents-guidelines-v5.4.3.pdf Embedding, which directly maps event mentions and arguments to their candidate types and roles using our pre-trained word sense embeddings.", "Table 5 makes the contrast clear: structural similarity (our approach) is much more effective than lexical similarity (baseline) for both trigger and argument classification.", "Also, as the number of seen types in training increases, the performance of the transfer model improves.", "We further evaluated the performance of our transfer approach on similar and distinct unseen types.", "The 33 subtypes defined in ACE fall within 8 coarse-grained main types, such as Life and Justice.", "Each subtype belongs to one main type.", "Subtypes that belong to the same main type tend to have similar structures.", "For example, Trial-Hearing and Charge-Indict have the same set of argument roles.", "For training our transfer model, we selected 4 subtypes of Justice: Arrest-Jail, Convict, Charge-Indict, Execute.", "For testing, we selected 3 other subtypes of Justice: Sentence, Appeal, Release-Parole.", "Additionally, we selected one subtype from each of the other seven main types for comparison.", "Table 6 shows that, when testing on a new unseen type, the more similar it is to the seen types, the better performance is achieved.", "ACE Event Identification & Classification The ACE2005 corpus includes the richest event annotations currently available for 33 types.", "However, in real-world scenarios, there may be thousands of event types of interest.", "To enrich the target event ontology and assess our transferable neural architecture on a large number of unseen types, when trained on limited annotations of seen types, we manually constructed a new event ontology which combined 33 ACE event types and argument roles, and 1,161 frames from FrameNet, except for the most generic frames such as Entity and Locale.", "Some ACE event types were easily aligned to frames, e.g., Die aligned to Death.", "Some frames were instead more accurately treated as inheritors of ACE types, such as Suicide-Attack, which inherits from Attack.", "We manually mapped the selected frames to ACE types.", "We then compared our approach with the following state-of-the-art supervised methods: • LSTM: A long short-term memory neural network (Hochreiter and Schmidhuber, 1997) based on distributed semantic features, similar .", "• Joint: A structured perceptron model based on symbolic semantic features (Li et al., 2013) .", "For our approach, we followed the experiment setting D in the previous section, using the same training and development data sets for the 10 seen types, but targeted all 1,194 event types in our new event ontology, instead of just the 33 ACE event types.", "For evaluation, we sampled 150 sentences from the remaining ACE05 data, including 129 annotated event mentions for the 23 unseen types.", "For both LSTM and Joint approaches, we used the entire ACE05 annotated data for 33 ACE event types for training except for the held-out 150 evaluation sentences.", "We first identified the candidate triggers and arguments, then mapped each of these to the target event ontology.", "We evaluated our model on their extracting of event mentions which were classified into 23 testing ACE types.", "Table 7 shows the per-formance.", "To further demonstrate the effectiveness of zero-shot learning in our framework and its impact in saving human annotation effort, we used the supervised LSTM approach for comparison.", "The training data of LSTM contained 3,464 sentences with 905 annotated event mentions for the 23 unseen event types.", "We divided these event annotations into 10 subsets and successively added one subset at a time (10% of annotations) into the training data of LSTM.", "Figure 4 shows the LSTM learning curve.", "By contrast, without any annotated mentions on the 23 unseen test event types in its training set, our transfer learning approach achieved performance comparable to that of the LSTM, which was trained on 3,000 sentences 5 with 500 annotated event mentions.", "Table 7 : Event Trigger and Argument Extraction Performance (%) on Unseen ACE Types.", "Impact of AMR Recall that we used AMR parsing output to identify triggers and arguments in constructing event structures.", "To assess the impact of the AMR parser (Wang et al., 2015a) on event extraction, we chose a subset of the ERE (Entity, Relation, Event) corpus (Song et al., 2015) which has ground-truth AMR annotations.", "This subset contains 304 documents with 1,022 annotated event mentions of 40 types.", "We selected the top-6 most popular event types (Arrest-Jail, Execute, Die, Meet, Sentence, Charge-Indict) with manual annotations of 548 event mentions as seen types.", "We sampled 500 negative event mentions from distinct types of clusters generated from the system based on ERE training sentences.", "We combined the annotated events for seen types and the negative event mentions, and used 90% for training and 10% for development.", "For evaluation, we selected 200 sentences from the remaining ERE subset, which contains 128 Attack event mentions and 40 Convict event mentions.", "Table 8 shows the event extraction performances based on groundtruth AMR and system AMR respectively.", "We also compared AMR analyses with Semantic Role Labeling (SRL) output (Palmer et al., 2010) by keeping only the core roles (e.g., :ARG0, :ARG1) from AMR annotations.", "As Table 8 shows, comparing the full AMR (top row) to this SRL proxy (middle row), the fine-grained AMR semantic relations such as :location, :instrument appear to be more informative for inferring event argument role labeling.", "Method Trigger Labeling Related Work Most previous event extraction methods have been based on supervised learning, using either symbolic features (Ji and Grishman, 2008; Miwa et al., 2009; Liao and Grishman, 2010; Liu et al., 2010; Hong et al., 2011; McClosky et al., 2011; Riedel and McCallum, 2011; Li et al., 2013; or distributional features (Chen et al., 2015; Nguyen and Grishman, 2015; Nguyen et al., 2016) derived from a large amount of training data, and treating event types and argument role labels as symbols.", "These approaches can achieve high quality for known event types, but cannot be applied to new types without additional annotation effort.", "In contrast, we provide a new angle on event extraction, modeling it as a generic grounding task by taking advantage of rich semantics of event types.", "Some other IE paradigms such as Open IE (Etzioni et al., 2005; Banko et al., 2007 Banko et al., , 2008 Etzioni et al., 2011; Ritter et al., 2012) , Preemptive IE (Shinyama and Sekine, 2006) , Ondemand IE (Sekine, 2006) , Liberal IE (Huang et al., , 2017 , and semantic frame-based event discovery (Kim et al., 2013) can discover many events without pre-defined event schema.", "These paradigms however rely on information redundancy, and so they are not effective when the input data only consists of a few sentences.", "Our work can discover events from any size of input corpus and can also be complementary with these paradigms.", "Our event extraction paradigm is similar to the task of entity linking (Ji and Grishman, 2011) in semantic mapping.", "However, entity linking aims to map entity mentions to the same concept, while our framework maps each event mention to a specific category.", "In addition, Bronstein et al.", "(2015) and Peng et al.", "(2016) employ an eventindependent similarity-based function for event trigger detection, which follows few-shot learning setting and requires some trigger examples as seeds.", "Lu and Roth (2012) design a structure pref-erence modeling framework, which can automatically predict argument roles without any annotated data, but it relies on manually constructed patterns.", "Zero-Shot learning has been widely applied in visual object classification Norouzi et al., 2013; Socher et al., 2013a; Chen et al., 2017; Li et al., 2017; Xian et al., 2017; Changpinyo et al., 2017) , fine-grained name tagging (Ma et al., 2016; Qu et al., 2016) , relation extraction (Verga et al., 2016; Levy et al., 2017) , semantic parsing (Bapna et al., 2017) and domain adaptation (Romera-Paredes and Torr, 2015; Kodirov et al., 2015; Peng et al., 2017) .", "In contrast to these tasks, for our case, the number of seen types in event extraction with manual annotations is quite limited.", "The most popular event schemas, such as ACE, define 33 event types while most visual object training sets contain more than 1,000 types.", "Therefore, methods proposed for zero-shot visual-object classification cannot be directly applied to event extraction due to overfitting.", "In this work, we designed a new loss function by creating \"negative\" training instances to avoid overfitting.", "Conclusions and Future Work In this work, we take a fresh look at the event extraction task and model it as a generic grounding problem.", "We propose a transferable neural architecture, which leverages existing humanconstructed event schemas and manual annotations for a small set of seen types, and transfers the knowledge from the existing types to the extraction of unseen types, to improve the scalability of event extraction as well as to save human effort.", "To the best of our knowledge, this work is the first time that zero-shot learning has been applied to event extraction.", "Without any annotation, our approach can achieve performance comparable to state-of-the-art supervised models trained on a large amount of labeled data.", "In the future, we will extend this framework to other Information Extraction problems." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "5.1", "5.2", "5.3", "6.1", "6.3", "6.4", "7", "8" ], "paper_header_content": [ "Introduction", "Approach Overview", "Trigger and Argument Identification", "Trigger and Type Structure Composition", "Trigger Classification for Seen Types", "Argument Classification for Seen Types", "Zero-Shot Classification for Unseen Types", "Hyper-Parameters", "ACE Event Identification & Classification", "Impact of AMR", "Related Work", "Conclusions and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168 ], "paper_content_text": [ "Introduction Learning policies for task-completion dialogue is often formulated as a reinforcement learning (RL) problem (Young et al., 2013; Levin et al., 1997) .", "However, applying RL to real-world dialogue systems can be challenging, due to the constraint that an RL learner needs an environment to operate in.", "In the dialogue setting, this requires a dialogue agent to interact with real users and adjust its policy in an online fashion, as illustrated in Figure 1(a) .", "Unlike simulation-based games such as Atari games (Mnih et al., 2015) and AlphaGo (Silver et al., 2016a (Silver et al., , 2017 where RL has made its greatest strides, task-completion dialogue systems may incur significant real-world cost in case of failure.", "Thus, except for very simple tasks (Singh et al., 2002; Gašić et al., 2010 Gašić et al., , 2011 Pietquin et al., 2011; Li et al., 2016a; Su et al., 2016b) , RL is too expensive to be applied to real users to train dialogue agents from scratch.", "One strategy is to convert human-interacting dialogue to a simulation problem (similar to Atari games), by building a user simulator using human conversational data (Schatzmann et al., 2007; Li et al., 2016b) .", "In this way, the dialogue agent can learn its policy by interacting with the simulator instead of real users (Figure 1(b) ).", "The simulator, in theory, does not incur any real-world cost and can provide unlimited simulated experience for reinforcement learning.", "The dialogue agent trained with such a user simulator can then be deployed to real users and further enhanced by only a small number of human interactions.", "Most of recent studies in this area have adopted this strategy (Su et al., 2016a; Lipton et al., 2016; Zhao and Eskenazi, 2016; Williams et al., 2017; Dhingra et al., 2017; Liu and Lane, 2017; Peng et al., 2017b; Budzianowski et al., 2017; Peng et al., 2017a) .", "However, user simulators usually lack the conversational complexity of human interlocutors, and the trained agent is inevitably affected by biases in the design of the simulator.", "Dhingra et al.", "(2017) demonstrated a significant discrepancy in a simulator-trained dialogue agent when evaluated with simulators and with real users.", "Even more challenging is the fact that there is no universally accepted metric to evaluate a user simulator (Pietquin and Hastie, 2013 Figure 1 : Three strategies of learning task-completion dialogue policies via RL.", "controversial whether training task-completion dialogue agent via simulated users is a valid approach.", "We propose a new strategy of learning dialogue policy by interacting with real users.", "Compared to previous works (Singh et al., 2002; Li et al., 2016a; Su et al., 2016b; Papangelis, 2012) , our dialogue agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial dialogue tasks.", "Our approach is based on the Dyna-Q framework (Sutton, 1990) where planning is integrated into policy learning for task-completion dialogue.", "Specifically, we incorporate a model of the environment, referred to as the world model, into the dialogue agent, which simulates the environment and generates simulated user experience.", "During the dialogue policy learning, real user experience plays two pivotal roles: first, it can be used to improve the world model and make it behave more like real users, via supervised learning; second, it can also be used to directly improve the dialogue policy via RL.", "The former is referred to as world model learning, and the latter direct reinforcement learning.", "Dialogue policy can be improved either using real experience directly (i.e., direct reinforcement learning) or via the world model indirectly (referred to as planning or indirect reinforcement learning).", "The interaction between world model learning, direct reinforcement learning and planning is illustrated in Figure 1 (c), following the Dyna-Q framework (Sutton, 1990) .", "The original papers on Dyna-Q and most its early extensions used tabular methods for both planning and learning (Singh, 1992; Peng and Williams, 1993; Moore and Atkeson, 1993; Kuvayev and Sutton, 1996) .", "This table-lookup representation limits its application to small problems only.", "Sutton et al.", "(2012) extends the Dyna architecture to linear function approximation, making it applicable to larger problems.", "In the dialogue setting, we are dealing with a much larger action-state space.", "Inspired by Mnih et al.", "(2015) , we propose Deep Dyna-Q (DDQ) by combining Dyna-Q with deep learning approaches to representing the state-action space by neural networks (NN).", "By employing the world model for planning, the DDQ method can be viewed as a model-based RL approach, which has drawn growing interest in the research community.", "However, most model-based RL methods (Tamar et al., 2016; Silver et al., 2016b; Gu et al., 2016; Racanière et al., 2017) are developed for simulation-based, synthetic problems (e.g., games), but not for human-in-the-loop, real-world problems.", "To these ends, our main contributions in this work are two-fold: • We present Deep Dyna-Q, which to the best of our knowledge is the first deep RL framework that incorporates planning for taskcompletion dialogue policy learning.", "• We demonstrate that a task-completion dialogue agent can efficiently adapt its policy on the fly, by interacting with real users via RL.", "This results in a significant improvement in success rate on a nontrivial task.", "the next action 2 based on the current state; (4) a model-based natural language generation (NLG) module for converting dialogue actions to natural language response (Wen et al.", "); and (5) a world model for generating simulated user actions and simulated rewards.", "As illustrated in Figure 1 (c), starting with an initial dialogue policy and an initial world model (both trained with pre-collected human conversational data), the training of the DDQ agent consists of three processes: (1) direct reinforcement learning, where the agent interacts with a real user, collects real experience and improves the dialogue policy; (2) world model learning, where the world model is learned and refined using real experience; and (3) planning, where the agent improves the dialogue policy using simulated experience.", "Although these three processes conceptually can occur simultaneously in the DDQ agent, we implement an iterative training procedure, as shown in Algorithm 1, where we specify the order in which they occur within each iteration.", "In what follows, we will describe these processes in details.", "Direct Reinforcement Learning In this process (lines 5-18 in Algorithm 1) we use the DQN method (Mnih et al., 2015) to improve the dialogue policy based on real experience.", "We consider task-completion dialogue as a Markov Decision Process (MDP), where the agent inter-acts with a user in a sequence of actions to accomplish a user goal.", "In each step, the agent observes the dialogue state s, and chooses the action a to execute, using an -greedy policy that selects a random action with probability or otherwise follows the greedy policy a = argmax a Q(s, a ; θ Q ).", "Q(s, a; θ Q ) which is the approximated value function, implemented as a Multi-Layer Perceptron (MLP) parameterized by θ Q .", "The agent then receives reward 3 r, observes next user response a u , and updates the state to s .", "Finally, we store the experience (s, a, r, a u , s ) in the replay buffer D u .", "The cycle continues until the dialogue terminates.", "We improve the value function Q(s, a; θ Q ) by adjusting θ Q to minimize the mean-squared loss function, defined as follows: L(θ Q ) = E (s,a,r,s )∼D u [(y i − Q(s, a; θ Q )) 2 ] y i = r + γ max a Q (s , a ; θ Q ) (1) where γ ∈ [0, 1] is a discount factor, and Q (.)", "is the target value function that is only periodically updated (line 42 in Algorithm 1).", "By differentiating the loss function with respect to θ Q , we arrive at the following gradient: ∇ θ Q L(θ Q ) = E (s,a,r,s )∼D u [(r+ γ max a Q (s , a ; θ Q ) − Q(s, a; θ Q )) ∇ θ Q Q(s, a; θ Q )] (2) As shown in lines 16-17 in Algorithm 1, in each iteration, we improve Q(.)", "using minibatch Deep Q-learning.", "Planning In the planning process (lines 23-41 in Algorithm 1), the world model is employed to generate simulated experience that can be used to improve dialogue policy.", "K in line 24 is the number of planning steps that the agent performs per step of direct reinforcement learning.", "If the world model is able to accurately simulate the environment, a big K can be used to speed up the policy learning.", "In DDQ, we use two replay buffers, D u for storing real experience and D s for simulated experience.", "Learning and planning are accomplished Algorithm 1 Deep Dyna-Q for Dialogue Policy Learning Require: N , , K, L, C, Z Ensure: Q(s, a; θQ), M (s, a; θM ) 1: initialize Q(s, a; θQ) and M (s, a; θM ) via pre-training on human conversational data 2: initialize Q (s, a; θ Q ) with θ Q = θQ 3: initialize real experience replay buffer D u using Reply Buffer Spiking (RBS), and simulated experience replay buffer D s as empty 4: for n=1:N do 5: # Direct Reinforcement Learning starts 6: user starts a dialogue with user action a u 7: generate an initial dialogue state s 8: while s is not a terminal state do 9: with probability select a random action a 10: otherwise select a = argmax a Q(s, a ; θQ) 11: execute a, and observe user response a u and reward r 12: update dialogue state to s 13: store (s, a, r, a u , s ) to D u 14: s = s 15: end while 16: sample random minibatches of (s, a, r, s ) from D u 17: update θQ via Z-step minibatch Q-learning according to Equation (2) 18: # Direct Reinforcement Learning ends 19: # World Model Learning starts 20: sample random minibatches of training samples (s, a, r, a u , s ) from D u 21: update θM via Z-step minibatch SGD of multi-task learning 22: # World Model Learning ends 23: # Planning starts 24: for k=1:K do 25: t = FALSE, l = 0 26: sample a user goal G 27: sample user action a u from G 28: generate an initial dialogue state s 29: while t is FALSE ∧ l ≤ L do 30: with probability select a random action a 31: otherwise select a = argmax a Q(s, a ; θQ) 32: execute a 33: world model responds with a u , r and t 34: update dialogue state to s 35: store (s, a, r, s ) to D s 36: l = l + 1, s = s 37: end while 38: sample random minibatches of (s, a, r, s ) from D s 39: update θQ via Z-step minibatch Q-learning according to Equation (2) 40 : end for 41: # Planning ends 42: every C steps reset θ Q = θQ 43: end for by the same DQN algorithm, operating on real experience in D u for learning and on simulated experience in D s for planning.", "Thus, here we only describe the way the simulated experience is generated.", "Similar to Schatzmann et al.", "(2007) , at the beginning of each dialogue, we uniformly draw a user goal G = (C, R), where C is a set of con-straints and R is a set of requests (line 26 in Algorithm 1).", "For movie-ticket booking dialogues, constraints are typically the name and the date of the movie, the number of tickets to buy, etc.", "Requests can contain these slots as well as the location of the theater, its start time, etc.", "Table 3 presents some sampled user goals and dialogues generated by simulated and real users, respectively.", "The first user action a u (line 27) can be either a request or an inform dialogueact.", "A request, such as request(theater; moviename=batman), consists of a request slot and multiple ( 1) constraint slots, uniformly sampled from R and C, respectively.", "An inform contains constraint slots only.", "The user action can also be converted to natural language via NLG, e.g., \"which theater will show batman?\"", "In each dialogue turn, the world model takes as input the current dialogue state s and the last agent action a (represented as an one-hot vector), and generates user response a u , reward r, and a binary variable t, which indicates whether the dialogue terminates (line 33).", "The generation is accomplished using the world model M (s, a; θ M ), a MLP shown in Figure 3 , as follows: h = tanh(W h (s, a) + b h ) r = W r h + b r a u = softmax(W a h + b a ) t = sigmoid(W t h + b t ) where (s, a) is the concatenation of s and a, and W and b are parameter matrices and vectors, respectively.", "Task-Specific Representation s: state a: agent action a u r t Shared layers Figure 3: The world model architecture.", "World Model Learning In this process (lines 19-22 in Algorithm 1), M (s, a; θ M ) is refined via minibatch SGD using real experience in the replay buffer D u .", "As shown in Figure 3 , M (s, a; θ M ) is a multi-task neural network (Liu et al., 2015) that combines two classification tasks of simulating a u and t, respectively, and one regression task of simulating r. The lower layers are shared across all tasks, while the top layers are task-specific.", "Experiments and Results We evaluate the DDQ method on a movie-ticket booking task in both simulation and human-in-theloop settings.", "Dataset Raw conversational data in the movie-ticket booking scenario was collected via Amazon Mechanical Turk.", "The dataset has been manually labeled based on a schema defined by domain experts, as shown in Table 4 , which consists of 11 dialogue acts and 16 slots.", "In total, the dataset contains 280 annotated dialogues, the average length of which is approximately 11 turns.", "Dialogue Agents for Comparison To benchmark the performance of DDQ, we have developed different versions of task-completion dialogue agents, using variations of Algorithm • The DQN(K) agents are learned by DQN, but with K times more real experiences than the DQN agent.", "DQN(K) is evaluated in the simulation setting only.", "Its performance can be viewed as the upper bound of its DDQ(K) counterpart, assuming that the world model in DDQ(K) perfectly matches real users.", "Implementation Details All the models in these agents (Q(s, a; θ Q ), M (s, a; θ M )) are MLPs with tanh activations.", "Each policy network Q(.)", "has one hidden layer with 80 hidden nodes.", "As shown in Figure 3 , the world model M (.)", "contains two shared hidden layers and three task-specific hidden layers, with 80 nodes in each.", "All the agents are trained by Algorithm 1 with the same set of hyper-parameters.", "-greedy is always applied for exploration.", "We set the discount factor γ = 0.95.", "The buffer sizes of both D u and D s are set to 5000.", "The target value function is updated at the end of each epoch.", "In each epoch, Q(.)", "and M (.)", "are refined using one-step (Z = 1) 16-tupleminibatch update.", "4 In planning, the maximum length of a simulated dialogue is 40 (L = 40).", "In addition, to make the dialogue training efficient, we also applied a variant of imitation learning, called Reply Buffer Spiking (RBS) (Lipton et al., 2016) .", "We built a naive but occasionally successful rule-based agent based on human conversational dataset (line 1 in Algorithm 1), and prefilled the real experience replay buffer D u with 100 dialogues of experience (line 2) before training for all the variants of agents.", "Simulated User Evaluation In this setting the dialogue agents are optimized by interacting with user simulators, instead of real users.", "Thus, the world model is learned to mimic user simulators.", "Although the simulator-trained agents are sub-optimal when applied to real users due to the discrepancy between simulators and real users, the simulation setting allows us to perform a detailed analysis of DDQ without much cost and to reproduce the experimental results easily.", "4 We found in our experiments that setting Z > 1 improves the performance of all agents, but does not change the conclusion of this study: DDQ consistently outperforms DQN by a statistically significant margin.", "Conceptually, the optimal value of Z used in planning is different from that in direct reinforcement learning, and should vary according to the quality of the world model.", "The better the world model is, the more aggressive update (thus bigger Z) is being used in planning.", "We leave it to future work to investigate how to optimize Z for planning in DDQ.", "User Simulator We adapted a publicly available user simulator (Li et al., 2016b) to the taskcompletion dialogue setting.", "During training, the simulator provides the agent with a simulated user response in each dialogue turn and a reward signal at the end of the dialogue.", "A dialogue is considered successful only when a movie ticket is booked successfully and when the information provided by the agent satisfies all the user's constraints.", "At the end of each dialogue, the agent receives a positive reward of 2 * L for success, or a negative reward of −L for failure, where L is the maximum number of turns in each dialogue, and is set to 40 in our experiments.", "Furthermore, in each turn, the agent receives a reward of −1, so that shorter dialogues are encouraged.", "Readers can refer to Appendix B for details on the user simulator.", "Results The main simulation results are reported in Table 1 and Figures 4 and 5.", "For each agent, we report its results in terms of success rate, average reward, and average number of turns (averaged over 5 repetitions of the experiments).", "Results show that the DDQ agents consistently outperform DQN with a statistically significant margin.", "Figure 4 shows the learning curves of different DDQ agents trained using different planning steps.", "Since the training of all RL agents started with RBS using the same rule-based agent, their performance in the first few epochs is very close.", "After that, performance improved for all values of K, but much more rapidly for larger values.", "and DDQ(10) took only 50 epochs.", "Intuitively, the optimal value of K needs to be determined by seeking the best trade-off between the quality of the world model and the amount of simulated experience that is useful for improving the dialogue agent.", "This is a non-trivial optimization problem because both the dialogue agent and the world model are updated constantly during training and the optimal K needs to be adjusted accordingly.", "For example, we find in our experiments that at the early stages of training, it is fine to perform planning aggressively by using large amounts of simulated experience even though they are of low quality, but in the late stages of training where the dialogue agent has been significantly improved, low-quality simulated experience is likely to hurt the performance.", "Thus, in our implementation of Algorithm 1, we use a heuristic 5 to reduce the value of K in the late stages of training (e.g., after 150 epochs in Figure 4 ) to mitigate the negative impact of low-qualify simulated experience.", "We leave it to future work how to optimize the planning step size during DDQ training in a principled way.", "Figure 5 shows that the quality of the world model has a significant impact on the agent's performance.", "The learning curve of DQN(10) indicates the best performance we can expect with a perfect world model.", "With a pre-trained world model, the performance of the DDQ agent improves more rapidly, although eventually, the DDQ and DDQ(rand-init θ M ) agents reach the same success rate after many epochs.", "The world model learning process is crucial to both the efficiency of dialogue policy learning and the final performance of the agent.", "For example, in the early stages (before 60 epochs), the performances of DDQ and DDQ(fixed θ M ) remain very close to each other, but DDQ reaches a success rate almost 5 The heuristic is not presented in Algorithm 1.", "Readers can refer to the released source code for details.", "10% better than DDQ(fixed θ M ) after 400 epochs.", "Human-in-the-Loop Evaluation In this setting, five dialogue agents (i.e., DQN, DDQ(10), DDQ(10, rand-init θ M ), DDQ(5), and DDQ(5, rand-init θ M )) are trained via RL by interacting with real human users.", "In each dialogue session, one of the agents was randomly picked to converse with a user.", "The user was presented with a user goal sampled from the corpus, and was instructed to converse with the agent to complete the task.", "The user had the choice of abandoning the task and ending the dialogue at any time, if she or he believed that the dialogue was unlikely to succeed or simply because the dialogue dragged on for too many turns.", "In such cases, the dialogue session is considered failed.", "At the end of each session, the user was asked to give explicit feedback whether the dialogue succeeded (i.e., whether the movie tickets were booked with all the user constraints satisfied).", "Each learning curve is trained with two runs, with each run generating 150 dialogues (and K * 150 additional simulated dialogues when planning is applied).", "In total, we collected 1500 dialogue sessions for training all five agents.", "The main results are presented in Table 2 and Simulation Sample Real User Sample movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"numberofpeople\":\"2\" \"theater\": \"?\"", "\"moviename\": \"deadpool\" \"starttime\": \"?\"", "\"city\": \"seattle\" \"date\": \"?\"", "} } } movie-ticket booking user goal: { \"request slots\": { \"constraint slots\": { \"ticket\": \"?\"", "\"date\":\"this weekend\" \"theater\": \"?\"", "\"numberofpeople\": \"1\" \"starttime\": \"?\"", "\"moviename\": \"batman\" } } } Table 3 presents four example dialogues produced by two dialogue agents interacting with simulated and human users, respectively.", "The DQN agent, after being trained with 100 dialogues, still behaved like a naive rule-based agent that re-quested information bit by bit in a fixed order.", "When the user did not answer the request explicitly (e.g., usr: which theater is available?", "), the agent failed to respond properly.", "On the other hand, with planning, the DDQ agent trained with 100 real dialogues is much more robust and can complete 50% of user tasks successfully.", "• A larger K leads to more aggressive planning and better results, as shown by DDQ(10) vs. DDQ(5).", "• Pre-training world model with human con-versational data improves the learning efficiency and the agent's performance, as shown by DDQ(5) vs. DDQ(5, rand-init θ M ), and DDQ(10) vs. DDQ(10, rand-init θ M ).", "Conclusion We propose a new strategy for a task-completion dialogue agent to learn its policy by interacting with real users.", "Compared to previous work, our agent learns in a much more efficient way, using only a small number of real user interactions, which amounts to an affordable cost in many nontrivial domains.", "Our strategy is based on the Deep Dyna-Q (DDQ) framework where planning is integrated into dialogue policy learning.", "The effectiveness of DDQ is validated by human-in-theloop experiments, demonstrating that a dialogue agent can efficiently adapt its policy on the fly by interacting with real users via deep RL.", "One interesting topic for future research is exploration in planning.", "We need to deal with the challenge of adapting the world model in a changing environment, as exemplified by the domain extension problem (Lipton et al., 2016) .", "As pointed out by Sutton and Barto (1998) , the general problem here is a particular manifestation of the conflict between exploration and exploitation.", "In a planning context, exploration means trying actions that may improve the world model, whereas exploitation means trying to behave in the optimal way given the current model.", "To this end, we want the agent to explore in the environment, but not so much that the performance would be greatly degraded.", "Additional Authors A Dataset Annotation Schema B User Simulator In the task-completion dialogue setting, the entire conversation is around a user goal implicitly, but the agent knows nothing about the user goal explicitly and its objective is to help the user to accomplish this goal.", "Generally, the definition of user goal contains two parts: • inform slots contain a number of slot-value pairs which serve as constraints from the user.", "• request slots contain a set of slots that user has no information about the values, but wants to get the values from the agent during the conversation.", "ticket is a default slot which always appears in the request slots part of user goal.", "To make the user goal more realistic, we add some constraints in the user goal: slots are split into two groups.", "Some of slots must appear in the user goal, we called these elements as Required slots.", "In the movie-booking scenario, it includes moviename, theater, starttime, date, numberofpeople; the rest slots are Optional slots, for example, theater chain, video format etc.", "We generated the user goals from the labeled dataset mentioned in Section 3.1, using two mechanisms.", "One mechanism is to extract all the slots (known and unknown) from the first user turns (excluding the greeting user turn) in the data, since usually the first turn contains some or all the required information from user.", "The other mechanism is to extract all the slots (known and unknown) that first appear in all the user turns, and then aggregate them into one user goal.", "We dump these user goals into a file as the user-goal database.", "Every time when running a dialogue, we randomly sample one user goal from this user goal database." ] }

{ "paper_header_number": [ "1", "2.1", "2.2", "2.3", "3", "3.1", "3.2", "3.3", "3.4", "4" ], "paper_header_content": [ "Introduction", "Direct Reinforcement Learning", "Planning", "World Model Learning", "Experiments and Results", "Dataset", "Dialogue Agents for Comparison", "Simulated User Evaluation", "Human-in-the-Loop Evaluation", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139 ], "paper_content_text": [ "Introduction In the Japanese-Chinese machine translation task, reordering is the main problem due to substantial differences in sentence structures between these two languages.", "For example, Japanese has a subject-object-verb (SOV) structure, while, Chinese has a subject-verb-object (SVO) structure.", "The pre-ordering technology is one way to handle this problem (Wu, et al., 2011) , but it needs to train a pre-ordering model.", "An hierarchical phrase-based (HPB) model (Chiang, 2005; Chiang, 2007 ) is a powerful method to cover any format of translation pairs by using synchronous context free grammar.", "Hiero grammars can capture complex nested translation relationships to handle reordering.", "However, due to its compromise on the efficiency of rule extraction and decoding, (a) a source language span limit is applied with 10, (b) the number of non-terminals in one rule is set to 2, (c) there is a prohibition of consecutive nonterminals on the source language side of a rule and With a traditional approach, the typical H-PB model fails to capture complex reordering information as shown in Figure 1 .", "By contrast, Fillmore (1968) has proposed case grammar, which is effectively proved and originally used in rule-based machine translation (RBMT) system (Yamabana,1997) .", "Furthermore, Kawahara (1994 Kawahara ( , 2002 defines the Japanese shallow CF that is widely and successfully used in Japanese dependency tasks provided by CoNLL-09 (Hajič, 2011) .", "Figure 2 shows the CF's ability to capture reordering information.", "In this paper, we describe effective approaches to introducing source language Japanese CF in the Japanese-Chinese translation task.", "Unlike previous work, we are the first to use Japanese CF information on the HPB model, and to transform CF information into SCFG style rules, which is suitable and useful in the original HPB decoder.", "By importing CF into the HPB model, we expand search space and introduce fine-grained rules.", "The remainder of this paper is organized as follows.", "After introducing Japanese CF,the proposed approach is introduced in Section 3; the experimental results and associate analysis are given Case Frame Unlike HPB model's format grammar, case grammar is linguistically sensible and is applied to semantically analyze sentence.", "Based on case grammar, a sentence will be analyzed using different deep case components (agent, instrument, experiencer, object, location, benefactive, factitive, goal, source and time).", "This way, Fillmore has defined the deep verb CF, where one example is shown in Figure 2 (a).", "Deep case is language independent.", "If two sentences from different languages have the exactly same meaning and description, they will have the same deep case grammar analysis.", "Figure 2 (a) shows the sentence \"today we will get him to the airport by car\" described respectively in Japanese and Chinese.", "Meanwhile, Figure 2 (a) shows deep case alignment between these two different languages.", "Deep case alignment in two different languages is one to one mapping.", "For example, in Figure 2 (a), \"私 達 は\" (we) is the agent in Japanese, mapping \" 我们 \" (we) (agent) in Chinese.", "The deep CF is well known, but it is rarely used in statistical machine translation due to the difficulty of the auto-analysis for all languages including Japanese.", "However, due to the explicit case in Japanese, Kawahara (2002) redefines the shallow verbal CF in Japanese shown in Figure 2 analysis.", "As a result, recent research has achieved high accuracy (more than 90%) on Japanese shallow CF analysis (Kawahara and Kurohashi, 2006) .", "Between the deep case and the Japanese shallow case, there is a many-to-many relation shown in Table 1 .", "In this paper, we will only use \"case frame\" to represent Japanese verbal shallow CF for short.", "The proposed approach A case frame is the linguistic concept, which provides linguistic guidance for derivation.", "Here, we present a method to alleviate complex reordering problems in the Japanese-Chinese machine translation task with case frame.", "Generally, we obtain both the case frame and the hiero-style SCFG from the training data, and then transfer the case frame rule (CFR) to SCFG style and use both of them in decoding with the SCFG.", "The method benefits from both hiero-style translation and linguistic information.", "In the rule extraction of our approach, we acquire case frame rules using fuzzy strategy and hiero-style rules using traditional HPB rule extraction method.", "In decoding, we use the traditional HPB decoder with CYK and cube pruning.", "Figure 3 shows an example of CFRs extraction processing from a pair of word-aligned Japanese-Chinese sentences with a source language CF, and their SCFG style.", "Case Frame Rules Extraction As described in section 2, the Japanese shallow case frame can be obtained through surface analysis.", "This way, we can extract case frame reordering rules from sentence pairs with alignment information as shown in Figure 4 , where original case frame rules are from o1 to o6.", "Given a source language case frame and related word alignment, one case frame is mapped to the case frame reordering rule set,where there are two kinds of rules: reordering rule and phrase rule.", "• Phrase rule: Each component in a case frame generates one phrase rule.", "We extract the phrase rule by following the traditional phrase-based model 's strategy (Och and Ney, 2004) .", "Each phrase rule has a case distinction associated with a shallow case in a case frame like r1 to r5 in Figure 3 .", "• Reordering rule: One case frame generates one reordering rule.", "For reordering rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B in target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All of the Transforming Case Frame Rule into SCFG style To make case frame rules directly accessible to the Hiero-style decoder with performs decoding with SCFG rules, we convert original case frame rules into SCFG style.", "And then, case frame rule is defined as SCFG-style, which is a little different from hiero rules.", "• Phrase rule transformation: We take o1 as an example transforming into r1 shown in Figure 5 (a).", "We use o1 's case distinction as case distinction of r1 's left.", "The source side of the r1 's right is source phrase in o1 and the target side is target phrase in o1.", "• Reordering rules transformation: We take o6 as an example transforming into r6 shown in Figure 5(b) .", "We also use o6 's verb case distinction as case distinction of r6 's left.", "(default X if there is no case distinction in this example).", "Each slot of o6 is transformed into related X with respective case distinction in r6.", "The target side of the rule 's right is target language 's reordering.", "It is clearly seen that if there is no non-terminals in the right of reordering rule, reordering rule is the same with phrase rule.", "In this way, each case frame rule is associated with exactly one SCFG rule.", "Therefore, we can obtain a fine-grained SCFG from case frames due to case distinction.", "On one hand, non-terminals associated with case are linguistically sensible.", "For example of r4, \"空港 まで\" with \"マデ\" case is translated to \"去 机场\" that means \"to airport\".", "On the other hand, it can capture complex reordering information.", "For example of r6, the source side of the rule's right means that \"ガ\" (who) \"時間\" (when) \"ヲ\" (whom) \"マデ\" (where) \"デ\" (how) \"送って 行き ます\" (send)in Japanese order, and the target side of the rule 's right means that \"時 間\" (when) \"ガ\" (who) \"デ\" (how) \"送\" (send) \"デ\" (whom) \"マデ\" (where) in Chinese order.", "For reordering the rule extraction, we need to compute the relatively order of target language span associated with each case slot.", "The order is relatively soft to the word alignment.", "For example, if a source language phrase A covers target span [2, 4] and the other source language phrase B covers target span [1, 3] , then the phrase A is relatively right to the phrase B on the target side; if a span is covered by the other one, the rule is forbidden during extraction.", "All the possible CFs with word alignment can be seen in Figure 4 , where only (c) rule is forbidden.", "Generally, we define the transformed case frame rules as SCFG style: X → ⟨γ, α, ∼⟩ (1) Where X is non-terminal, γ and α are both strings of terminals and non-terminals as the same with SCFG in the HPB model.", "Compared with SCFG in the HPB model, the only difference is that non-terminals are distinguished by case as shown in Figure 3 from r1 to r6.", "Decoding Both transformed case frame rules and HPB rules can be applied using traditional Hiero decoders with a slight modification.", "Here we follow the description of Hiero decoding by Chiang (2007) .", "The source sentence is parsed under the Hiero grammar using the CYK algorithm.", "Each cell in the CYK grid is associated with a list of rules that applies to its span from the bottom up.", "For each derivation, we apply cube pruning (Chiang,2007) and beam search technology.", "This procedure accommodates traditional HPB rules directly.", "We use traditional HPB rules for translation as shown in Figure 6(a) .", "For example, the traditional rule can be applied in the span (14, 16) .", "Since the span (4, 18) is longer than 10 words, the traditional rule cannot be applied in the span.", "We move our focus towards case frame reordering rules, and analyze sentences and obtain all the case frames, and then for each CF, we match rules to the span related to the CF.", "If a match is found, the CYK cell for the span is selected, and that rule is added to the list of rules in the selected CYK cell as shown in Figure 6 (b).", "For example, the span (1,18) can be matched with r6.", "The complex reordering can be captured by r6.", "It is clear that the HPB rules have non-terminals without any distinction and the case frame rules have non-terminals with case distinction.", "Generally, there are two kinds of non-terminals: X and X with case.", "During decoding, we respectively use three kinds of constraints on case frame rule matching: Without constraints ignore all the case distinction in case frame rules, so case frame rule format is the same with HPB rules.", "In this way, we just expand SCFG.", "Soft constraints admit the match between different case distinctions by adding extra dynamic feature -soft count.", "For example, X with \"ヲ\" is allowed to match X with \"マデ\" by adding 1 to soft count.", "Hard constraints only admit the completed and exact match.", "On one hand, we admit X to match all of the X with or without distinction, on the other hand, we only allow X with distinction to match X with the same distinction.", "Features The baseline feature set used in this work consists of 7 features, including a strong 5-gram language model, bidirectional translation probabilities, bidirectional lexical probabilities, and a word count, a glue rule count.", "In the CF reordering rule, bidirectional translation probabilities and bidirection-al lexical probabilities are also used during decoding.", "In addition, we introduce several features for applying case frame rules, and we adopt these features to log-linear model during decoding.", "• Rule type indicators.", "For soft or hard constraint, we consider two indicator features, indicating case frame rules, case frame reordering rules.", "Case frame rules indicator feature is used to distinguish case frame rules and original HPB rules.", "Case frame reordering rules indicator feature is used to distinguish phrase rules and reordering rules in case frame rule set.", "• Dynamic soft constraints.", "For soft constraints, we consider the soft constraints.", "Note that when X with case mismatches X with other different case, we add dynamic soft constraints count for this mismatching instead of prohibition.", "Evaluation Experimental Setup We report results for this Japanese-Chinese task.", "We use two data sets, where one uses news from the 7th China Workshop on Machine Translation (CWMT) including 280 thousand sentence pairs for training, 500 sentence pairs for parameter optimization and 900 sentence pairs for testing, the other, from Asian Scientific Paper Excerpt Corpus-Japanese to Chinese (ASPEC-JC) includes 680 thousand pairs for training, 2090 sentence pairs for parameter optimization and 1800 sentence pairs for testing.", "The source side sentences are parsed by KNP (Kurohashi and Nagao, 1994) into chunk dependency structures whose nodes are at chunk-level.", "Also we achieve corresponding case frame analysis from byproduct of KNP.", "The word alignment is obtained by running GIZA++ (Och and Ney, 2003) on the corpus in both direction and applying \"grow-diag-and\"refinement (Koehn et al., 2003) .", "We apply SRI Language Modeling Toolkit (Stolcke, 2002) to train a 5-gram language model for target side sentences.", "Results For comparison, we also manually modify the extracted case frame rules of development and test data with case frame information according to the Japanese and Chinese grammar.", "We report machine translation performance in Table 2 using case insensitive BLEU-4 metric (Papineni et al., 2002) , considering the balance of the performance of lexical and phrase.", "The experiments are organized as follows: • exp1: we use the NiuTrans (Xiao, 2012) hierarchical phrase-based model as strong baseline system.", "• exp2: we transform CFRs into SCFG-style rules without any case distinction, and add these rule into exp1 system.", "Analysis Finally, we discuss an example of real translation from our test set.", "See Figure 7 for translations generated from different systems.", "The Japanese input sentence contains \"…下さ れた\" which is usually translated into \"下达… \" (i.e.", "a transformed CF reordering rule \"X → X (下さ れた, 下达\" X)) .", "However, because the \"…下さ れた\" pattern spans 12 words and that is beyond the span limit, our baseline is unable to apply this desired rule and so it chooses the wrong reordering translation.", "When importing CF reordering rule which captures the CF \"(を)下さ れた\" , we can transform the CF reordering rule into one that is SCFG-style and achieve right reordering information.", "• Better reordering Main structure in Japanese structure is SOV-style, which is different from Chinese SVO-style.", "Reordering problem is significant in Japanese-Chinese translation, especially with long phrase for S and/or V. Compared with hierarchical phrase-based rules, case frame rules have better phrase reordering.", "In the example as shown in Figure 8 , the source sentence main centered verbs contain the word \"確認(confirm)\" and the word \"集合(gather)\".", "The Hiero result mistakenly treats that objective phrase as subjective (SOV), thus results in translation with different structure from source sentence.", "Conversely, our system captures this component relations in case frame and translates it into the SVO structure.", "• Better exical translation results Moreover, we also find that our system can get better lexical translation results, for instance, the result of the word \"時間厳守(punctuality)\",as indicated in Figure 8 .", "Related Work Recently linguistically-motivated models have been intensively investigated in MT.", "In particular, source tree-based models (Liu et al., 2006; Huang et al., 2006; Eisner, 2003; Zhang et al., 2008; Liu et al., 2009a; Xie et al., 2011) have received growing interest due to their excellent ability to model source language syntax for better lexical selection and reordering.", "Alternatively, the hierarchical phrase-based approach (Chiang, 2005) considers the underlying hierarchical structures of sentences but does not require linguistically syntactic trees on either language ′ s side.", "There are several lines of work for augmenting hierarchical phrase-based systems with the use of source language linguistic information.", "Xiao (2014) incorporates source syntax into the hierarchical phrase-based model.", "They develop procedures for joint decoding and optimization within a single system by transforming tree-to-string rules into SCFG rules.", "By enlarging SCFG grammar, they perform well on Chinese-English tasks.", "Our approach is motivated by high-precision Japanese case analysis, and aims to augment the search space of Hiero with linguistically-motivated hypotheses.", "Moreover, we consider hiero as the backbone model and only introduce and transform Japanese CF into SCFG rules where they can contribute.", "Another related line of work is to introduce pre-ordering approach for Japanese main structure.", "Wu (2011) and Sudoh (2013) propose several methods to train pre-ordering model for pre-ordering.", "We note that, we have no need to train extra pre-ordering models for the Japanese main structure, and we only use the highprecision Japanese explicit case analysis to improve Japanese-Chinese translation performance described in this paper.", "Conclusion and Future Work We have presented an approach to improving Hiero-style systems by augmenting the SCFG with Japanese case frame rules.", "The input case frame are used to introduce new linguisticallysensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong Hiero baseline in the Japanese-to-Chinese task.", "This paper presented an approach to improve H-PB model systems by augmenting the SCFG with Japanese CFRs.", "The CF are used to introduce new linguistically-sensible hypotheses into the translation search space while maintaining the Hiero robustness qualities and avoiding computational explosions.", "We obtain significant improvements over a strong HPB baseline in the Japanese-to-Chinese task.", "We will try to improve the performance of our system with soft constraint or hard constraint using case frame rules, and we will challenge to resolve the problem of tense, aspect and some special grammatical sentences of Japanese to Chinese translation.", "Training Program of Innovation and Entrepreneurship for Undergraduates." ] }

{ "paper_header_number": [ "1", "2", "3", "3.1", "3.2", "3.3", "3.4", "4.1", "4.2", "4.3", "5", "6" ], "paper_header_content": [ "Introduction", "Case Frame", "The proposed approach", "Case Frame Rules Extraction", "Transforming Case Frame Rule into SCFG style", "Decoding", "Features", "Experimental Setup", "Results", "Analysis", "Related Work", "Conclusion and Future Work" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129 ], "paper_content_text": [ "Introduction The CL-SciSumm task provides resources to encourage research in a promising direction of scientific paper summarization, which considers the set of citation sentences (i.e., \"citances\") that reference a specific paper as a (community created) summary of a topic or paper [21] .", "Citances for a reference paper are considered a synopses of its key points and also its key contributions and importance within an academic community [19] .", "The advantage of using citances is that they are embedded with meta-commentary and offer a contextual, interpretative layer to the cited text.", "The drawback, however, is that though a collection of citances offers a view of the cited paper, it does not consider the context of the target user [9] [24] , verify the claim of the citation or provide context from the reference paper, in terms the type of information cited or where it is in the referenced paper [8] .", "CL-SciSumm explores summarization of scientific research, for the computational linguistics research domain.", "It encourages the incorporation of new kinds of information in automatic scientific paper summarization, such as the facets of research information being summarized the research paper.", "Our previous task suggested that scholars in CL typically cite methods information from other papers.", "CL-SciSumm also encourages the use of citing mini-summaries written in other papers, by other scholars, when they refer to the paper.", "It is anticipated that these selected facts would closely reflect the most important contributions and applications of the paper.", "These insights have been explored in a smaller scope by previous work.", "We propose that further explorations can help to advance the state of the art.", "Furthermore, we expect that the CL-SciSumm Task could spur the creation of new resources and tools, to automate the synthesis and updating of automatic summaries of CL research papers.", "Previous work in scientific summarization has attempted to automatically generate multi-document summaries by instantiating a hierarchical topic tree [6] , generating model citation sentences [17] or implementing a literature review framework [8] .", "However, the limited availability of evaluation resources and humancreated summaries constrains research in this area.", "In 2014, the CL-SciSumm Pilot task was conducted as a part of the larger BioMedSumm Task at TAC 5 .", "In 2016, our proposal was not successful with ACL; fortunately it was accepted as a part of the BIRNDL workshop [15] at JCDL-2016 6 .", "The development and dissemination of the CL-SciSumm dataset and the related Shared Task has been generously supported by the Microsoft Research Asia (MSRA) Research Grant 2016.", "Task Given: A topic consisting of a Reference Paper (RP) and up to ten Citing Papers (CPs) that all contain citations to the RP.", "In each CP, the text spans (i.e., citances) have been identified that pertain to a particular citation to the RP.", "Task 1A: For each citance, identify the spans of text (cited text spans) in the RP that most accurately reflect the citance.", "These are of the granularity of a sentence fragment, a full sentence, or several consecutive sentences (no more than 5).", "Task 1B: For each cited text span, identify what facet of the paper it belongs to, from a predefined set of facets.", "Task 2: Finally, generate a structured summary of the RP from the cited text spans of the RP.", "The length of the summary should not exceed 250 words.", "This was an optional bonus task.", "Evaluation: Participants were required to submit their system outputs from the test set to the task organizers.", "An automatic evaluation script was used to measure system performance for Task 1a, in terms of the sentence id overlaps between the sentences identified in system output, versus the gold standard created by human annotators.", "Task 1b was evaluated as a proportion of the correctly classified discourse facets by the system, contingent on the expected response of Task 1a.", "Task 2 was optional, and evaluated using the ROUGE-N [12] scores between the system output and three types of gold standard summaries of the research paper.", "Data: The dataset comprises ten pairs of training sets, development and test sets.", "Each pair comprises the annotated citing sentences for a research paper and the discourse facets being referenced, and summaries of the research paper.", "CL-SciSumm Pilot 2014 The CL Summarization Pilot Task [7] was conducted as a part of the Biomed-Summ Track at the Text Analysis Conference 2014 (TAC 2014) 7 .", "Ten pairs of annotated citing sentences and summaries were made available to the participants, who reported their performance on the same Tasks described above, as a cross-validation over the same dataset.", "System outputs for Task 1a were scored using word overlaps with the gold standard measured by the ROUGE-L score.", "Task 1b was scored using precision, recall and F 1 .", "Task 2 was an optional task where system summaries were evaluated against the abstract using ROUGE-L. No centralized evaluation was performed.", "All scores were self-reported.", "Three teams submitted their system outputs.", "clair umich was a supervised system using lexical, syntactic and WordNet based features; MQ system used information retrieval inspired ranking methods; TALN.UPF used various TF-IDF scores.", "During this task, the participants reported several errors in the dataset including text encoding and inconsistencies in the text offsets.", "The annotators also reported flaws in the xml encoding, and problems in the OCR export to XML.", "These issues hindered system building and evaluation.", "Accordingly, changes were made to the annotation file format and the XML transformation process in the current task.", "Development The CL-SciSumm 2016 task included the original training dataset of the Pilot Task, to encourage teams from the previous edition to participate.", "It also incorporated a new development corpus of ten sets for system training, and a separate test corpus of ten sets for evaluation.", "Additionally, it provided three types of summaries for each set in each corpus -the abstract, written by the authors of the research paper the community summary, collated from the reference spans of its citances human-written summary written by the annotators of the CL-SciSumm annotation effort 7 http://www.nist.gov/tac/2014 For the general procedure followed to construct the CL-SciSumm corpus, please see [7] .", "There are two differences in the selection of citing papers (CP) for the training corpus, as compared to the development and test corpora.", "Firstly, the minimum numbers of CP provided in the former, which was 3, was increased to 8 in the construction of the latter.", "Secondly, the maximum number of CPs provided in the former was 10, but this limit was removed in the construction of the latter, so that up to 60 CPs have been provided for a single RP.", "This was done to have more citances of which potentially more would mention the RP in greater detail.", "This would also produce a wider perspective in the community summary.", "Annotation The annotators of the development and test corpora were five postgraduate students in Applied Linguistics, from University of Hyderabad, India.", "They were selected out of a larger pool of over twenty-five participants, who were all trained to annotate an RP and its CPs on their personal laptops, using the Knowtator 8 annotation package of the Protege editing environment 9 .", "The annotation scheme was unchanged from what was followed by [7] : Given each RP and its associated CPs, the annotation group was instructed to find citations to the RP in each CP.", "Specifically, the citation text, citation marker, reference text, and discourse facet were identified for each citation of the RP found in the CP.", "Inadvertently, we included the gold standard annotations for Task 1a and 1b when we released the test corpus.", "We alerted the participating teams to this mistake and requested them not to use that information for training their systems.", "Overview of Approaches The following paragraphs discuss the approaches followed by the participating systems, in no particular order.", "Except for the top performing systems in each of the sub-tasks, we do not provide detailed relative performance information for each system, in this paper.", "The evaluation scripts have been provided at the CL-SciSumm Github respository 10 where the participants may run their own evaluation and report the results.", "The approach by [14] used the Transdisciplinary Scientific Lexicon (TSL) developed by [5] to build a profile for each discourse facet in citances and reference spans.", "Then a similarity function developed by [16] was used to select the best-matching reference span with the same facet as the citance.", "For Task 2, the authors used Maximal Marginal Relevance [3] to choose sentences so that they brought new information to the summary.", "Nomoto [20] proposed a hybrid model for Task 2, comprising TFIDF and a tripartite neural network.", "Stochastic gradient descent was performed on a training data comprising of triples of citance, the true reference and the set of false references for the citance.", "Sentence selection was based on a dissimilarity score similar to MMR.", "Mao et al.", "[11] used an SVM classifier with a topical lexicon to identify the best matching reference spans for a citance, using ifd similarity, Jaccard similarity and context similarity.", "They finally submitted six system runs, each following a variant of similarities and approaches -the fusion method, the Jaccard Cascade method, the Jaccard Focused method, the SVM method and two voting methods.", "Klampfl et al.", "[10] developed three different approaches based on summarization and classification techniques.", "They applied a modified version of an unsupervised summarization technique, termed it TextSentenceRank, to the reference document.", "Their second method incorporates similarities of sentences to the citation on a textual level, and employed classification to select from candidates previously extracted through the original TextSentenceRank algorithm.", "Their third method used unsupervised summarization of the relevant sub-part of the document that was previously selected in a supervised manner.", "Saggion et al.", "[23] reported their results for the linear regression implementation of WEKA used together with the GATE system.", "They trained their model to learn the weights of different features with respect to the relevance of cited text spans and the relevance to a community-based summary.", "Two runs were submitted, using SUMMA [22] to score and extract all matched sentences and only the top sentences respectively.", "Lu et al.", "[13] regarded Task 1a as a ranking problem, applying Learning to Rank strategies.", "In contrast, the group cast Task 1b as a standard text classification problem, where novel feature engineering was the team's focus.", "Along this vein, the group considered features of both citation contexts and cited spans.", "Aggarwal and Sharma [1] propose several heuristics derived from bigram overlap counts between citances and reference text to identify the reference text span for each citance.", "This score is used to rank and select sentences from the reference text as output.", "Baki et al.", "[18] used SVM with subset tree kernel, a type of convolution kernel.", "Computed similarities between three tree representations of the citance and reference text formed the convolution kernel.", "Their set-up scored better than their TF-IDF baseline method.", "They submitted three system runs with this approach.", "The PolyU system [2] , for Task 1a, use SVM-rank with lexical and document structural features to rank reference text sentences for every citance.", "Task 1b is solved using a decision tree classifier.", "Finally, they model summarization as a query-focussed summarization with citances as queries.", "They generate summaries (Task 2) by improvising on a Manifold Ranking method (see [2] for details).", "Finally, the system submitted by Conroy and Davis [4] attempted to solve Task 2 with an adaptation of a system developed for the TAC 2014 BioMedSumm Task 11 .", "They provided the results from a simple vector space model, wherein they used a TF representation of the text and non-negative matrix factorization (NNMF) to estimate the latent weights of the terms for scientific document summarization.", "They also provide the results from two language models based on the distribution of words in human-written summaries.", "System Runs Performance of systems for Task 1a was measured by the number of sentences output by the system that overlap with the sentences in the human annotated reference text span (see section 4.1).", "These numbers were then used to calculate the precision, recall and F 1 score for each system.", "As Task 1b is a multi-label classification, this task was also scored by metrics -precision, recall and F 1 score.", "Nine systems submitted outputs for Task 1.", "The following plots rank the systems for Task 1 by their F 1 scores.", "In the figures, all the systems have been identified by their participant number.", "Only the top performing systems for Tasks 1a, 1b and 2 have been identified by name in sections 6 and 7.", "Task 2, to create a summary of the reference paper from citances and the reference paper text, was evaluated against 3 types of gold standard summaries: the reference paper's abstract, a community summary and a human summary.", "A Java Implementation of ROUGE 12 was used to compare the gold summaries against summaries generated by systems.", "We calculated ROUGE-2 and ROUGE-4 F 1 scores for the system summaries against each of the 3 summary types.", "ROUGE-1 and ROUGE-3, which showed similar results have been omitted from this paper.", "Four of the nine system that did Task 1 also did the bonus Task 2.", "Following are the plots with their performance measured by ROUGE-2 and ROUGE-4 against the 3 gold standard summary types.", "For Task 1a, the best performance was shown by sys16, developed by [2] .", "The next best performance was shown by sys8 [11] and sys6 [20] .", "For Task 1b, the best performance was shown by sys8 [11] , followed by the systems sys16 [2] and sys10 [23] .", "For Task 2, the system by [11] , sys8, performed the best against abstract and community summaries, while sys16 [2] performed well on community summaries.", "The system by sys15 [1] , was also a strong performer on these tasks.", "On human summaries, the best performance was seen from sys3 [4] .", "The F 1 scores of all systems on Tasks 1a and 1b were generally low.", "However, the systems ranked in the first 3 places, did significantly better than systems ranked in the last 3 places.", "On Task 2, all systems except sys16 performed better when evaluated against abstracts, than against other summary types.", "Furthermore, system performances did not differ significantly from one another when evaluated against human and community summaries.", "However, when evaluated against abstracts, the best performing system significantly outperforms systems ranked in the lower half.", "Conclusion Ten systems participated in the CL-SciSumm Task 2016.", "A variety of heuristical, lexical and supervised approaches were used.", "Two of the best performing systems in Task 1a and 1b were also participants in the CL-SciSumm Pilot Task.", "The results from Task 2 suggest that automatic summarization systems may be adaptable to different domains, as we observed that the system by [4] , which had originally been developed for biomedical human summaries, outperformed the others.", "We also note that systems performing well on Tasks 1a and 1b also do well in generating community summaries -this supports our expectations about the Shared Task, and validates the need to push the state-of-the-art in scientific summarization.", "In future work, other methods of evaluation can be used for comparing the performance of the different approaches, and a deeper analysis can lead to new insights about which approaches work well with certain kinds of data.", "However, such an inquiry was beyond the scope of this overview paper.", "We deem our Task a success, as it has spurred the interest of the community and the development of tools and approaches for scientific summarization.", "We are investigating other potential subtasks which could be added into our purview.", "We are also scouting for other related research problems, of relevance to the scientific summarization community." ] }

{ "paper_header_number": [ "1", "2", "3", "4", "4.1", "5", "6", "7" ], "paper_header_content": [ "Introduction", "Task", "CL-SciSumm Pilot 2014", "Development", "Annotation", "Overview of Approaches", "System Runs", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265 ], "paper_content_text": [ "Introduction In this paper we study two bilingual tasks that strongly depend on bilingual word embeddings (BWEs).", "Previously, specialized domain adaptation approaches to such tasks were proposed.", "We instead show experimentally that a simple adaptation process involving only unlabeled text is highly effective.", "We then show that a semisupervised classification method from computer vision can be applied successfully for further gains in cross-lingual classification.", "Our BWE adaptation method is delightfully simple.", "We begin by adapting monolingual word embeddings to the target domain for source and target languages by simply building them using both general and target-domain unlabeled data.", "As a second step we use post-hoc mapping (Mikolov et al., 2013b) , i.e., we use a seed lexicon to transform the word embeddings of the two languages into the same vector space.", "We show experimentally for the first time that the domain-adapted bilingual word embeddings we produce using this extremely simple technique are highly effective.", "We study two quite different tasks and domains, where resources are lacking, showing that our simple technique performs well for both of them: cross-lingual twitter sentiment classification and medical bilingual lexicon induction.", "In previous work, task-dependent approaches were used for this type of domain adaptation.", "Our approach is simple and task independent.", "Second, we adapt the semi-supervised image classification system of Häusser et al.", "(2017) for NLP problems for the first time.", "This approach is broadly applicable to many NLP classification tasks where unlabeled data is available.", "We tailor it to both of our cross-lingual tasks.", "The system exploits unlabeled data during the training of classifiers by learning similar features for similar labeled and unlabeled training examples, thereby extracting information from unlabeled examples as well.", "As we show experimentally, the system further improves cross-lingual knowledge transfer for both of our tasks.", "After combining both techniques, the results of sentiment analysis are competitive with systems that use annotated data in the target language, an impressive result considering that we require no target-language annotated data.", "The method also yields impressive improvements for bilingual lexicon induction compared with baselines trained on in-domain data.", "We show that this system requires the high-quality domain-adapted bilingual word embeddings we previously created to use unlabeled data well.", "Previous Work 2.1 Bilingual Word Embeddings Many approaches have been proposed for creating high quality BWEs using different bilingual signals.", "Following Mikolov et al.", "(2013b) , many authors (Faruqui and Dyer, 2014; Xing et al., 2015; Lazaridou et al., 2015; Vulić and Korhonen, 2016) map monolingual word embeddings (MWEs) into the same bilingual space.", "Others leverage parallel texts (Hermann and Blunsom, 2014; or create artificial cross-lingual corpora using seed lexicons or document alignments (Vulić and Moens, 2015; Duong et al., 2016) to train BWEs.", "In contrast, our aim is not to improve the intrinsic quality of BWEs, but to adapt BWEs to specific domains to enhance their performance on bilingual tasks in these domains.", "Faruqui et al.", "(2015) , Gouws and Søgaard (2015) , Rothe et al.", "(2016) have previously studied domain adaptation of bilingual word embeddings, showing it to be highly effective for improving downstream tasks.", "However, importantly, their proposed methods are based on specialized domain lexicons (such as, e.g., sentiment lexicons) which contain task specific word relations.", "Our delightfully simple approach is, in contrast, effectively task independent (in that it only requires unlabeled in-domain text), which is an important strength.", "Cross-Lingual Sentiment Analysis Sentiment analysis is widely applied, and thus ideally we would have access to high quality supervised models in all human languages.", "Unfortunately, good quality labeled datasets are missing for many languages.", "Training models on resource rich languages and applying them to resource poor languages is therefore highly desirable.", "Crosslingual sentiment classification (CLSC) tackles this problem (Mihalcea et al., 2007; Banea et al., 2010; Wan, 2009; Lu et al., 2011; Balamurali and Joshi, 2012; Gui et al., 2013) .", "Recent CLSC approaches use BWEs as features of deep learning architectures which allows us to use a model for target-language sentiment classification, even when the model was trained only using sourcelanguage supervised training data.", "Following this approach we perform CLSC on Spanish tweets using English training data.", "Even though Spanish is not resource-poor we simulate this by using only English annotated data.", "Xiao and Guo (2013) proposed a cross-lingual log-bilinear document model to learn distributed representations of words, which can capture both the semantic similarities of words across languages and the predictive information with respect to the classification task.", "Similarly, Tang and Wan (2014) jointly embedded texts in different languages into a joint semantic space representing sentiment.", "Zhou et al.", "(2014) employed aligned sentences in the BWE learning process, but in the sentiment classification process only representations in the source language are used for training, and representations in the target language are used for predicting labels.", "An important weakness of these three works was that aligned sentences were required.", "Some work has trained sentiment-specific BWEs using annotated sentiment information in both languages (Zhou et al., 2015 (Zhou et al., , 2016 , which is desirable, but this is not applicable to our scenario.", "Our goal is to adapt BWEs to a specific domain without requiring additional task-specific engineering or knowledge sources beyond having access to plentiful target-language in-domain unlabeled text.", "Both of the approaches we study in this work fit this criterion, the delightfully simple method for adapting BWEs can improve the performance of any off-the-shelf classifier that is based on BWEs, while the broadly applicable semi-supervised approach of Häusser et al.", "(2017) can improve the performance of any off-the-shelf classifier.", "Bilingual Lexicon Induction (BLI) BLI is an important task that has been addressed by a large amount of previous work.", "The goal of BLI is to automatically extract word translation pairs using BWEs.", "While BLI is often used to provide an intrinsic evaluation of BWEs (Lazaridou et al., 2015; Vulić and Moens, 2015; Vulić and Korhonen, 2016) it is also useful for tasks such as machine translation (Madhyastha and España Bohnet, 2017) .", "Most work on BLI using BWEs focuses on frequent words in high-resource domains such as parliament proceedings or news texts.", "Recently Heyman et al.", "(2017) tackled BLI of words in the medical domain.", "This task is useful for many applications such as terminology extraction or OOV mining for machine translation of medical texts.", "Heyman et al.", "(2017) show that when only a small amount of medical data is available, BLI using BWEs tends to perform poorly.", "Especially BWEs obtained using post-hoc mapping (Mikolov et al., 2013b; Lazaridou et al., 2015) fail on this task.", "Consequently, Heyman et al.", "(2017) build BWEs using aligned documents and then engineer a specialized classification-based approach to BLI.", "In contrast, our delightfully simple approach to create high-quality BWEs for the medical domain requires only monolingual data.", "We show that our adapted BWEs yield impressive improvements over non-adapted BWEs in this task with both cosine similarity and with the classifier of Heyman et al.", "(2017) .", "In addition, we show that the broadly applicable method can push performance further using easily accessible unlabeled data.", "Adaptation of BWEs BWEs trained on general domain texts usually result in lower performance when used in a system for a specific domain.", "There are two reasons for this.", "(i) Vocabularies of specific domains contain words that are not used in the general case, e.g., names of medicines or diseases.", "(ii) The meaning of a word varies across domains; e.g., \"apple\" mostly refers to a fruit in general domains, but is an electronic device in many product reviews.", "The delightfully simple method adapts general domain BWEs in a way that preserves the semantic knowledge from general domain data and leverages monolingual domain specific data to create domain-specific BWEs.", "Our domain-adaptation approach is applicable to any language-pair in which monolingual data is available.", "Unlike other methods, our approach is task independent: it only requires unlabeled in-domain target language text.", "Approach To create domain adapted BWEs, we first train MWEs (monolingual word embeddings) in both languages and then map those into the same space using post-hoc mapping (Mikolov et al., 2013b) .", "We train MWEs for both languages by concatenating monolingual out-of-domain and in-domain data.", "The out-of-domain data allows us to create accurate distributed representations of common vocabulary while the in-domain data embeds domain specific words.", "We then map the two MWEs using a small seed lexicon to create the adapted BWEs.", "Because post-hoc mapping only requires a seed lexicon as bilingual signal it can easily be used with (cheap) monolingual data.", "For post-hoc mapping, we use Mikolov et al.", "(2013b)'s approach.", "This model assumes a W ∈ R d 1 ×d 2 matrix which maps vectors from the source to the target MWEs where d 1 and d 2 are the embedding space dimensions.", "A seed lexicon of (x i , y i ) ∈ L ⊆ R d 1 ×R d 2 pairs is needed where x i and y i are source and target MWEs.", "W can be learned using ridge regression by minimizing the L 2 -regularized mapping error between the source x i and the target y i vectors: min W i ||W x i − y i || 2 2 + λ||W || 2 2 (1) where λ is the regularization weight.", "Based on the source embedding x, we then compute a target embedding as W x.", "We create MWEs with word2vec skipgram (Mikolov et al., 2013a) 1 and estimate W with scikit-learn (Pedregosa et al., 2011) .", "We use default parameters.", "Cross-Lingual Sentiment Classification In CLSC, an important application of BWEs, we train a supervised sentiment model on training data available in the source (a resource rich language) and apply it to the target (a resource poor language, for which there is typically no training data available).", "Because BWEs embed source and target words in the same space, annotations in the source (represented as BWEs) enable transfer learning.", "For CLSC of tweets, a drawback of BWEs trained on non-twitter data is that they do not produce embeddings for twitter-specific vocabulary, e.g., slang words like English coool and (Mexican) Spanish chido, resulting in lost information when a sentiment classifier uses them.", "Training Data for Twitter Specific BWEs As comparable non-twitter data we use OpenSubtitles (Lison and Tiedemann, 2016) which contains 49.2M English and Spanish subtitle sentences respectively (Subtitle).", "The reason behind choosing Subtitles is that although it is out-of-domain it contains slang words similar to tweets thus serving as a strong baseline in our setup.", "We experiment with two monolingual twitter data sets: (i) 22M tweets: Downloaded 2 English (17.2M) and Spanish (4.8M) tweets using the public Twitter Streaming API 3 with language filters en and es (ii) a BACKGROUND corpus of 296K English and 150K Spanish (non-annotated) tweets released with the test data of the RepLab task (Amigó et al., 2013) described below All twitter data was tokenized using Bird et al.", "(2009) and lowercased.", "User names, URLs, numbers, emoticons and punctuation were removed.", "As lexicon for the mapping, we use the BNC word frequency list (Kilgarriff, 1997) , a list of 6,318 frequent English lemmas and their Spanish translations, obtained from Google Translate.", "Note that we do not need a domain-specific lexicon in order to get good quality adapted BWEs.", "Training Data for Sentiment Classifiers For sentiment classification, we use data from the RepLab 2013 shared task (Amigó et al., 2013) .", "The data is annotated with positive, neutral and negative labels and contains English and Spanish tweets.", "We used the official English training set (26.6K tweets) and the Spanish test set (14.9K) in the resource-poor setup.", "We only use the 7.2K Spanish labeled training data for comparison reasons in §6.2, which we will discuss later.", "The shared task was on target-level sentiment analysis, i.e., given a pair (document, target entity), the gold annotation is based on whether the sentiment expressed by the document is about the target.", "For example: I cried on the back seat of my BMW!", "where BMW is the target would be negative in the sentence-level scenario.", "However, it is neutral in the target-level case because the negative sentiment is not related to BMW.", "The reason for using this dataset is that it contains comparable English and Spanish tweets annotated for sentiment.", "There are other twitter datasets for English (Nakov et al., 2016) and Spanish (Garcıa-Cumbreras et al., 2016), but they were downloaded at different times and were annotated using different annotation methodologies, thus impeding a clean and consistent evaluation.", "Sentiment Systems For evaluating our adapted BWEs on the RepLab dataset we used a target-aware sentiment classifier introduced by Zhang et al.", "(2016) .", "The network first embeds input words using pre-trained 3 dev.twitter.com/streaming/overview BWEs and feeds them to a bi-directional gated neural network.", "Pooling is applied on the hidden representations of the left and right context of the target mention respectively.", "Finally, gated neurons are used to model the interaction between the target mention and its surrounding context.", "During training we hold our pre-trained BWEs fixed and keep the default parameters of the model.", "We also implement Kim (2014)'s CNN-nonstatic system, which does not use the target information in a given document (target-ignorant).", "The network first embeds input words using pretrained BWEs and feeds them to a convolutional layer with multiple window sizes.", "Max pooling is applied on top of convolution followed by a fully connected network with one hidden layer.", "We used this system as well because it performed comparably to the target-aware system.", "The reason for this is that only 1% of the used data contains more than one target and out of these rare cases only 14% have differing sentiment labels in the same sentence, which are the difficult cases of target-level sentiment analysis.", "We used the default parameters as described in (Kim, 2014) with the exception of using 1000 feature maps and 30 epochs, based on our initial experiments.", "Word embeddings are fixed during the training just as for the target-aware classifier.", "Results As we previously explained we evaluate our adaptation method on the task of target-level sentiment classification using both target-aware and target-ignorant classifiers.", "For all experiments, our two baselines are off-the-shelf classifiers using non-adapted BWEs, i.e., BWEs trained only using Subtitles.", "Our goal is to show that our BWE adaptation method can improve the performance of such classifiers.", "We train our adapted BWEs on the concatenation of Subtitle and 22M tweets or BACKGROUND respectively.", "In addition, we also report results with BWEs trained only on tweets.", "To train the sentiment classifiers we use the English Replab training set and we evaluate on the Spanish test set.", "To show the performance that can be reached in a monolingual setup, we report results obtained by using annotated Spanish sentiment data instead of English (oracle).", "We train two oracle sentiment classifiers using (i) MWEs trained on only the Spanish part of Subtitle and (ii) BWEs trained on Subtitle using posthoc mapping.", "The difference between the two is that the embeddings of (ii) are enriched with English words which can be beneficial for the classification of Spanish tweets because they often contain a few English words.", "We do not compare with word embedding adaptation methods relying on specialized resources.", "The point of our work is to study task-independent methods and to the best of our knowledge ours is the first such attempt.", "Similarly, we do not compare against machine translation based sentiment classifiers (e.g., (Zhou et al., 2016) ) because for their adaptation in-domain parallel data would be needed.", "Table 1 gives results for both classifiers.", "It shows that the adaptation of Subtitle based BWEs with data from Twitter (22M tweets and BACK-GROUND) clearly outperforms the Baseline in all cases.", "The target-aware system performed poorly with the baseline BWEs and could benefit significantly from the adaptation approach.", "The target-ignorant performed better with the baseline BWEs but could also benefit from the adaptation.", "Comparing results with the Twitter-dataset-only based BWEs, the 22M tweets performed better even though the BACKGROUND dataset is from the same topic as the RepLab train and test sets.", "Our conjecture is that the latter is too small to create good BWEs.", "In combination with Subtitles, 22M tweets also yields better results than when combined with BACKGROUND.", "Although the best accuracy was reached using the 22M tweetsonly based BWEs, it is only slightly better then the adapted Subtitles+22M tweets based BWEs.", "In §6 we show that both the semantic knowledge from Subtitles and the domain-specific information from tweets are needed to further improve results.", "Comparing the two classifiers we can say that they performed similarly in terms of their best results.", "On the other hand, the target-ignorant system had better results on average.", "This might seem surprising at first because the system does not use the target as information.", "But considering the characteristics of RepLab, i.e., that the number of tweets that contains multiple targets is negligible, using the target offers no real advantage.", "Although we did not focus on the impact of the seed lexicon size, we ran post-hoc mapping with different sizes during our preliminary experiments.", "With 1,000 and 100 word pairs in the lexicon the target-ignorant system suffered 0.5% and 4.0% drop in average of our setups respectively.", "To summarize the result: using adapted BWEs for the Twitter CLSC task improves the performance of off-the-shelf classifiers.", "Medical Bilingual Lexicon Induction Another interesting downstream task for BWEs is bilingual lexicon induction.", "Given a list of words in a source language, the goal of BLI is to mine translations for each word in a chosen target language.", "The medical bilingual lexicon induction task proposed in (Heyman et al., 2017) aims to mine medical words using BWEs trained on a very small amount of English and Dutch monolingual medical data.", "Due to the lack of resources in this domain, good quality BWEs are hard to build using in-domain data only.", "We show that by enriching BWEs with general domain knowledge (in the form of general domain monolingual corpora) better results can be achieved on this medical domain task.", "Experimental Setup We evaluate our improved BWEs on the dataset provided by Heyman et al.", "(2017) .", "The monolingual medical data consists of English and Dutch medical articles from Wikipedia.", "The English (resp.", "Dutch) articles contain 52,336 (resp.", "21,374) sentences.", "A total of 7,368 manually annotated word translation pairs occurring in the English (source) and Dutch (target) monolingual corpora are provided as gold data.", "This set is split 64%/16%/20% into trn/dev/test.", "20% of the English words have multiple translations.", "Given an English word, the task is to find the correct Dutch translation.", "As monolingual general-domain data we use Table 2 : We report F 1 results for medical BLI with the cosine similarity and the classifier based systems.", "We present baseline and our proposed domain adaptation method using both general and medical lexicons.", "the English and Dutch data from Europarl (v7) (Koehn, 2005) , a corpus of 2 million sentence pairs.", "Although Europarl is a parallel corpus, we use it in a monolingual way and shuffle each side of the corpus before training.", "By using massive cheap data we create high-quality MWEs in each language which are still domain-specific (due to inclusion of medical data).", "To obtain an out-ofdomain seed lexicon, we translated the English words in BNC to Dutch using Google Translate (just as we did before for the Twitter CLSC task).", "We then use the out-of-domain BNC and the indomain medical seed lexicons in separate experiments to create BWEs with post-hoc mapping.", "Note, we did not concatenate the two lexicons because (i) they have a small common subset of source words which have different target words, thus having a negative effect on the mapping and (ii) we did not want to modify the medical seed lexicon because it was taken from previous work.", "BLI Systems To perform BLI we use two methods.", "Because BWEs represent words from different languages in a shared space, BLI can be performed via cosine similarity in this space.", "In other words, given a BWE representing two languages V s and V t , the translation of each word s ∈ V s can be induced by taking the word t ∈ V t whose representation x t in the BWE is closest to the representation x s .", "As the second approach we use a classifier based system proposed by Heyman et al.", "(2017) .", "This neural network based system is comprised of two main modules.", "The first is a character-level LSTM which aims to learn orthographic similarity of word pairs.", "The other is the concatenation of the embeddings of the two words using embedding layers with the aim of learning the similarity among semantic representations of the words.", "Dense layers are applied on top of the two modules before the output soft-max layer.", "The classifier is trained using positive and negative word pair examples and a pre-trained word embedding model.", "Negative examples are randomly generated for each positive one in the training lexicon.", "We used default parameters as reported by Heyman et al.", "(2017) except for the t classification thresholds (used at prediction time).", "We finetuned these on dev.", "We note that the system works with pre-trained MWEs as well (and report these as official baseline results) but it requires BWEs for candidate generation at prediction time, thus we use BWEs for the system's input for all experiments.", "In preliminary work, we had found that MWE and BWE results are similar.", "Results Heyman et al.", "(2017)'s results are our baseline.", "Table 2 compares its performance with our adapted BWEs, with both cosine similarity and classification based systems.", "\"top\" F 1 scores are based on the most probable word as prediction only; \"all\" F 1 scores use all words as prediction whose probability is above the threshold.", "It can be seen that the cosine similarity based system using adapted BWEs clearly outperforms the nonadapted BWEs which were trained in a resource poor setup.", "4 Moreover, the best performance was reached using the general seed lexicon for the mapping which is due to the fact that general domain words have better quality embeddings in the MWE models, which in turn gives a better quality mapping.", "The classification based system performs significantly better comparing to cosine similarity by exploiting the seed lexicon better.", "Using adapted BWEs as input word embeddings for the system further improvements were achieved which shows the better quality of our BWEs.", "Simulating an even poorer setup by using a general lexicon, the performance gain of the classifier is lower.", "This shows the significance of the medical seed lexicon for this system.", "On the other hand, adapted BWEs have better performance compared to non-adapted ones using the best translation while they have just slightly lower F 1 using multiple translations.", "This result shows that while with adapted BWEs the system predicts better \"top\" translations, it has a harder time when predicting \"all\" due to the increased vocabulary size.", "To summarize: we have shown that adapted BWEs increase performance for this task and domain; and they do so independently of the taskspecific system that is used.", "Semi-Supervised Learning In addition to the experiments that show our BWEadaptation method's task and language independence, we investigate ways to further incorporate unlabeled data to overcome data sparsity.", "Häusser et al.", "(2017) introduce a semisupervised method for neural networks that makes associations from the vector representation of labeled samples to those of unlabeled ones and back.", "This lets the learning exploit unlabeled samples as well.", "While Häusser et al.", "(2017) use their model for image classification, we adapt it to CLSC of tweets and medical BLI.", "We show that our semisupervised model requires adapted BWEs to be effective and yields significant improvements.", "This innovative method is general and can be applied to any classification when unlabeled text is available.", "Model Häusser et al.", "(2017)'s basic assumption is that the embeddings of labeled and unlabeled samplesi.e., the representations in the neural network on which the classification layer is applied -are similar within the same class.", "To achieve this, walking cycles are introduced: a cycle starts from a labeled sample, goes to an unlabeled one and ends at a labeled one.", "A cycle is correct if the start and end samples are in the same class.", "The probability of going from sample A to B is proportional to the cosine similarity of their embeddings.", "To maximize the number of correct cycles, two loss functions are employed: Walker loss and Visit loss.", "Walker loss penalizes incorrect walks and encourages a uniform probability distribution of walks to the correct class.", "It is defined as: L walker := H(T, P aba ) (2) where H is the cross-entropy function, P aba ij is the probability that a cycle starts from sample i and ends at j and T is the uniform target distribution: T ij := 1/(#c(i)) if c(i) = c(j) 0 otherwise (3) where c(i) is the class of sample i and #c(i) is the number of occurrences of c(i) in the labeled set.", "Visit loss encourages cycles to visit all unlabeled samples, rather than just those which are the most similar to labeled samples.", "It is defined as: L visit := H(V, P visit ) P visit j := P ab ij i (4) V j := 1 U where H is cross-entropy, P ab ij is the probability that a cycle starts from sample i and goes to j and U is the number of unlabeled samples.", "The total loss during training is the sum of the walker, visit and classification (cross-entropy between predicted and gold labels) losses which is minimized using Adam (Kingma and Ba, 2015) .", "We adapt this model (including the two losses) to sentiment classification, focusing on the targetignorant classifier, and the classifier based approach for BLI.", "We will call these systems semisup 5 .", "Due to the fact that we initialize the embedding layers for both classifiers with BWEs the models are able to make some correct cycles at the beginning of the training and improve them later on.", "We will describe the labeled and unlabeled datasets used in the subsequent sections below.", "We use Häusser et al.", "(2017) 's implementation of the losses, with 1.0, 0.5 and 1.0 weights for the walker, visit and classification losses, respectively, for CLSC based on preliminary experiments.", "We fine-tuned the weights for BLI on dev for each experiment.", "Table 3 : Accuracy on CLSC of the adapted BWE approach with the semisup (target-ignorant with additional loss functions) system comparing to the target-ignorant in brackets.", "Semi-Supervised CLSC As in §4.4, we use pre-trained BWEs to initialize the classifier and use English sentiment training data as the labeled set.", "Furthermore, we use the Spanish sentiment training data as the unlabeled set, ignoring its annotation.", "This setup is very similar to real-word low-resource scenarios: unlabeled target-language tweets are easy to download while labeled English ones are available.", "Table 3 gives results for adapted BWEs and shows that semisup helps only when word embeddings are adapted to the Twitter domain.", "As mentioned earlier, semisup compares labeled and unlabeled samples based on their vector representations.", "By using BWEs based on only Subtitles, we lose too many embeddings of similar English and Spanish tweets.", "On the other hand, if we use only tweet-based BWEs we lose good quality semantic knowledge which can be learned from more standard text domains.", "By combining the two domains we were able to capture both sides.", "For Sub-title+22M tweets, we even get very close to the best oracle (BWE Subtitle) in Table 1 getting only 0.27% less accuracy -an impressive result keeping in mind that we did not use labeled Spanish data.", "The RepLab dataset contains tweets from 4 topics: automotive, banking, university, music.", "We manually analyzed similar tweets from the labeled and unlabeled sets.", "We found that when using semisup, English and Spanish tweets from the same topics are more similar in the embedding space than occurs without the additional losses.", "Topics differ in how they express sentiment -this may explain why semisup increases performance for RepLab.", "Adding supervision.", "To show how well semisup can exploit the unlabeled data we used both English and Spanish sentiment training data together to train the sentiment classifiers.", "Table 4 shows that by using annotated data in both languages we get clearly better results than when using only one language.", "Tables 3 and 4 show that for Subtitle+22M tweets based BWEs, the semisup approach achieved high improvement (2.17%) comparing to targetignorant with English training data only, while it achieved lower improvement (0.97%) with the Subtitle+BACKGROUND based BWEs.", "On the other hand, adding labeled Spanish data caused just a slight increase comparing to semisup with Subtitle+22M tweets based BWEs (0.59%), while in case of Subtitle+BACKGROUND we got significant additional improvement (2.61%).", "This means that with higher quality BWEs, unlabeled target-language data can be exploited better.", "It can also be seen that the target-aware system outperformed the target-ignorant system using additional labeled target-language data.", "The reason could be that it is a more complex network and therefore needs more data to reach high performance.", "The results in table 4 are impressive: our targetlevel system is strongly competitive with the official shared task results.", "We achieved high accuracy on the Spanish test set by using only English training data.", "Comparing our best system which used all training data to the official results (Amigó et al., 2013) we would rank 2 nd even though our system is not fine-tuned for the Re-pLab dataset.", "Furthermore, we also outperformed the oracles when using annotated data from both languages which shows the additional advantage of using BWEs.", "Semi-Supervised BLI For BLI experiments with semisup we used word pairs from the medical seed lexicon as the labeled set (with negative word pairs generated as described in §5.2).", "As opposed to CLSC and the work of (Häusser et al., 2017) , for this task we do not have an unlabeled set, and therefore we need to generate it.", "We developed two scenarios.", "For the first, BNC, we generate a general unlabeled set using English words from the BNC lexicon and generate 10 pairs out of each word by using the 5 most similar Dutch words based on the corresponding BWEs and 5 random Dutch words.", "For the second scenario, medical, we generate an in-domain unlabeled set by generating for each English word in the medical lexicon the 3 most similar Dutch words based on BWEs and for each of these we use the 5 most similar English words (ignoring the words which are in the original medical lexicon) and 5 negative words.", "The idea behind these methods is to automatically generate an unlabeled set that hopefully has a similar positive and negative word pair distribution to the distribution in the labeled set.", "Table 5 show that adding semisup to the classifier further increases performance for BLI as well.", "For the baseline system, when using only in-domain text for creating BWEs, only the medical unlabeled set was effective, general domain word pairs could not be exploited due to the lack of general semantic knowledge in the BWE model.", "On the other hand, by using our domain adapted BWEs, which contain both general domain and in-domain semantical knowledge, we can exploit word pairs from both domains.", "Results for adapted BWEs increased in 3 out of 4 cases, where the only exception is when using multiple translations for a given source word (which may have been caused by the bigger vocabulary size).", "Results in These results show that adapted BWEs are needed to exploit unlabeled data well which leads to an impressive overall 3.71 increase compared with the best result in previous work (Heyman et al., 2017) , by using only unlabeled data.", "Conclusion Bilingual word embeddings trained on general domain data yield poor results in out-of-domain tasks.", "We presented experiments on two different low-resource task/domain combinations.", "Our delightfully simple task independent method to adapt BWEs to a specific domain uses unlabeled monolingual data only.", "We showed that with the support of adapted BWEs the performance of offthe-shelf methods can be increased for both crosslingual Twitter sentiment classification and medical bilingual lexicon induction.", "Furthermore, by adapting the broadly applicable semi-supervised approach of Häusser et al.", "(2017) (which until now has only been applied in computer vision) we were able to effectively exploit unlabeled data to further improve performance.", "We showed that, when also using high-quality adapted BWEs, the performance of the semi-supervised systems can be significantly increased by using unlabeled data at classifier training time.", "In addition, CLSC results are competitive with a system that uses targetlanguage labeled data, even when we use no such target-language labeled data." ] }

{ "paper_header_number": [ "1", "2", "2.2", "2.3", "3", "3.1", "4", "4.1", "4.2", "4.3", "4.4", "5", "5.1", "5.2", "5.3", "6", "6.1", "6.2", "6.3", "7" ], "paper_header_content": [ "Introduction", "Previous Work 2.1 Bilingual Word Embeddings", "Cross-Lingual Sentiment Analysis", "Bilingual Lexicon Induction (BLI)", "Adaptation of BWEs", "Approach", "Cross-Lingual Sentiment Classification", "Training Data for Twitter Specific BWEs", "Training Data for Sentiment Classifiers", "Sentiment Systems", "Results", "Medical Bilingual Lexicon Induction", "Experimental Setup", "BLI Systems", "Results", "Semi-Supervised Learning", "Model", "Semi-Supervised CLSC", "Semi-Supervised BLI", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266 ], "paper_content_text": [ "Introduction Text Simplification (TS) is generally defined as the conversion of a sentence into one or more simpler sentences.", "It has been shown useful both as a preprocessing step for tasks such as Machine Translation (MT; Mishra et al., 2014; Štajner and Popović, 2016) and relation extraction (Niklaus et al., 2016) , as well as for developing reading aids, e.g.", "for people with dyslexia (Rello et al., 2013) or non-native speakers (Siddharthan, 2002) .", "TS includes both structural and lexical operations.", "The main structural simplification operation is sentence splitting, namely rewriting a single sentence into multiple sentences while preserving its meaning.", "While recent improvement in TS has been achieved by the use of neural MT (NMT) approaches (Nisioi et al., 2017; Zhang and Lapata, 2017) , where TS is consid-ered a case of monolingual translation, the sentence splitting operation has not been addressed by these systems, potentially due to the rareness of this operation in the training corpora (Narayan and Gardent, 2014; Xu et al., 2015) .", "We show that the explicit integration of sentence splitting in the simplification system could also reduce conservatism, which is a grave limitation of NMT-based TS systems (Alva-Manchego et al., 2017) .", "Indeed, experimenting with a stateof-the-art neural system (Nisioi et al., 2017) , we find that 66% of the input sentences remain unchanged, while none of the corresponding references is identical to the source.", "Human and automatic evaluation of the references (against other references), confirm that the references are indeed simpler than the source, indicating that the observed conservatism is excessive.", "Our methods for performing sentence splitting as pre-processing allows the TS system to perform other structural (e.g.", "deletions) and lexical (e.g.", "word substitutions) operations, thus increasing both structural and lexical simplicity.", "For combining linguistically informed sentence splitting with data-driven TS, two main methods have been proposed.", "The first involves handcrafted syntactic rules, whose compilation and validation are laborious (Shardlow, 2014) .", "For example, Siddharthan and Angrosh (2014) used 111 rules for relative clauses, appositions, subordination and coordination.", "Moreover, syntactic splitting rules, which form a substantial part of the rules, are usually language specific, requiring the development of new rules when ported to other languages (Aluísio and Gasperin, 2010; Seretan, 2012; Hung et al., 2012; Barlacchi and Tonelli, 2013 , for Portuguese, French, Vietnamese, and Italian respectively).", "The second method uses linguistic information for detecting potential splitting points, while splitting probabilities are learned us-ing a parallel corpus.", "For example, in the system of Narayan and Gardent (2014) (henceforth, HYBRID) , the state-of-the-art for joint structural and lexical TS, potential splitting points are determined by event boundaries.", "In this work, which is the first to combine structural semantics and neural methods for TS, we propose an intermediate way for performing sentence splitting, presenting Direct Semantic Splitting (DSS), a simple and efficient algorithm based on a semantic parser which supports the direct decomposition of the sentence into its main semantic constituents.", "After splitting, NMT-based simplification is performed, using the NTS system.", "We show that the resulting system outperforms HY-BRID in both automatic and human evaluation.", "We use the UCCA scheme for semantic representation (Abend and Rappoport, 2013) , where the semantic units are anchored in the text, which simplifies the splitting operation.", "We further leverage the explicit distinction in UCCA between types of Scenes (events), applying a specific rule for each of the cases.", "Nevertheless, the DSS approach can be adapted to other semantic schemes, like AMR (Banarescu et al., 2013) .", "We collect human judgments for multiple variants of our system, its sub-components, HYBRID and similar systems that use phrase-based MT.", "This results in a sizable human evaluation benchmark, which includes 28 systems, totaling at 1960 complex-simple sentence pairs, each annotated by three annotators using four criteria.", "1 This benchmark will support the future analysis of TS systems, and evaluation practices.", "Previous work is discussed in §2, the semantic and NMT components we use in §3 and §4 respectively.", "The experimental setup is detailed in §5.", "Our main results are presented in §6, while §7 presents a more detailed analysis of the system's sub-components and related settings.", "Related Work MT-based sentence simplification.", "Phrasebased Machine Translation (PBMT; Koehn et al., 2003) was first used for TS by Specia (2010) , who showed good performance on lexical simplification and simple rewriting, but under-prediction of other operations.", "Štajner et al.", "(2015) took a similar approach, finding that it is beneficial to use training data where the source side is highly similar to the target.", "Other PBMT for TS systems include the work of Coster and Kauchak (2011b) , which uses Moses (Koehn et al., 2007) , the work of Coster and Kauchak (2011a) , where the model is extended to include deletion, and PBMT-R (Wubben et al., 2012) , where Levenshtein distance to the source is used for re-ranking to overcome conservatism.", "The NTS NMT-based system (Nisioi et al., 2017) (henceforth, N17) reported superior performance over PBMT in terms of BLEU and human evaluation scores, and serves as a component in our system (see Section 4).", "took a similar approach, adding lexical constraints to an NMT model.", "Zhang and Lapata (2017) combined NMT with reinforcement learning, using SARI (Xu et al., 2016) , BLEU, and cosine similarity to the source as the reward.", "None of these models explicitly addresses sentence splitting.", "Alva-Manchego et al.", "(2017) proposed to reduce conservatism, observed in PBMT and NMT systems, by first identifying simplification operations in a parallel corpus and then using sequencelabeling to perform the simplification.", "However, they did not address common structural operations, such as sentence splitting, and claimed that their method is not applicable to them.", "Xu et al.", "(2016) used Syntax-based Machine Translation (SBMT) for sentence simplification, using a large scale paraphrase dataset (Ganitketitch et al., 2013) for training.", "While it does not target structural simplification, we include it in our evaluation for completeness.", "Structural sentence simplification.", "Syntactic hand-crafted sentence splitting rules were proposed by Chandrasekar et al.", "(1996) , Siddharthan (2002) , Siddhathan (2011) in the context of rulebased TS.", "The rules separate relative clauses and coordinated clauses and un-embed appositives.", "In our method, the use of semantic distinctions instead of syntactic ones reduces the number of rules.", "For example, relative clauses and appositives can correspond to the same semantic category.", "In syntax-based splitting, a generation module is sometimes added after the split (Siddharthan, 2004) , addressing issues such as reordering and determiner selection.", "In our model, no explicit regeneration is applied to the split sentences, which are fed directly to an NMT system.", "Glavaš andŠtajner (2013) used a rule-based system conditioned on event extraction and syntax for defining two simplification models.", "The eventwise simplification one, which separates events to separate output sentences, is similar to our semantic component.", "Differences are in that we use a single semantic representation for defining the rules (rather than a combination of semantic and syntactic criteria), and avoid the need for complex rules for retaining grammaticality by using a subsequent neural component.", "Combined structural and lexical TS.", "Earlier TS models used syntactic information for splitting.", "Zhu et al.", "(2010) used syntactic information on the source side, based on the SBMT model of Yamada and Knight (2001) .", "Syntactic structures were used on both sides in the model of Woodsend and Lapata (2011) , based on a quasi-synchronous grammar (Smith and Eisner, 2006) , which resulted in 438 learned splitting rules.", "The model of Siddharthan and Angrosh (2014) is similar to ours in that it combines linguistic rules for structural simplification and statistical methods for lexical simplification.", "However, we use 2 semantic splitting rules instead of their 26 syntactic rules for relative clauses and appositions, and 85 syntactic rules for subordination and coordination.", "Narayan and Gardent (2014) argued that syntactic structures do not always capture the semantic arguments of a frame, which may result in wrong splitting boundaries.", "Consequently, they proposed a supervised system (HYBRID) that uses semantic structures (Discourse Semantic Representations, (Kamp, 1981) ) for sentence splitting and deletion.", "Splitting candidates are pairs of event variables associated with at least one core thematic role (e.g., agent or patient).", "Semantic annotation is used on the source side in both training and test.", "Lexical simplification is performed using the Moses system.", "HYBRID is the most similar system to ours architecturally, in that it uses a combination of a semantic structural component and an MT component.", "Narayan and Gardent (2016) proposed instead an unsupervised pipeline, where sentences are split based on a probabilistic model trained on the semantic structures of Simple Wikipedia as well as a language model trained on the same corpus.", "Lexical simplification is there performed using the unsupervised model of Biran et al.", "(2011) .", "As their BLEU and adequacy scores are lower than HYBRID's, we use the latter for comparison.", "Stajner and Glavaš (2017) combined rule-based simplification conditioned on event extraction, to-gether with an unsupervised lexical simplifier.", "They tackle a different setting, and aim to simplify texts (rather than sentences), by allowing the deletion of entire input sentences.", "Split and Rephrase.", "recently proposed the Split and Rephrase task, focusing on sentence splitting.", "For this purpose they presented a specialized parallel corpus, derived from the WebNLG dataset .", "The latter is obtained from the DBPedia knowledge base (Mendes et al., 2012) using content selection and crowdsourcing, and is annotated with semantic triplets of subject-relation-object, obtained semi-automatically.", "They experimented with five systems, including one similar to HY-BRID, as well as sequence-to-sequence methods for generating sentences from the source text and its semantic forms.", "The present paper tackles both structural and lexical simplification, and examines the effect of sentence splitting on the subsequent application of a neural system, in terms of its tendency to perform other simplification operations.", "For this purpose, we adopt a semantic corpus-independent approach for sentence splitting that can be easily integrated in any simplification system.", "Another difference is that the semantic forms in Split and Rephrase are derived semi-automatically (during corpus compilation), while we automatically extract the semantic form, using a UCCA parser.", "Direct Semantic Splitting Semantic Representation UCCA (Universal Cognitive Conceptual Annotation; Abend and Rappoport, 2013) is a semantic annotation scheme rooted in typological and cognitive linguistic theory (Dixon, 2010b (Dixon, ,a, 2012 Langacker, 2008) .", "It aims to represent the main semantic phenomena in the text, abstracting away from syntactic forms.", "UCCA has been shown to be preserved remarkably well across translations (Sulem et al., 2015) and has also been successfully used for the evaluation of machine translation (Birch et al., 2016) and, recently, for the evaluation of TS (Sulem et al., 2018) and grammatical error correction (Choshen and Abend, 2018) .", "Formally, UCCA structures are directed acyclic graphs whose nodes (or units) correspond either to the leaves of the graph or to several elements viewed as a single entity according to some semantic or cognitive consideration.", "A Scene is UCCA's notion of an event or a frame, and is a unit that corresponds to a movement, an action or a state which persists in time.", "Every Scene contains one main relation, which can be either a Process or a State.", "Scenes contain one or more Participants, interpreted in a broad sense to include locations and destinations.", "For example, the sentence \"He went to school\" has a single Scene whose Process is \"went\".", "The two Participants are \"He\" and \"to school\".", "Scenes can have several roles in the text.", "First, they can provide additional information about an established entity (Elaborator Scenes), commonly participles or relative clauses.", "For example, \"(child) who went to school\" is an Elaborator Scene in \"The child who went to school is John\" (\"child\" serves both as an argument in the Elaborator Scene and as the Center).", "A Scene may also be a Participant in another Scene.", "For example, \"John went to school\" in the sentence: \"He said John went to school\".", "In other cases, Scenes are annotated as Parallel Scenes (H), which are flat structures and may include a Linker (L), as in: \"When L [he arrives] H , [he will call them] H \".", "With respect to units which are not Scenes, the category Center denotes the semantic head.", "For example, \"dogs\" is the Center of the expression \"big brown dogs\", and \"box\" is the center of \"in the box\".", "There could be more than one Center in a unit, for example in the case of coordination, where all conjuncts are Centers.", "We define the minimal center of a UCCA unit u to be the UCCA graph's leaf reached by starting from u and iteratively selecting the child tagged as Center.", "For generating UCCA's structures we use TUPA, a transition-based parser (Hershcovich et al., 2017) (specifically, the TUPA BiLST M model).", "TUPA uses an expressive set of transitions, able to support all structural properties required by the UCCA scheme.", "Its transition classifier is based on an MLP that receives a BiLSTM encoding of elements in the parser state (buffer, stack and intermediate graph), given word embeddings and other features.", "The Semantic Rules For performing DSS, we define two simple splitting rules, conditioned on UCCA's categories.", "We currently only consider Parallel Scenes and Elaborator Scenes, not separating Participant Scenes, in order to avoid splitting in cases of nominalizations or indirect speech.", "For example, the sentence \"His arrival surprised everyone\", which has, in addition to the Scene evoked by \"surprised\", a Participant Scene evoked by \"arrival\", is not split here.", "Rule #1.", "Parallel Scenes of a given sentence are extracted, separated in different sentences, and concatenated according to the order of appearance.", "More formally, given a decomposition of a sentence S into parallel Scenes Sc 1 , Sc 2 , · · · Sc n (indexed by the order of the first token), we obtain the following rule, where \"|\" is the sentence delimiter: S −→ Sc1|Sc2| · · · |Scn As UCCA allows argument sharing between Scenes, the rule may duplicate the same sub-span of S across sentences.", "For example, the rule will convert \"He came back home and played piano\" into \"He came back home\"|\"He played piano.\"", "Rule #2.", "Given a sentence S, the second rule extracts Elaborator Scenes and corresponding minimal centers.", "Elaborator Scenes are then concatenated to the original sentence, where the Elaborator Scenes, except for the minimal center they elaborate, are removed.", "Pronouns such as \"who\", \"which\" and \"that\" are also removed.", "Formally, if {(Sc 1 , C 1 ) · · · (Sc n , C n )} are the Elaborator Scenes of S and their corresponding minimal centers, the rewrite is: S −→ S − n i=1 (Sci − Ci)|Sc1| · · · |Scn where S −A is S without the unit A.", "For example, this rule converts the sentence \"He observed the planet which has 14 known satellites\" to \"He observed the planet| Planet has 14 known satellites.\".", "Article regeneration is not covered by the rule, as its output is directly fed into the NMT component.", "After the extraction of Parallel Scenes and Elaborator Scenes, the resulting simplified Parallel Scenes are placed before the Elaborator Scenes.", "See Figure 1 .", "Neural Component The split sentences are run through the NTS stateof-the-art neural TS system (Nisioi et al., 2017) , built using the OpenNMT neural machine translation framework (Klein et al., 2017) .", "The architecture includes two LSTM layers, with hidden states of 500 units in each, as well as global attention combined with input feeding (Luong et al., 2015) .", "Training is done with a 0.3 dropout probability (Srivastava et al., 2014) .", "This model uses alignment probabilities between the predictions and the original sentences, rather than characterbased models, to retrieve the original words.", "We here consider the w2v initialization for NTS (N17), where word2vec embeddings of size 300 are trained on Google News (Mikolov et al., 2013a) and local embeddings of size 200 are trained on the training simplification corpus (Řehůřek and Sojka, 2010; Mikolov et al., 2013b) .", "Local embeddings for the encoder are trained on the source side of the training corpus, while those for the decoder are trained on the simplified side.", "For sampling multiple outputs from the system, beam search is performed during decoding by generating the first 5 hypotheses at each step ordered by the log-likelihood of the target sentence given the input sentence.", "We here explore both the highest (h1) and fourth-ranked (h4) hypotheses, which we show to increase the SARI score and to be much less conservative.", "2 We thus experiment with two variants of the neural component, denoted by NTS-h1 and NTS-h4.", "The pipeline application of the rules and the neural system results in two corresponding models: SENTS-h1 and SENTS-h4.", "Experimental Setup Corpus All systems are tested on the test corpus of Xu et al.", "(2016) , 3 comprising 359 sentences from the PWKP corpus (Zhu et al., 2010) Neural component.", "We use the NTS-w2v model 6 provided by N17, obtained by training on the corpus of Hwang et al.", "(2015) and tuning on the corpus of Xu et al.", "(2016) .", "The training set is based on manual and automatic alignments between standard English Wikipedia and Simple English Wikipedia, including both good matches and partial matches whose similarity score is above the 0.45 scale threshold (Hwang et al., 2015) .", "The total size of the training set is about 280K aligned sentences, of which 150K sentences are full matches and 130K are partial matches.", "7 Comparison systems.", "We compare our findings to HYBRID, which is the state of the art for joint structural and lexical simplification, imple-mented by Zhang and Lapata (2017) .", "8 We use the released output of HYBRID, trained on a corpus extracted from Wikipedia, which includes the aligned sentence pairs from Kauchak (2013) , the aligned revision sentence pairs in Woodsend and Lapata (2011) , and the PWKP corpus, totaling about 296K sentence pairs.", "The tuning set is the same as for the above systems.", "In order to isolate the effect of NMT, we also implement SEMoses, where the neural-based component is replaced by the phrase-based MT system Moses, 9 which is also used in HYBRID.", "The training, tuning and test sets are the same as in the case of SENTS.", "MGIZA 10 is used for word alignment.", "The KenLM language model is trained using the target side of the training corpus.", "Additional baselines.", "We report human and automatic evaluation scores for Identity (where the output is identical to the input), for Simple Wikipedia where the output is the corresponding aligned sentence in the PWKP corpus, and for the SBMT-SARI system, tuned against SARI (Xu et al., 2016) , which maximized the SARI score on this test set in previous works (Nisioi et al., 2017; Zhang and Lapata, 2017) .", "Automatic evaluation.", "The automatic metrics used for the evaluation are: (1) BLEU (Papineni et al., 2002) (2) SARI (System output Against References and against the Input sentence; Xu et al., 2016) , which compares the n-grams of the system output with those of the input and the human references, separately evaluating the quality of words that are added, deleted and kept by the systems.", "(3) F add : the addition component of the SARI score (F-score); (4) F keep : the keeping component of the SARI score (F-score); (5) P del : the deletion component of the SARI score (precision).", "11 Each metric is computed against the 8 available references.", "We also assess system conservatism, reporting the percentage of sentences copied from the input (%Same), the averaged Levenshtein distance from the source (LD SC , which considers additions, deletions, and substitutions), and the number of source sentences that are split (#Split).", "12 Human evaluation.", "Human evaluation is carried out by 3 in-house native English annotators, who rated the different input-output pairs for the different systems according to 4 parameters: Grammaticality (G), Meaning preservation (M), Simplicity (S) and Structural Simplicity (StS).", "Each input-output pair is rated by all 3 annotators.", "Elicitation questions are given in Table 1 .", "As the selection process of the input-output pairs in the test corpus of Xu et al.", "(2016) , as well as their crowdsourced references, are explicitly biased towards lexical simplification, the use of human evaluation permits us to evaluate the structural aspects of the system outputs, even where structural operations are not attested in the references.", "Indeed, we show that system outputs may receive considerably higher structural simplicity scores than the source, in spite of the sample selection bias.", "Following previous work (e.g., Narayan and Gardent, 2014; Xu et al., 2016; Nisioi et al., 2017) , Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "Note that in the first question, the input sentence is not taken into account.", "The grammaticality of the input is assessed by evaluating the Identity transformation (see Table 2 ), providing a baseline for the grammaticality scores of the other systems.", "Following N17, a -2 to +2 scale is used for measuring simplicity, where a 0 score indicates that the input and the output are equally complex.", "This scale, compared to the standard 1 to 5 scale, permits a better differentiation between cases where simplicity is hurt (the output is more complex than the original) and between cases where the output is as simple as the original, for example in the case of the identity transformation.", "Structural simplicity is also evaluated with a -2 to +2 scale.", "The question for eliciting StS is accompanied with a negative example, showing a case of lexical simplification, where a complex word is replaced by a simple one (the other questions appear without examples).", "A positive example is not included so as not to bias the annotators by revealing the nature of the operations we focus on (splitting and deletion).", "We follow N17 in applying human evaluation on the first 70 sentences of the test corpus.", "13 The resulting corpus, totaling 1960 sentence pairs, each annotated by 3 annotators, also include the additional experiments described in Section 7 as well as the outputs of the NTS and SENTS systems used with the default initialization.", "The inter-annotator agreement, using Cohen's quadratic weighted κ (Cohen, 1968) , is computed as the average agreement of the 3 annotator pairs.", "The obtained rates are 0.56, 0.75, 0.47 and 0.48 for G, M, S and StS respectively.", "System scores are computed by averaging over the 3 annotators and the 70 sentences.", "G Is the output fluent and grammatical?", "M Does the output preserve the meaning of the input?", "S Is the output simpler than the input?", "StS Is the output simpler than the input, ignoring the complexity of the words?", "Table 2 : Human evaluation of the different NMT-based systems.", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "The highest score in each column appears in bold.", "Structural simplification systems are those that explicitly model structural operations.", "Results Human evaluation.", "Results are presented in Table 2 .", "First, we can see that the two SENTS systems outperform HYBRID in terms of G, M, and S. SENTS-h1 is the best scoring system, under all human measures.", "In comparison to NTS, SENTS scores markedly higher on the simplicity judgments.", "Meaning preservation and grammaticality are lower for SENTS, which is likely due to the more conservative nature of NTS.", "Interestingly, the application of the splitting rules by themselves does not yield a considerably simpler sentence.", "This likely stems from the rules not necessarily yielding grammatical sentences (NTS often serves as a grammatical error corrector over it), and from the incorporation of deletions, which are also structural operations, and are performed by the neural system.", "An example of high structural simplicity scores for SENTS resulting from deletions is presented in Table 5 , together with the outputs of the other systems and the corresponding human evaluation scores.", "NTS here performs lexical simplification, replacing the word \"incursions\" by \"raids\" or \"attacks\"'.", "On the other hand, the high StS scores obtained by DSS and SEMoses are due to sentence splittings.", "Automatic evaluation.", "Results are presented in Table 3 .", "Identity obtains much higher BLEU scores than any other system, suggesting that BLEU may not be informative in this setting.", "SARI seems more informative, and assigns the lowest score to Identity and the second highest to the reference.", "Both SENTS systems outperform HYBRID in terms of SARI and all its 3 sub-components.", "The h4 setting (hypothesis #4 in the beam) is generally best, both with and without the splitting rules.", "Comparing SENTS to using NTS alone (without splitting), we see that SENTS obtains higher SARI scores when hypothesis #1 is used and that NTS obtains higher scores when hypothesis #4 is used.", "This may result from NTS being more conservative than SENTS (and HYBRID), which is rewarded by SARI (conservatism is indicated by the %Same column).", "Indeed for h1, %Same is reduced from around 66% for NTS, to around 7% for SENTS.", "Conservatism further decreases when h4 is used (for both NTS and SENTS).", "Examining SARI's components, we find that SENTS outperforms NTS on F add , and is comparable (or even superior for h1 setting) to NTS on P del .", "The superior SARI score of NTS over SENTS is thus entirely a result of a superior F keep , which is easier for a conservative system to maximize.", "Comparing HYBRID with SEMoses, both of which use Moses, we find that SEMoses obtains higher BLEU and SARI scores, as well as G and M human scores, and splits many more sentences.", "HYBRID scores higher on the human simplicity measures.", "We note, however, that applying NTS alone is inferior to HYBRID in terms of simplicity, and that both components are required to obtain high simplicity scores (with SENTS).", "We also compare the sentence splitting component used in our systems (namely DSS) to that used in HYBRID, abstracting away from deletionbased and lexical simplification.", "We therefore apply DSS to the test set (554 sentences) of the Table 4 : Automatic and human evaluation for the different combinations of Moses and DSS.", "The automatic metrics as well as the lexical and structural properties reported (%Same: proportion of sentences copied from the input; LDSC: Averaged Levenshtein distance from the source; #Split: number of split sentences) concern the 359 sentences of the test corpus.", "Human evaluation, with the G, M, S, and StS parameters, is applied to the first 70 sentences of the corpus.", "The highest score in each column appears in bold.", "WEB-SPLIT corpus (See Section 2), which focuses on sentence splitting.", "We compare our results to those reported for a variant of HYBRID used without the deletion module, and trained on WEB-SPLIT .", "DSS gets a higher BLEU score (46.45 vs. 39.97) and performs more splittings (number of output sentences per input sentence of 1.73 vs. 1.26).", "Additional Experiments Replacing the parser by manual annotation.", "In order to isolate the influence of the parser on the results, we implement a semi-automatic version of the semantic component, which uses manual UCCA annotation instead of the parser, focusing of the first 70 sentences of the test corpus.", "We employ a single expert UCCA annotator and use the UCCAApp annotation tool .", "Results are presented in Table 6 , for both SENTS and SEMoses.", "In the case of SEMoses, meaning preservation is improved when manual UCCA annotation is used.", "On the other hand, simplicity degrades, possibly due to the larger number of Scenes marked by the human annotator (TUPA tends to under-predict Scenes).", "This effect doesn't show with SENTS, where trends are similar to the automatic parses case, and high simplicity scores are obtained.", "This demonstrates that UCCA parsing technology is sufficiently mature to be used to carry out structural simplification.", "We also directly evaluate the performance of the parser by computing F1, Recall and Precision DAG scores (Hershcovich et al., 2017) , against the manual UCCA annotation.", "14 We obtain for primary edges (i.e.", "edges that form a tree structure) scores of 68.9 %, 70.5%, and 67.4% for F1, Recall and Precision respectively.", "For remotes edges (i.e.", "additional edges, forming a DAG), the scores are 45.3%, 40.5%, and 51.5%.", "These results are comparable with the out-of-domain results reported by Hershcovich et al.", "(2017) .", "Experiments on Moses.", "We test other variants of SEMoses, where phrase-based MT is used instead of NMT.", "Specifically, we incorporate semantic information in a different manner by implementing two additional models: (1) SETrain1-Moses, where a new training corpus is obtained by applying the splitting rules to the target side of the G M S StS Identity In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the incursions of other Viking groups.", "5.00 5.00 0.00 0.00 Simple Wikipedia In return, Rollo swore fealty to Charles, converted to Christianity, and swore to defend the northern region of France against raids by other Viking groups.", "4.67 5.00 1.00 0.00 SBMT-SARI In return, Rollo swore fealty to Charles, converted to Christianity, and set out to defend the north of France from the raids of other viking groups.", "4.67 4.67 0.67 0.00 NTS-h1 In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the raids of other Viking groups.", "5.00 5.00 1.00 0.00 NTS-h4 In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the attacks of other Viking groups.", "4.67 5.00 1.00 0.00 DSS Rollo swore fealty to Charles.", "Rollo converted to Christianity.", "Rollo undertook to defend the northern region of France against the incursions of other viking groups.", "4.00 4.33 1.33 1.33 HYBRID In return Rollo swore, and undertook to defend the region of France., Charles, converted 2.33 2.00 0.33 0.33 SEMoses Rollo swore put his seal to Charles.", "Rollo converted to Christianity.", "Rollo undertook to defend the northern region of France against the incursions of other viking groups.", "3.33 4.00 1.33 1.33 SENTS-h1 Rollo swore fealty to Charles.", "5.00 2.00 2.00 2.00 SENTS-h4 Rollo swore fealty to Charles and converted to Christianity.", "5.00 2.67 1.33 1.33 Table 5 : System outputs for one of the test sentences with the corresponding human evaluation scores (averaged over the 3 annotators).", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "Table 6 : Human evaluation using manual UCCA annotation.", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "X m refers to the semi-automatic version of the system X. training corpus; (2) SETrain2-Moses, where the rules are applied to the source side.", "The resulting parallel corpus is concatenated to the original training corpus.", "We also examine whether training a language model (LM) on split sentences has a positive effect, and train the LM on the split target side.", "For each system X, the version with the LM trained on split sentences is denoted by X LM .", "We repeat the same human and automatic evaluation protocol as in §6, presenting results in Table 4 .", "Simplicity scores are much higher in the case of SENTS (that uses NMT), than with Moses.", "The two best systems according to SARI are SEMoses and SEMoses LM which use DSS.", "In fact, they resemble the performance of DSS applied alone (Tables 2 and 3) , which confirms the high degree of conservatism observed by Moses in simplification (Alva-Manchego et al., 2017) .", "Indeed, all Moses-based systems that don't apply DSS as preprocessing are conservative, obtaining high scores for BLEU, grammaticality and meaning preservation, but low scores for simplicity.", "Training the LM on split sentences shows little improvement.", "Conclusion We presented the first simplification system combining semantic structures and neural machine translation, showing that it outperforms existing lexical and structural systems.", "The proposed approach addresses the over-conservatism of MTbased systems for TS, which often fail to modify the source in any way.", "The semantic component performs sentence splitting without relying on a specialized corpus, but only an off-theshelf semantic parser.", "The consideration of sentence splitting as a decomposition of a sentence into its Scenes is further supported by recent work on structural TS evaluation (Sulem et al., 2018) , which proposes the SAMSA metric.", "The two works, which apply this assumption to different ends (TS system construction, and TS evaluation), confirm its validity.", "Future work will leverage UCCA's cross-linguistic applicability to support multi-lingual TS and TS pre-processing for MT." ] }

{ "paper_header_number": [ "1", "2", "3.1", "3.2", "4", "5", "6", "7", "8" ], "paper_header_content": [ "Introduction", "Related Work", "Semantic Representation", "The Semantic Rules", "Neural Component", "Experimental Setup", "Results", "Additional Experiments", "Conclusion" ] }

[]

{ "paper_content_id": [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266 ], "paper_content_text": [ "Introduction Text Simplification (TS) is generally defined as the conversion of a sentence into one or more simpler sentences.", "It has been shown useful both as a preprocessing step for tasks such as Machine Translation (MT; Mishra et al., 2014; Štajner and Popović, 2016) and relation extraction (Niklaus et al., 2016) , as well as for developing reading aids, e.g.", "for people with dyslexia (Rello et al., 2013) or non-native speakers (Siddharthan, 2002) .", "TS includes both structural and lexical operations.", "The main structural simplification operation is sentence splitting, namely rewriting a single sentence into multiple sentences while preserving its meaning.", "While recent improvement in TS has been achieved by the use of neural MT (NMT) approaches (Nisioi et al., 2017; Zhang and Lapata, 2017) , where TS is consid-ered a case of monolingual translation, the sentence splitting operation has not been addressed by these systems, potentially due to the rareness of this operation in the training corpora (Narayan and Gardent, 2014; Xu et al., 2015) .", "We show that the explicit integration of sentence splitting in the simplification system could also reduce conservatism, which is a grave limitation of NMT-based TS systems (Alva-Manchego et al., 2017) .", "Indeed, experimenting with a stateof-the-art neural system (Nisioi et al., 2017) , we find that 66% of the input sentences remain unchanged, while none of the corresponding references is identical to the source.", "Human and automatic evaluation of the references (against other references), confirm that the references are indeed simpler than the source, indicating that the observed conservatism is excessive.", "Our methods for performing sentence splitting as pre-processing allows the TS system to perform other structural (e.g.", "deletions) and lexical (e.g.", "word substitutions) operations, thus increasing both structural and lexical simplicity.", "For combining linguistically informed sentence splitting with data-driven TS, two main methods have been proposed.", "The first involves handcrafted syntactic rules, whose compilation and validation are laborious (Shardlow, 2014) .", "For example, Siddharthan and Angrosh (2014) used 111 rules for relative clauses, appositions, subordination and coordination.", "Moreover, syntactic splitting rules, which form a substantial part of the rules, are usually language specific, requiring the development of new rules when ported to other languages (Aluísio and Gasperin, 2010; Seretan, 2012; Hung et al., 2012; Barlacchi and Tonelli, 2013 , for Portuguese, French, Vietnamese, and Italian respectively).", "The second method uses linguistic information for detecting potential splitting points, while splitting probabilities are learned us-ing a parallel corpus.", "For example, in the system of Narayan and Gardent (2014) (henceforth, HYBRID) , the state-of-the-art for joint structural and lexical TS, potential splitting points are determined by event boundaries.", "In this work, which is the first to combine structural semantics and neural methods for TS, we propose an intermediate way for performing sentence splitting, presenting Direct Semantic Splitting (DSS), a simple and efficient algorithm based on a semantic parser which supports the direct decomposition of the sentence into its main semantic constituents.", "After splitting, NMT-based simplification is performed, using the NTS system.", "We show that the resulting system outperforms HY-BRID in both automatic and human evaluation.", "We use the UCCA scheme for semantic representation (Abend and Rappoport, 2013) , where the semantic units are anchored in the text, which simplifies the splitting operation.", "We further leverage the explicit distinction in UCCA between types of Scenes (events), applying a specific rule for each of the cases.", "Nevertheless, the DSS approach can be adapted to other semantic schemes, like AMR (Banarescu et al., 2013) .", "We collect human judgments for multiple variants of our system, its sub-components, HYBRID and similar systems that use phrase-based MT.", "This results in a sizable human evaluation benchmark, which includes 28 systems, totaling at 1960 complex-simple sentence pairs, each annotated by three annotators using four criteria.", "1 This benchmark will support the future analysis of TS systems, and evaluation practices.", "Previous work is discussed in §2, the semantic and NMT components we use in §3 and §4 respectively.", "The experimental setup is detailed in §5.", "Our main results are presented in §6, while §7 presents a more detailed analysis of the system's sub-components and related settings.", "Related Work MT-based sentence simplification.", "Phrasebased Machine Translation (PBMT; Koehn et al., 2003) was first used for TS by Specia (2010) , who showed good performance on lexical simplification and simple rewriting, but under-prediction of other operations.", "Štajner et al.", "(2015) took a similar approach, finding that it is beneficial to use training data where the source side is highly similar to the target.", "Other PBMT for TS systems include the work of Coster and Kauchak (2011b) , which uses Moses (Koehn et al., 2007) , the work of Coster and Kauchak (2011a) , where the model is extended to include deletion, and PBMT-R (Wubben et al., 2012) , where Levenshtein distance to the source is used for re-ranking to overcome conservatism.", "The NTS NMT-based system (Nisioi et al., 2017) (henceforth, N17) reported superior performance over PBMT in terms of BLEU and human evaluation scores, and serves as a component in our system (see Section 4).", "took a similar approach, adding lexical constraints to an NMT model.", "Zhang and Lapata (2017) combined NMT with reinforcement learning, using SARI (Xu et al., 2016) , BLEU, and cosine similarity to the source as the reward.", "None of these models explicitly addresses sentence splitting.", "Alva-Manchego et al.", "(2017) proposed to reduce conservatism, observed in PBMT and NMT systems, by first identifying simplification operations in a parallel corpus and then using sequencelabeling to perform the simplification.", "However, they did not address common structural operations, such as sentence splitting, and claimed that their method is not applicable to them.", "Xu et al.", "(2016) used Syntax-based Machine Translation (SBMT) for sentence simplification, using a large scale paraphrase dataset (Ganitketitch et al., 2013) for training.", "While it does not target structural simplification, we include it in our evaluation for completeness.", "Structural sentence simplification.", "Syntactic hand-crafted sentence splitting rules were proposed by Chandrasekar et al.", "(1996) , Siddharthan (2002) , Siddhathan (2011) in the context of rulebased TS.", "The rules separate relative clauses and coordinated clauses and un-embed appositives.", "In our method, the use of semantic distinctions instead of syntactic ones reduces the number of rules.", "For example, relative clauses and appositives can correspond to the same semantic category.", "In syntax-based splitting, a generation module is sometimes added after the split (Siddharthan, 2004) , addressing issues such as reordering and determiner selection.", "In our model, no explicit regeneration is applied to the split sentences, which are fed directly to an NMT system.", "Glavaš andŠtajner (2013) used a rule-based system conditioned on event extraction and syntax for defining two simplification models.", "The eventwise simplification one, which separates events to separate output sentences, is similar to our semantic component.", "Differences are in that we use a single semantic representation for defining the rules (rather than a combination of semantic and syntactic criteria), and avoid the need for complex rules for retaining grammaticality by using a subsequent neural component.", "Combined structural and lexical TS.", "Earlier TS models used syntactic information for splitting.", "Zhu et al.", "(2010) used syntactic information on the source side, based on the SBMT model of Yamada and Knight (2001) .", "Syntactic structures were used on both sides in the model of Woodsend and Lapata (2011) , based on a quasi-synchronous grammar (Smith and Eisner, 2006) , which resulted in 438 learned splitting rules.", "The model of Siddharthan and Angrosh (2014) is similar to ours in that it combines linguistic rules for structural simplification and statistical methods for lexical simplification.", "However, we use 2 semantic splitting rules instead of their 26 syntactic rules for relative clauses and appositions, and 85 syntactic rules for subordination and coordination.", "Narayan and Gardent (2014) argued that syntactic structures do not always capture the semantic arguments of a frame, which may result in wrong splitting boundaries.", "Consequently, they proposed a supervised system (HYBRID) that uses semantic structures (Discourse Semantic Representations, (Kamp, 1981) ) for sentence splitting and deletion.", "Splitting candidates are pairs of event variables associated with at least one core thematic role (e.g., agent or patient).", "Semantic annotation is used on the source side in both training and test.", "Lexical simplification is performed using the Moses system.", "HYBRID is the most similar system to ours architecturally, in that it uses a combination of a semantic structural component and an MT component.", "Narayan and Gardent (2016) proposed instead an unsupervised pipeline, where sentences are split based on a probabilistic model trained on the semantic structures of Simple Wikipedia as well as a language model trained on the same corpus.", "Lexical simplification is there performed using the unsupervised model of Biran et al.", "(2011) .", "As their BLEU and adequacy scores are lower than HYBRID's, we use the latter for comparison.", "Stajner and Glavaš (2017) combined rule-based simplification conditioned on event extraction, to-gether with an unsupervised lexical simplifier.", "They tackle a different setting, and aim to simplify texts (rather than sentences), by allowing the deletion of entire input sentences.", "Split and Rephrase.", "recently proposed the Split and Rephrase task, focusing on sentence splitting.", "For this purpose they presented a specialized parallel corpus, derived from the WebNLG dataset .", "The latter is obtained from the DBPedia knowledge base (Mendes et al., 2012) using content selection and crowdsourcing, and is annotated with semantic triplets of subject-relation-object, obtained semi-automatically.", "They experimented with five systems, including one similar to HY-BRID, as well as sequence-to-sequence methods for generating sentences from the source text and its semantic forms.", "The present paper tackles both structural and lexical simplification, and examines the effect of sentence splitting on the subsequent application of a neural system, in terms of its tendency to perform other simplification operations.", "For this purpose, we adopt a semantic corpus-independent approach for sentence splitting that can be easily integrated in any simplification system.", "Another difference is that the semantic forms in Split and Rephrase are derived semi-automatically (during corpus compilation), while we automatically extract the semantic form, using a UCCA parser.", "Direct Semantic Splitting Semantic Representation UCCA (Universal Cognitive Conceptual Annotation; Abend and Rappoport, 2013) is a semantic annotation scheme rooted in typological and cognitive linguistic theory (Dixon, 2010b (Dixon, ,a, 2012 Langacker, 2008) .", "It aims to represent the main semantic phenomena in the text, abstracting away from syntactic forms.", "UCCA has been shown to be preserved remarkably well across translations (Sulem et al., 2015) and has also been successfully used for the evaluation of machine translation (Birch et al., 2016) and, recently, for the evaluation of TS (Sulem et al., 2018) and grammatical error correction (Choshen and Abend, 2018) .", "Formally, UCCA structures are directed acyclic graphs whose nodes (or units) correspond either to the leaves of the graph or to several elements viewed as a single entity according to some semantic or cognitive consideration.", "A Scene is UCCA's notion of an event or a frame, and is a unit that corresponds to a movement, an action or a state which persists in time.", "Every Scene contains one main relation, which can be either a Process or a State.", "Scenes contain one or more Participants, interpreted in a broad sense to include locations and destinations.", "For example, the sentence \"He went to school\" has a single Scene whose Process is \"went\".", "The two Participants are \"He\" and \"to school\".", "Scenes can have several roles in the text.", "First, they can provide additional information about an established entity (Elaborator Scenes), commonly participles or relative clauses.", "For example, \"(child) who went to school\" is an Elaborator Scene in \"The child who went to school is John\" (\"child\" serves both as an argument in the Elaborator Scene and as the Center).", "A Scene may also be a Participant in another Scene.", "For example, \"John went to school\" in the sentence: \"He said John went to school\".", "In other cases, Scenes are annotated as Parallel Scenes (H), which are flat structures and may include a Linker (L), as in: \"When L [he arrives] H , [he will call them] H \".", "With respect to units which are not Scenes, the category Center denotes the semantic head.", "For example, \"dogs\" is the Center of the expression \"big brown dogs\", and \"box\" is the center of \"in the box\".", "There could be more than one Center in a unit, for example in the case of coordination, where all conjuncts are Centers.", "We define the minimal center of a UCCA unit u to be the UCCA graph's leaf reached by starting from u and iteratively selecting the child tagged as Center.", "For generating UCCA's structures we use TUPA, a transition-based parser (Hershcovich et al., 2017) (specifically, the TUPA BiLST M model).", "TUPA uses an expressive set of transitions, able to support all structural properties required by the UCCA scheme.", "Its transition classifier is based on an MLP that receives a BiLSTM encoding of elements in the parser state (buffer, stack and intermediate graph), given word embeddings and other features.", "The Semantic Rules For performing DSS, we define two simple splitting rules, conditioned on UCCA's categories.", "We currently only consider Parallel Scenes and Elaborator Scenes, not separating Participant Scenes, in order to avoid splitting in cases of nominalizations or indirect speech.", "For example, the sentence \"His arrival surprised everyone\", which has, in addition to the Scene evoked by \"surprised\", a Participant Scene evoked by \"arrival\", is not split here.", "Rule #1.", "Parallel Scenes of a given sentence are extracted, separated in different sentences, and concatenated according to the order of appearance.", "More formally, given a decomposition of a sentence S into parallel Scenes Sc 1 , Sc 2 , · · · Sc n (indexed by the order of the first token), we obtain the following rule, where \"|\" is the sentence delimiter: S −→ Sc1|Sc2| · · · |Scn As UCCA allows argument sharing between Scenes, the rule may duplicate the same sub-span of S across sentences.", "For example, the rule will convert \"He came back home and played piano\" into \"He came back home\"|\"He played piano.\"", "Rule #2.", "Given a sentence S, the second rule extracts Elaborator Scenes and corresponding minimal centers.", "Elaborator Scenes are then concatenated to the original sentence, where the Elaborator Scenes, except for the minimal center they elaborate, are removed.", "Pronouns such as \"who\", \"which\" and \"that\" are also removed.", "Formally, if {(Sc 1 , C 1 ) · · · (Sc n , C n )} are the Elaborator Scenes of S and their corresponding minimal centers, the rewrite is: S −→ S − n i=1 (Sci − Ci)|Sc1| · · · |Scn where S −A is S without the unit A.", "For example, this rule converts the sentence \"He observed the planet which has 14 known satellites\" to \"He observed the planet| Planet has 14 known satellites.\".", "Article regeneration is not covered by the rule, as its output is directly fed into the NMT component.", "After the extraction of Parallel Scenes and Elaborator Scenes, the resulting simplified Parallel Scenes are placed before the Elaborator Scenes.", "See Figure 1 .", "Neural Component The split sentences are run through the NTS stateof-the-art neural TS system (Nisioi et al., 2017) , built using the OpenNMT neural machine translation framework (Klein et al., 2017) .", "The architecture includes two LSTM layers, with hidden states of 500 units in each, as well as global attention combined with input feeding (Luong et al., 2015) .", "Training is done with a 0.3 dropout probability (Srivastava et al., 2014) .", "This model uses alignment probabilities between the predictions and the original sentences, rather than characterbased models, to retrieve the original words.", "We here consider the w2v initialization for NTS (N17), where word2vec embeddings of size 300 are trained on Google News (Mikolov et al., 2013a) and local embeddings of size 200 are trained on the training simplification corpus (Řehůřek and Sojka, 2010; Mikolov et al., 2013b) .", "Local embeddings for the encoder are trained on the source side of the training corpus, while those for the decoder are trained on the simplified side.", "For sampling multiple outputs from the system, beam search is performed during decoding by generating the first 5 hypotheses at each step ordered by the log-likelihood of the target sentence given the input sentence.", "We here explore both the highest (h1) and fourth-ranked (h4) hypotheses, which we show to increase the SARI score and to be much less conservative.", "2 We thus experiment with two variants of the neural component, denoted by NTS-h1 and NTS-h4.", "The pipeline application of the rules and the neural system results in two corresponding models: SENTS-h1 and SENTS-h4.", "Experimental Setup Corpus All systems are tested on the test corpus of Xu et al.", "(2016) , 3 comprising 359 sentences from the PWKP corpus (Zhu et al., 2010) Neural component.", "We use the NTS-w2v model 6 provided by N17, obtained by training on the corpus of Hwang et al.", "(2015) and tuning on the corpus of Xu et al.", "(2016) .", "The training set is based on manual and automatic alignments between standard English Wikipedia and Simple English Wikipedia, including both good matches and partial matches whose similarity score is above the 0.45 scale threshold (Hwang et al., 2015) .", "The total size of the training set is about 280K aligned sentences, of which 150K sentences are full matches and 130K are partial matches.", "7 Comparison systems.", "We compare our findings to HYBRID, which is the state of the art for joint structural and lexical simplification, imple-mented by Zhang and Lapata (2017) .", "8 We use the released output of HYBRID, trained on a corpus extracted from Wikipedia, which includes the aligned sentence pairs from Kauchak (2013) , the aligned revision sentence pairs in Woodsend and Lapata (2011) , and the PWKP corpus, totaling about 296K sentence pairs.", "The tuning set is the same as for the above systems.", "In order to isolate the effect of NMT, we also implement SEMoses, where the neural-based component is replaced by the phrase-based MT system Moses, 9 which is also used in HYBRID.", "The training, tuning and test sets are the same as in the case of SENTS.", "MGIZA 10 is used for word alignment.", "The KenLM language model is trained using the target side of the training corpus.", "Additional baselines.", "We report human and automatic evaluation scores for Identity (where the output is identical to the input), for Simple Wikipedia where the output is the corresponding aligned sentence in the PWKP corpus, and for the SBMT-SARI system, tuned against SARI (Xu et al., 2016) , which maximized the SARI score on this test set in previous works (Nisioi et al., 2017; Zhang and Lapata, 2017) .", "Automatic evaluation.", "The automatic metrics used for the evaluation are: (1) BLEU (Papineni et al., 2002) (2) SARI (System output Against References and against the Input sentence; Xu et al., 2016) , which compares the n-grams of the system output with those of the input and the human references, separately evaluating the quality of words that are added, deleted and kept by the systems.", "(3) F add : the addition component of the SARI score (F-score); (4) F keep : the keeping component of the SARI score (F-score); (5) P del : the deletion component of the SARI score (precision).", "11 Each metric is computed against the 8 available references.", "We also assess system conservatism, reporting the percentage of sentences copied from the input (%Same), the averaged Levenshtein distance from the source (LD SC , which considers additions, deletions, and substitutions), and the number of source sentences that are split (#Split).", "12 Human evaluation.", "Human evaluation is carried out by 3 in-house native English annotators, who rated the different input-output pairs for the different systems according to 4 parameters: Grammaticality (G), Meaning preservation (M), Simplicity (S) and Structural Simplicity (StS).", "Each input-output pair is rated by all 3 annotators.", "Elicitation questions are given in Table 1 .", "As the selection process of the input-output pairs in the test corpus of Xu et al.", "(2016) , as well as their crowdsourced references, are explicitly biased towards lexical simplification, the use of human evaluation permits us to evaluate the structural aspects of the system outputs, even where structural operations are not attested in the references.", "Indeed, we show that system outputs may receive considerably higher structural simplicity scores than the source, in spite of the sample selection bias.", "Following previous work (e.g., Narayan and Gardent, 2014; Xu et al., 2016; Nisioi et al., 2017) , Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "Note that in the first question, the input sentence is not taken into account.", "The grammaticality of the input is assessed by evaluating the Identity transformation (see Table 2 ), providing a baseline for the grammaticality scores of the other systems.", "Following N17, a -2 to +2 scale is used for measuring simplicity, where a 0 score indicates that the input and the output are equally complex.", "This scale, compared to the standard 1 to 5 scale, permits a better differentiation between cases where simplicity is hurt (the output is more complex than the original) and between cases where the output is as simple as the original, for example in the case of the identity transformation.", "Structural simplicity is also evaluated with a -2 to +2 scale.", "The question for eliciting StS is accompanied with a negative example, showing a case of lexical simplification, where a complex word is replaced by a simple one (the other questions appear without examples).", "A positive example is not included so as not to bias the annotators by revealing the nature of the operations we focus on (splitting and deletion).", "We follow N17 in applying human evaluation on the first 70 sentences of the test corpus.", "13 The resulting corpus, totaling 1960 sentence pairs, each annotated by 3 annotators, also include the additional experiments described in Section 7 as well as the outputs of the NTS and SENTS systems used with the default initialization.", "The inter-annotator agreement, using Cohen's quadratic weighted κ (Cohen, 1968) , is computed as the average agreement of the 3 annotator pairs.", "The obtained rates are 0.56, 0.75, 0.47 and 0.48 for G, M, S and StS respectively.", "System scores are computed by averaging over the 3 annotators and the 70 sentences.", "G Is the output fluent and grammatical?", "M Does the output preserve the meaning of the input?", "S Is the output simpler than the input?", "StS Is the output simpler than the input, ignoring the complexity of the words?", "Table 2 : Human evaluation of the different NMT-based systems.", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "The highest score in each column appears in bold.", "Structural simplification systems are those that explicitly model structural operations.", "Results Human evaluation.", "Results are presented in Table 2 .", "First, we can see that the two SENTS systems outperform HYBRID in terms of G, M, and S. SENTS-h1 is the best scoring system, under all human measures.", "In comparison to NTS, SENTS scores markedly higher on the simplicity judgments.", "Meaning preservation and grammaticality are lower for SENTS, which is likely due to the more conservative nature of NTS.", "Interestingly, the application of the splitting rules by themselves does not yield a considerably simpler sentence.", "This likely stems from the rules not necessarily yielding grammatical sentences (NTS often serves as a grammatical error corrector over it), and from the incorporation of deletions, which are also structural operations, and are performed by the neural system.", "An example of high structural simplicity scores for SENTS resulting from deletions is presented in Table 5 , together with the outputs of the other systems and the corresponding human evaluation scores.", "NTS here performs lexical simplification, replacing the word \"incursions\" by \"raids\" or \"attacks\"'.", "On the other hand, the high StS scores obtained by DSS and SEMoses are due to sentence splittings.", "Automatic evaluation.", "Results are presented in Table 3 .", "Identity obtains much higher BLEU scores than any other system, suggesting that BLEU may not be informative in this setting.", "SARI seems more informative, and assigns the lowest score to Identity and the second highest to the reference.", "Both SENTS systems outperform HYBRID in terms of SARI and all its 3 sub-components.", "The h4 setting (hypothesis #4 in the beam) is generally best, both with and without the splitting rules.", "Comparing SENTS to using NTS alone (without splitting), we see that SENTS obtains higher SARI scores when hypothesis #1 is used and that NTS obtains higher scores when hypothesis #4 is used.", "This may result from NTS being more conservative than SENTS (and HYBRID), which is rewarded by SARI (conservatism is indicated by the %Same column).", "Indeed for h1, %Same is reduced from around 66% for NTS, to around 7% for SENTS.", "Conservatism further decreases when h4 is used (for both NTS and SENTS).", "Examining SARI's components, we find that SENTS outperforms NTS on F add , and is comparable (or even superior for h1 setting) to NTS on P del .", "The superior SARI score of NTS over SENTS is thus entirely a result of a superior F keep , which is easier for a conservative system to maximize.", "Comparing HYBRID with SEMoses, both of which use Moses, we find that SEMoses obtains higher BLEU and SARI scores, as well as G and M human scores, and splits many more sentences.", "HYBRID scores higher on the human simplicity measures.", "We note, however, that applying NTS alone is inferior to HYBRID in terms of simplicity, and that both components are required to obtain high simplicity scores (with SENTS).", "We also compare the sentence splitting component used in our systems (namely DSS) to that used in HYBRID, abstracting away from deletionbased and lexical simplification.", "We therefore apply DSS to the test set (554 sentences) of the Table 4 : Automatic and human evaluation for the different combinations of Moses and DSS.", "The automatic metrics as well as the lexical and structural properties reported (%Same: proportion of sentences copied from the input; LDSC: Averaged Levenshtein distance from the source; #Split: number of split sentences) concern the 359 sentences of the test corpus.", "Human evaluation, with the G, M, S, and StS parameters, is applied to the first 70 sentences of the corpus.", "The highest score in each column appears in bold.", "WEB-SPLIT corpus (See Section 2), which focuses on sentence splitting.", "We compare our results to those reported for a variant of HYBRID used without the deletion module, and trained on WEB-SPLIT .", "DSS gets a higher BLEU score (46.45 vs. 39.97) and performs more splittings (number of output sentences per input sentence of 1.73 vs. 1.26).", "Additional Experiments Replacing the parser by manual annotation.", "In order to isolate the influence of the parser on the results, we implement a semi-automatic version of the semantic component, which uses manual UCCA annotation instead of the parser, focusing of the first 70 sentences of the test corpus.", "We employ a single expert UCCA annotator and use the UCCAApp annotation tool .", "Results are presented in Table 6 , for both SENTS and SEMoses.", "In the case of SEMoses, meaning preservation is improved when manual UCCA annotation is used.", "On the other hand, simplicity degrades, possibly due to the larger number of Scenes marked by the human annotator (TUPA tends to under-predict Scenes).", "This effect doesn't show with SENTS, where trends are similar to the automatic parses case, and high simplicity scores are obtained.", "This demonstrates that UCCA parsing technology is sufficiently mature to be used to carry out structural simplification.", "We also directly evaluate the performance of the parser by computing F1, Recall and Precision DAG scores (Hershcovich et al., 2017) , against the manual UCCA annotation.", "14 We obtain for primary edges (i.e.", "edges that form a tree structure) scores of 68.9 %, 70.5%, and 67.4% for F1, Recall and Precision respectively.", "For remotes edges (i.e.", "additional edges, forming a DAG), the scores are 45.3%, 40.5%, and 51.5%.", "These results are comparable with the out-of-domain results reported by Hershcovich et al.", "(2017) .", "Experiments on Moses.", "We test other variants of SEMoses, where phrase-based MT is used instead of NMT.", "Specifically, we incorporate semantic information in a different manner by implementing two additional models: (1) SETrain1-Moses, where a new training corpus is obtained by applying the splitting rules to the target side of the G M S StS Identity In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the incursions of other Viking groups.", "5.00 5.00 0.00 0.00 Simple Wikipedia In return, Rollo swore fealty to Charles, converted to Christianity, and swore to defend the northern region of France against raids by other Viking groups.", "4.67 5.00 1.00 0.00 SBMT-SARI In return, Rollo swore fealty to Charles, converted to Christianity, and set out to defend the north of France from the raids of other viking groups.", "4.67 4.67 0.67 0.00 NTS-h1 In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the raids of other Viking groups.", "5.00 5.00 1.00 0.00 NTS-h4 In return, Rollo swore fealty to Charles, converted to Christianity, and undertook to defend the northern region of France against the attacks of other Viking groups.", "4.67 5.00 1.00 0.00 DSS Rollo swore fealty to Charles.", "Rollo converted to Christianity.", "Rollo undertook to defend the northern region of France against the incursions of other viking groups.", "4.00 4.33 1.33 1.33 HYBRID In return Rollo swore, and undertook to defend the region of France., Charles, converted 2.33 2.00 0.33 0.33 SEMoses Rollo swore put his seal to Charles.", "Rollo converted to Christianity.", "Rollo undertook to defend the northern region of France against the incursions of other viking groups.", "3.33 4.00 1.33 1.33 SENTS-h1 Rollo swore fealty to Charles.", "5.00 2.00 2.00 2.00 SENTS-h4 Rollo swore fealty to Charles and converted to Christianity.", "5.00 2.67 1.33 1.33 Table 5 : System outputs for one of the test sentences with the corresponding human evaluation scores (averaged over the 3 annotators).", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "Table 6 : Human evaluation using manual UCCA annotation.", "Grammaticality (G) and Meaning preservation (M) are measured using a 1 to 5 scale.", "A -2 to +2 scale is used for measuring simplicity (S) and structural simplicity (StS) of the output relative to the input sentence.", "X m refers to the semi-automatic version of the system X. training corpus; (2) SETrain2-Moses, where the rules are applied to the source side.", "The resulting parallel corpus is concatenated to the original training corpus.", "We also examine whether training a language model (LM) on split sentences has a positive effect, and train the LM on the split target side.", "For each system X, the version with the LM trained on split sentences is denoted by X LM .", "We repeat the same human and automatic evaluation protocol as in §6, presenting results in Table 4 .", "Simplicity scores are much higher in the case of SENTS (that uses NMT), than with Moses.", "The two best systems according to SARI are SEMoses and SEMoses LM which use DSS.", "In fact, they resemble the performance of DSS applied alone (Tables 2 and 3) , which confirms the high degree of conservatism observed by Moses in simplification (Alva-Manchego et al., 2017) .", "Indeed, all Moses-based systems that don't apply DSS as preprocessing are conservative, obtaining high scores for BLEU, grammaticality and meaning preservation, but low scores for simplicity.", "Training the LM on split sentences shows little improvement.", "Conclusion We presented the first simplification system combining semantic structures and neural machine translation, showing that it outperforms existing lexical and structural systems.", "The proposed approach addresses the over-conservatism of MTbased systems for TS, which often fail to modify the source in any way.", "The semantic component performs sentence splitting without relying on a specialized corpus, but only an off-theshelf semantic parser.", "The consideration of sentence splitting as a decomposition of a sentence into its Scenes is further supported by recent work on structural TS evaluation (Sulem et al., 2018) , which proposes the SAMSA metric.", "The two works, which apply this assumption to different ends (TS system construction, and TS evaluation), confirm its validity.", "Future work will leverage UCCA's cross-linguistic applicability to support multi-lingual TS and TS pre-processing for MT." ] }

{ "paper_header_number": [ "1", "2", "3.1", "3.2", "4", "5", "6", "7", "8" ], "paper_header_content": [ "Introduction", "Related Work", "Semantic Representation", "The Semantic Rules", "Neural Component", "Experimental Setup", "Results", "Additional Experiments", "Conclusion" ] }

[]