Datasets:

sobamchan

/

aclsum

like 0

P06-1112

In this paper , we explore correlation of dependency relation paths to rank candidate answers in answer extraction . Using the correlation measure , we compare dependency relations of a candidate answer and mapped question phrases in sentence with the corresponding relations in question . Different from previous studies , we propose an approximate phrase mapping algorithm and incorporate the mapping score into the correlation measure . The correlations are further incorporated into a Maximum Entropy-based ranking model which estimates path weights from training . Experimental results show that our method significantly outperforms state-ofthe-art syntactic relation-based methods by up to 20 % in MRR . Answer Extraction is one of basic modules in open domain Question Answering ( QA ) . It is to further process relevant sentences extracted with Passage / Sentence Retrieval and pinpoint exact answers using more linguistic-motivated analysis . Since QA turns to find exact answers rather than text snippets in recent years , answer extraction becomes more and more crucial . Typically , answer extraction works in the following steps : • Recognize expected answer type of a question . • Annotate relevant sentences with various types of named entities . • Regard the phrases annotated with the expected answer type as candidate answers . • Rank candidate answers . In the above work flow , answer extraction heavily relies on named entity recognition ( NER ) . On one hand , NER reduces the number of candidate answers and eases answer ranking . On the other hand , the errors from NER directly degrade answer extraction performance . To our knowledge , most top ranked QA systems in TREC are supported by effective NER modules which may identify and classify more than 20 types of named entities ( NE ) , such as abbreviation , music , movie , etc . However , developing such named entity recognizer is not trivial . Up to now , we have n't found any paper relevant to QA-specific NER development . So , it is hard to follow their work . In this paper , we just use a general MUC-based NER , which makes our results reproducible . A general MUC-based NER ca n't annotate a large number of NE classes . In this case , all noun phrases in sentences are regarded as candidate answers , which makes candidate answer sets much larger than those filtered by a well developed NER . The larger candidate answer sets result in the more difficult answer extraction . Previous methods working on surface word level , such as density-based ranking and pattern matching , may not perform well . Deeper linguistic analysis has to be conducted . This paper proposes a statistical method which exploring correlation of dependency relation paths to rank candidate answers . It is motivated by the observation that relations between proper answers and question phrases in candidate sentences are always similar to the corresponding relations in question . For example , the question " What did Alfred Nobel invent ? " and the candidate sentence " ... in the will of Swedish industrialist Alfred Nobel , who invented dynamite . " For each question , firstly , dependency relation paths are defined and extracted from the question and each of its candidate sentences . Secondly , the paths from the question and the candidate sentence are paired according to question phrase mapping score . Thirdly , correlation between two paths of each pair is calculated by employing Dynamic Time Warping algorithm . The input of the calculation is correlations between dependency relations , which are estimated from a set of training path pairs . Lastly , a Maximum Entropy-based ranking model is proposed to incorporate the path correlations and rank candidate answers . Furthermore , sentence supportive measure are presented according to correlations of relation paths among question phrases . It is applied to re-rank the candidate answers extracted from the different candidate sentences . Considering phrases may provide more accurate information than individual words , we extract dependency relations on phrase level instead of word level . The experiment on TREC questions shows that our method significantly outperforms a densitybased method by 50 % in MRR and three stateof-the-art syntactic-based methods by up to 20 % in MRR . Furthermore , we classify questions by judging whether NER is used . We investigate how these methods perform on the two question sets . The results indicate that our method achieves better performance than the other syntactic-based methods on both question sets . Especially for more difficult questions , for which NER may not help , our method improves MRR by up to 31 % . The paper is organized as follows . Section 2 discusses related work and clarifies what is new in this paper . Section 3 presents relation path correlation in detail . Section 4 and 5 discuss how to incorporate the correlations for answer ranking and re-ranking . Section 6 reports experiment and results . In this paper , we propose a relation path correlation-based method to rank candidate answers in answer extraction . We extract and pair relation paths from questions and candidate sentences . Next , we measure the relation path correlation in each pair based on approximate phrase mapping score and relation sequence alignment , which is calculated by DTW algorithm . Lastly , a ME-based ranking model is proposed to incorporate the path correlations and rank candidate answers . The experiment on TREC questions shows that our method significantly outperforms a density-based method by 50 % in MRR and three state-of-the-art syntactic-based methods by up to 20 % in MRR . Furthermore , the method is especially effective for difficult questions , for which NER may not help . Therefore , it may be used to further enhance state-of-the-art QA systems even if they have a good NER . In the future , we are to further evaluate the method based on the overall performance of a QA system and adapt it to sentence retrieval task .

A generally accessible NER system for QA systems produces a larger answer candidate set which would be hard for current surface word-level ranking methods.

They propose a statistical method which takes correlations of dependency relation paths computed by the Dynamic Time Wrapping algorithm into account for ranking candidate answers.

The proposed method outperforms state-of-the-art syntactic relation-based methods by up to 20% and shows it works even better on harder questions where NER performs poorly.

2020.acl-main.528

Recently , many works have tried to augment the performance of Chinese named entity recognition ( NER ) using word lexicons . As a representative , Lattice-LSTM ( Zhang and Yang , 2018 ) has achieved new benchmark results on several public Chinese NER datasets . However , Lattice-LSTM has a complex model architecture . This limits its application in many industrial areas where real-time NER responses are needed . In this work , we propose a simple but effective method for incorporating the word lexicon into the character representations . This method avoids designing a complicated sequence modeling architecture , and for any neural NER model , it requires only subtle adjustment of the character representation layer to introduce the lexicon information . Experimental studies on four benchmark Chinese NER datasets show that our method achieves an inference speed up to 6.15 times faster than those of state-ofthe-art methods , along with a better performance . The experimental results also show that the proposed method can be easily incorporated with pre-trained models like BERT . 1 Named Entity Recognition ( NER ) is concerned with the identification of named entities , such as persons , locations , and organizations , in unstructured text . NER plays an important role in many downstream tasks , including knowledge base construction ( Riedel et al . , 2013 ) , information retrieval ( Chen et al . , 2015 ) , and question answering ( Diefenbach et al . , 2018 ) . In languages where words are naturally separated ( e.g. , English ) , NER has been conventionally formulated as a sequence labeling problem , and the state-of-the-art results have been achieved using neural-network-based models ( Huang et al . , 2015 ; Chiu and Nichols , 2016 ; Liu et al . , 2018 ) . Compared with NER in English , Chinese NER is more difficult since sentences in Chinese are not naturally segmented . Thus , a common practice for Chinese NER is to first perform word segmentation using an existing CWS system and then apply a word-level sequence labeling model to the segmented sentence ( Yang et al . , 2016 ; He and Sun , 2017b ) . However , it is inevitable that the CWS system will incorrectly segment query sentences . This will result in errors in the detection of entity boundary and the prediction of entity category in NER . Therefore , some approaches resort to performing Chinese NER directly at the character level , which has been empirically proven to be effective ( He and Wang , 2008 ; Liu et al . , 2010 ; Li et al . , 2014 ; Liu et al . , 2019 ; Sui et al . , 2019 ; Gui et al . , 2019b ; Ding et al . , 2019 ) . A drawback of the purely character-based NER method is that the word information is not fully exploited . With this consideration , Zhang and Yang , ( 2018 ) proposed Lattice-LSTM for incorporating word lexicons into the character-based NER model . Moreover , rather than heuristically choosing a word for the character when it matches multiple words in the lexicon , the authors proposed to preserve all words that match the character , leaving the subsequent NER model to determine which word to apply . To realize this idea , they introduced an elaborate modification to the sequence modeling layer of the LSTM-CRF model ( Huang et al . , 2015 ) . Experimental studies on four Chinese NER datasets have verified the effectiveness of Lattice-LSTM . However , the model architecture of Lattice-LSTM is quite complicated . In order to introduce lexicon information , Lattice-LSTM adds several additional edges between nonadjacent characters in the input sequence , which significantly slows its training and inference speeds . In addition , it is difficult to transfer the structure of Lattice-LSTM to other neural-network architectures ( e.g. , convolutional neural networks and transformers ) that may be more suitable for some specific tasks . In this work , we propose a simpler method to realize the idea of Lattice-LSTM , i.e. , incorporating all the matched words for each character to a character-based NER model . The first principle of our model design is to achieve a fast inference speed . To this end , we propose to encode lexicon information in the character representations , and we design the encoding scheme to preserve as much of the lexicon matching results as possible . Compared with Lattice-LSTM , our method avoids the need for a complicated model architecture , is easier to implement , and can be quickly adapted to any appropriate neural NER model by adjusting the character representation layer . In addition , ablation studies show the superiority of our method in incorporating more complete and distinct lexicon information , as well as introducing a more effective word-weighting strategy . The contributions of this work can be summarized as follows : • We propose a simple but effective method for incorporating word lexicons into the character representations for Chinese NER . • The proposed method is transferable to different sequence-labeling architectures and can be easily incorporated with pre-trained models like BERT ( Devlin et al . , 2018 ) . We performed experiments on four public Chinese NER datasets . The experimental results show that when implementing the sequence modeling layer with a single-layer Bi-LSTM , our method achieves considerable improvements over the state-of-theart methods in both inference speed and sequence labeling performance . In this work , we addressed the computational efficiency of utilizing word lexicons in Chinese NER . To obtain a high-performing Chinese NER system with a fast inference speed , we proposed a novel method to incorporate the lexicon information into the character representations . Experimental studies on four benchmark Chinese NER datasets reveal that our method can achieve a much faster inference speed and better performance than the compared state-of-the-art methods .

Named entity recognition in Chinese requires word segmentation causes errors or character-level model with lexical features that is complex and expensive.

They propose to encode lexicon features into character representations so it can keep the system simpler and achieve faster inference than previous models.

The proposed efficient character-based LSTM method with lexical features achieves 6.15 times faster inference speed and better performance than previous models.

P19-1352

Word embedding is central to neural machine translation ( NMT ) , which has attracted intensive research interest in recent years . In NMT , the source embedding plays the role of the entrance while the target embedding acts as the terminal . These layers occupy most of the model parameters for representation learning . Furthermore , they indirectly interface via a soft-attention mechanism , which makes them comparatively isolated . In this paper , we propose shared-private bilingual word embeddings , which give a closer relationship between the source and target embeddings , and which also reduce the number of model parameters . For similar source and target words , their embeddings tend to share a part of the features and they cooperatively learn these common representation units . Experiments on 5 language pairs belonging to 6 different language families and written in 5 different alphabets demonstrate that the proposed model provides a significant performance boost over the strong baselines with dramatically fewer model parameters . With the introduction of ever more powerful architectures , neural machine translation ( NMT ) has become the most promising machine translation method ( Kalchbrenner and Blunsom , 2013 ; Sutskever et al . , 2014 ; Bahdanau et al . , 2015 ) . For word representation , different architecturesincluding , but not limited to , recurrence-based ( Chen et al . , 2018 ) , convolution-based ( Gehring et al . , 2017 ) and transformation-based ( Vaswani et al . , 2017 ) NMT models-have been taking advantage of the distributed word embeddings to capture the syntactic and semantic properties of words ( Turian et al . , 2010 ) . Figure 1 : Comparison between ( a ) standard word embeddings and ( b ) shared-private word embeddings . In ( a ) , the English word " Long " and the German word " Lange " , which have similar lexical meanings , are represented by two private d-dimension vectors . While in ( b ) , the two word embeddings are made up of two parts , indicating the shared ( lined nodes ) and the private ( unlined nodes ) features . This enables the two words to make use of common representation units , leading to a closer relationship between them . NMT usually utilizes three matrices to represent source embeddings , target input embeddings , and target output embeddings ( also known as pre-softmax weight ) , respectively . These embeddings occupy most of the model parameters , which constrains the improvements of NMT because the recent methods become increasingly memory-hungry ( Vaswani et al . , 2017 ; Chen et al . , 2018 ) . 1 Even though converting words into subword units ( Sennrich et al . , 2016b ) , nearly 55 % of model parameters are used for word representation in the Transformer model ( Vaswani et al . , 2017 ) . To overcome this difficulty , several methods are proposed to reduce the parameters used for word representation of NMT . Press and Wolf ( 2017 ) propose two weight tying ( WT ) methods , called decoder WT and three-way WT , to substantially reduce the parameters of the word embeddings . Decoder WT ties the target input embedding and target output embedding , which has become the new de facto standard of practical NMT ( Sen- Figure 2 : Shared-private bilingual word embeddings perform between the source and target words or sub-words ( a ) with similar lexical meaning , ( b ) with same word form , and ( c ) without any relationship . Different sharing mechanisms are adapted into different relationship categories . This strikes the right balance between capturing monolingual and bilingual characteristics . The closeness of relationship decides the portion of features to be used for sharing . Words with similar lexical meaning tend to share more features , followed by the words with the same word form , and then the unrelated words , as illustrated by the lined nodes . nrich et al . , 2017 ) . Three-way WT uses only one matrix to represent the three word embeddings , where the source and target words that have the same word form tend to share a word vector . This method can also be adapted to sub-word NMT with a shared source-target sub-word vocabulary and it performs well in language pairs with many of the same characters , such as English-German and English-French ( Vaswani et al . , 2017 ) . Unfortunately , this method is not applicable to languages that are written in different alphabets , such as Chinese-English ( Hassan et al . , 2018 ) . Another challenge facing the source and target word embeddings of NMT is the lack of interactions . This degrades the attention performance , leading to some unaligned translations that hurt the translation quality . Hence , Kuang et al . ( 2018 ) propose to bridge the source and target embeddings , which brings better attention to the related source and target words . Their method is applicable to any language pairs , providing a tight interaction between the source and target word pairs . However , their method requires additional components and model parameters . In this work , we aim to enhance the word representations and the interactions between the source and target words , while using even fewer parameters . To this end , we present a languageindependent method , which is called sharedprivate bilingual word embeddings , to share a part of the embeddings of a pair of source and target words that have some common characteristics ( i.e. similar words should have similar vectors ) . Figure 1 illustrates the difference between the standard word embeddings and shared-private word embeddings of NMT . In the proposed method , each source ( or target ) word is represented by a word embedding that consists of the shared features and the private features . The shared features can also be regarded as the prior alignments connecting the source and target words . The private features allow the words to better learn the monolingual characteristics . Meanwhile , the features shared by the source and target embeddings result in a significant reduction of the number of parameters used for word representations . The experimental results on 6 translation datasets of different scales show that our model with fewer parameters yields consistent improvements over the strong Transformer baselines . In this work , we propose a novel sharing technique to improve the learning of word embeddings for NMT . Each word embedding is composed of shared and private features . The shared features act as a prior alignment guidance for the attention model to improve the quality of attention . Meanwhile , the private features enable the words to better capture the monolingual characteristics , result in an improvement of the overall translation quality . According to the degree of relevance between a parallel word pair , the word pairs are categorized into three different groups and the number of shared features is different . Our experimental results show that the proposed method outperforms the strong Transformer baselines while using fewer model parameters .

Word embeddings occupy a large amount of memory, and weight tying does not mitigate this issue for distant language pairs on translation tasks.

They propose a language independet method where a model shares embeddings between source and target only when words have some common characteristics.

Experiments on machine translation datasets involving multiple language families and scripts show that the proposed model outperforms baseline models while using fewer parameters.

D12-1061

This paper explores log-based query expansion ( QE ) models for Web search . Three lexicon models are proposed to bridge the lexical gap between Web documents and user queries . These models are trained on pairs of user queries and titles of clicked documents . Evaluations on a real world data set show that the lexicon models , integrated into a ranker-based QE system , not only significantly improve the document retrieval performance but also outperform two state-of-the-art log-based QE methods . Term mismatch is a fundamental problem in Web search , where queries and documents are composed using different vocabularies and language styles . Query expansion ( QE ) is an effective strategy to address the problem . It expands a query issued by a user with additional related terms , called expansion terms , so that more relevant documents can be retrieved . In this paper we explore the use of clickthrough data and translation models for QE . We select expansion terms for a query according to how likely it is that the expansion terms occur in the title of a document that is relevant to the query . Assuming that a query is parallel to the titles of documents clicked for that query ( Gao et al . 2010a ) , three lexicon models are trained on query-title pairs extracted from clickthrough data . The first is a word model that learns the translation probability between single words . The second model uses lexi-calized triplets to incorporate word dependencies for translation . The third is a bilingual topic model , which represents a query as a distribution of hidden topics and learns the translation between a query and a title term at the semantic level . We will show that the word model provides a rich set of expansion candidates while the triplet and topic models can effectively select good expansion terms , and that a ranker-based QE system which incorporates all three of these models not only significantly improves Web search result but outperforms other log-based QE methods that are stateof-the-art . There is growing interest in applying user logs to improve QE . A recent survey is due to Baeze-Yates and Ribeiro-Neto ( 2011 ) . Below , we briefly discuss two log-based QE methods that are closest to ours and are re-implemented in this study for comparison . Both systems use the same type of log data that we used to train the lexicon models . The term correlation model of Cui et al . ( 2002 ; 2003 ) is to our knowledge the first to explore querydocument relations for direct extraction of expansion terms for Web search . The method outperforms traditional QE methods that do not use log data e.g. the local analysis model of Xu and Croft ( 1996 ) . In addition , as pointed out by Cui et al . ( 2003 ) there are three important advantages that make log-based QE a promising technology to improve the performance of commercial search engines . First , unlike traditional QE methods that are based on relevance feedback , log-based QE derives expansion terms from search logs , allowing term correlations to be pre-computed offline . Compared to methods that are based on thesauri either compiled manually ( Prager et al . 2001 ) or derived au-tomatically from document collections ( Jing and Croft 1994 ) , the log-based method is superior in that it explicitly captures the correlation between query terms and document terms , and thus can bridge the lexical gap between them more effectively . Second , since search logs retrain querydocument pairs clicked by millions of users , the term correlations reflect the preference of the majority of users . Third , the term correlations evolve along with the accumulation of user logs , thus can reflect updated user interests at a specific time . However , as pointed out by Riezler et al . ( 2008 ) , Cui et al . 's correlation-based method suffers low precision of QE partly because the correlation model does not explicitly capture context information and is susceptible to noise . Riezler et al . developed a QE system by retraining a standard phrase-based statistical machine translation ( SMT ) system using query-snippet pairs extracted from clickthrough data ( Riezler et al . 2008 ; Riezler and Liu 2010 ) . The SMT-based system can produce cleaner , more relevant expansion terms because rich context information useful for filtering noisy expansions is captured by combining language model and phrase translation model in its decoder . Furthermore , in the SMT system all component models are properly smoothed using sophisticated techniques to avoid sparse data problems while the correlation model relies on pure counts of term frequencies . However , the SMT system is used as a black box in their experiments . So the relative contribution of different SMT components is not verified empirically . In this study we break this black box in order to build a better , simpler QE system . We will show that the proposed lexicon models outperform significantly the term correlation model , and that a simpler QE system that incorporates the lexicon models can beat the sophisticated , black-box SMT system . In this paper we extend the previous log-based QE methods in two directions . First , we formulate QE as the problem of translating a source language of queries into a target language of documents , represented as titles . This allows us to adapt the established techniques developed for SMT to QE . Specially , we propose three lexicon models based on terms , lexicalized triplets , and topics , respectively . These models are trained on pairs of user queries and the titles of clicked documents using EM . Second , we present a ranker-based QE system , the heart of which is a MRF-based ranker in which the lexicon models are incorporated as features . We perform experiments on the Web search task using a real world data set . Results show that the proposed system outperforms significantly other stateof-the-art QE systems . This study is part of a bigger , ongoing project , aiming to develop a real-time QE system for Web search , where simplicity is the key to the success . Thus , what we learned from this study is particularly encouraging . We demonstrate that with large amounts of clickthrough data for model training , simple lexicon models can achieve state-of-the-art QE performance , and that the MRF-based ranker provides a simple and flexible framework to incorporate a variety of features capturing different types of term dependencies in such an effective way that the Web search performance can be directly optimized .

Term mismatches between a query and documents hinder retrievals of relevant documents and black box statistical machine translation models are used to expand queries.

They propose to train lexicon query expansion models by using transaction logs that contain pairs of queries and titles of clicked documents.

The proposed query expansion model enables retrieval systems to significantly outperform models with previous expansion models while being more transparent.

N07-1011

Traditional noun phrase coreference resolution systems represent features only of pairs of noun phrases . In this paper , we propose a machine learning method that enables features over sets of noun phrases , resulting in a first-order probabilistic model for coreference . We outline a set of approximations that make this approach practical , and apply our method to the ACE coreference dataset , achieving a 45 % error reduction over a comparable method that only considers features of pairs of noun phrases . This result demonstrates an example of how a firstorder logic representation can be incorporated into a probabilistic model and scaled efficiently . Noun phrase coreference resolution is the problem of clustering noun phrases into anaphoric sets . A standard machine learning approach is to perform a set of independent binary classifications of the form " Is mention a coreferent with mention b ? " This approach of decomposing the problem into pairwise decisions presents at least two related difficulties . First , it is not clear how best to convert the set of pairwise classifications into a disjoint clustering of noun phrases . The problem stems from the transitivity constraints of coreference : If a and b are coreferent , and b and c are coreferent , then a and c must be coreferent . This problem has recently been addressed by a number of researchers . A simple approach is to perform the transitive closure of the pairwise decisions . However , as shown in recent work ( McCallum and Wellner , 2003 ; Singla and Domingos , 2005 ) , better performance can be obtained by performing relational inference to directly consider the dependence among a set of predictions . For example , McCallum and Wellner ( 2005 ) apply a graph partitioning algorithm on a weighted , undirected graph in which vertices are noun phrases and edges are weighted by the pairwise score between noun phrases . A second and less studied difficulty is that the pairwise decomposition restricts the feature set to evidence about pairs of noun phrases only . This restriction can be detrimental if there exist features of sets of noun phrases that can not be captured by a combination of pairwise features . As a simple example , consider prohibiting coreferent sets that consist only of pronouns . That is , we would like to require that there be at least one antecedent for a set of pronouns . The pairwise decomposition does not make it possible to capture this constraint . In general , we would like to construct arbitrary features over a cluster of noun phrases using the full expressivity of first-order logic . Enabling this sort of flexible representation within a statistical model has been the subject of a long line of research on first-order probabilistic models ( Gaifman , 1964 ; Halpern , 1990 ; Paskin , 2002 ; Poole , 2003 ; Richardson and Domingos , 2006 ) . Conceptually , a first-order probabilistic model can be described quite compactly . A configuration of the world is represented by a set of predi- cates , each of which has an associated real-valued parameter . The likelihood of each configuration of the world is proportional to a combination of these weighted predicates . In practice , however , enumerating all possible configurations , or even all the predicates of one configuration , can result in intractable combinatorial growth ( de Salvo Braz et al . , 2005 ; Culotta and McCallum , 2006 ) . In this paper , we present a practical method to perform training and inference in first-order models of coreference . We empirically validate our approach on the ACE coreference dataset , showing that the first-order features can lead to an 45 % error reduction . We have presented learning and inference procedures for coreference models using first-order features . By relying on sampling methods at training time and approximate inference methods at testing time , this approach can be made scalable . This results in a coreference model that can capture features over sets of noun phrases , rather than simply pairs of noun phrases . This is an example of a model with extremely flexible representational power , but for which exact inference is intractable . The simple approximations we have described here have enabled this more flexible model to outperform a model that is simplified for tractability . A short-term extension would be to consider features over entire clusterings , such as the number of clusters . This could be incorporated in a ranking scheme , as in Ng ( 2005 ) . Future work will extend our approach to a wider variety of tasks . The model we have described here is specific to clustering tasks ; however a similar formulation could be used to approach a number of language processing tasks , such as parsing and relation extraction . These tasks could benefit from first-order features , and the present work can guide the approximations required in those domains . Additionally , we are investigating more sophisticated inference algorithms that will reduce the greediness of the search procedures described here .

Existing approaches treat noun phrase coreference resolution as a set of independent binary classifications limiting the features to be only pairs of noun phrases.

They propose a machine learning method that uses sets of noun phrases as features that are coupled with a sampling method to enable scalability.

Evaluation on the ACE coreference dataset, the proposed method achieves a 45% error reduction over a previous method.

2021.acl-long.67

Bilingual lexicons map words in one language to their translations in another , and are typically induced by learning linear projections to align monolingual word embedding spaces . In this paper , we show it is possible to produce much higher quality lexicons with methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . Directly applying a pipeline that uses recent algorithms for both subproblems significantly improves induced lexicon quality and further gains are possible by learning to filter the resulting lexical entries , with both unsupervised and semisupervised schemes . Our final model outperforms the state of the art on the BUCC 2020 shared task by 14 F 1 points averaged over 12 language pairs , while also providing a more interpretable approach that allows for rich reasoning of word meaning in context . Further analysis of our output and the standard reference lexicons suggests they are of comparable quality , and new benchmarks may be needed to measure further progress on this task . 1 Bilingual lexicons map words in one language to their translations in another , and can be automatically induced by learning linear projections to align monolingual word embedding spaces ( Artetxe et al . , 2016 ; Smith et al . , 2017 ; Lample et al . , 2018 , inter alia ) . Although very successful in practice , the linear nature of these methods encodes unrealistic simplifying assumptions ( e.g. all translations of a word have similar embeddings ) . In this paper , we show it is possible to produce much higher quality lexicons without these restrictions by introducing new methods that combine ( 1 ) unsupervised bitext mining and ( 2 ) unsupervised word alignment . We show that simply pipelining recent algorithms for unsupervised bitext mining ( Tran et al . , 2020 ) and unsupervised word alignment ( Sabet et al . , 2020 ) significantly improves bilingual lexicon induction ( BLI ) quality , and that further gains are possible by learning to filter the resulting lexical entries . Improving on a recent method for doing BLI via unsupervised machine translation ( Artetxe et al . , 2019 ) , we show that unsupervised mining produces better bitext for lexicon induction than translation , especially for less frequent words . These core contributions are established by systematic experiments in the class of bitext construction and alignment methods ( Figure 1 ) . Our full induction algorithm filters the lexicon found via the initial unsupervised pipeline . The filtering can be either fully unsupervised or weakly-supervised : for the former , we filter using simple heuristics and global statistics ; for the latter , we train a multi-layer perceptron ( MLP ) to predict the probability of a word pair being in the lexicon , where the features are global statistics of word alignments . In addition to BLI , our method can also be directly adapted to improve word alignment and reach competitive or better alignment accuracy than the state of the art on all investigated language pairs . We find that improved alignment in sentence representations ( Tran et al . , 2020 ) leads to better contextual word alignments using local similarity ( Sabet et al . , 2020 ) . Our final BLI approach outperforms the previous state of the art on the BUCC 2020 shared task ( Rapp et al . , 2020 ) by 14 F 1 points averaged over 12 language pairs . Manual analysis shows that most of our false positives are due to the incompleteness of the reference and that our lexicon is comparable to the reference lexicon and the output of a supervised system . Because both of our key building blocks make use of the pretrainined contextual representations from mBART ( Liu et al . , Word Alignment Statistical Feature Extraction cooccurrence(good , guten ) = 2 one-to-one align(good , guten ) = 2 many-to-one align(good , guten ) = 0 cosine_similarity(good , guten ) = 0.8 inner_product(good , guten ) = 1.8 count(good ) = 2 count(guten ) = 2 We present a direct and effective framework for BLI with unsupervised bitext mining and word alignment , which sets a new state of the art on the task . From the perspective of pretrained multilingual models ( Conneau et al . , 2019 ; Liu et al . , 2020 ; Tran et al . , 2020 , inter alia ) , our work shows that they have successfully captured information about word translation that can be extracted using similarity based alignment and refinement . Although BLI is only about word types , it strongly benefits from contextualized reasoning at the token level . umich.edu/ ˜mihalcea / wpt ( en-fr and ro-en ) ; https : //web.eecs.umich.edu/ ˜mihalcea / wpt05 ( enhi )

Existing methods to induce bilingual lexicons use linear projections to align word embeddings that are based on unrealistic simplifying assumptions.

They propose to use both unsupervised bitext mining and unsupervised word alignment methods to produce higher quality lexicons.

The proposed method achieves the state-of-the-art in the bilingual lexical induction task while keeping the interpretability of their pipeline.

D18-1065

In this paper we show that a simple beam approximation of the joint distribution between attention and output is an easy , accurate , and efficient attention mechanism for sequence to sequence learning . The method combines the advantage of sharp focus in hard attention and the implementation ease of soft attention . On five translation and two morphological inflection tasks we show effortless and consistent gains in BLEU compared to existing attention mechanisms . In structured input-output models as used in tasks like translation and image captioning , the attention variable decides which part of the input aligns to the current output . Many attention mechanisms have been proposed ( Xu et al . , 2015 ; Bahdanau et al . , 2014 ; Luong et al . , 2015 ; Martins and Astudillo , 2016 ) but the de facto standard is a soft attention mechanism that first assigns attention weights to input encoder states , then computes an attention weighted ' soft ' aligned input state , which finally derives the output distribution . This method is end to end differentiable and easy to implement . Another less popular variant is hard attention that aligns each output to exactly one input state but requires intricate training to teach the network to choose that state . When successfully trained , hard attention is often found to be more accurate ( Xu et al . , 2015 ; Zaremba and Sutskever , 2015 ) . In NLP , a recent success has been in a monotonic hard attention setting in morphological inflection tasks ( Yu et al . , 2016 ; Aharoni and Goldberg , 2017 ) . For general seq2seq learning , methods like Sparse-Max ( Martins and Astudillo , 2016 ) and local attention ( Luong et al . , 2015 ) were proposed to bridge the gap between soft and hard attention . * Both authors contributed equally to this work In this paper we propose a surprisingly simpler alternative based on the original joint distribution between output and attention , of which existing soft and hard attention mechanisms are approximations . The joint model couples input states individually to the output like in hard attention , but it combines the advantage of end-to-end trainability of soft attention . When the number of input states is large , we propose to use a simple approximation of the full joint distribution called Beam-joint . This approximation is also easily trainable and does not suffer from the high variance of Monte-Carlo sampling gradients of hard attention . We evaluated our model on five translation tasks and increased BLEU by 0.8 to 1.7 over soft attention , which in turn was better than hard and the recent Sparsemax ( Martins and Astudillo , 2016 ) attention . More importantly , the training process was as easy as soft attention . For further support , we also evaluate on two morphological inflection tasks and got gains over soft and hard attention . In this paper we showed a simple yet effective approximation of the joint attention-output distribution in sequence to sequence learning . Our joint model consistently provides higher accuracy without significant running time overheads in five translation and two morphological inflection tasks . An interesting direction for future work is to extend beam-joint to multi-head attention architectures as in ( Vaswani et al . , 2017 ; Xu Chen , 2018 ) .

Softmax attention models are popular because of their differentiable and easy to implement nature while hard attention models outperform them when successfully trained.

They propose a method to approximate the joint attention-output distribution which provides sharp attention as hard attention and easy implementation as soft attention.

The proposed approach outperforms soft attention models and recent hard attention and Sparsemax models on five translation tasks and also on morphological inflection tasks.

2022.acl-long.304

Contrastive learning has achieved impressive success in generation tasks to militate the " exposure bias " problem and discriminatively exploit the different quality of references . Existing works mostly focus on contrastive learning on the instance-level without discriminating the contribution of each word , while keywords are the gist of the text and dominant the constrained mapping relationships . Hence , in this work , we propose a hierarchical contrastive learning mechanism , which can unify hybrid granularities semantic meaning in the input text . Concretely , we first propose a keyword graph via contrastive correlations of positive-negative pairs to iteratively polish the keyword representations . Then , we construct intra-contrasts within instance-level and keyword-level , where we assume words are sampled nodes from a sentence distribution . Finally , to bridge the gap between independent contrast levels and tackle the common contrast vanishing problem , we propose an inter-contrast mechanism that measures the discrepancy between contrastive keyword nodes respectively to the instance distribution . Experiments demonstrate that our model outperforms competitive baselines on paraphrasing , dialogue generation , and storytelling tasks . Generation tasks such as storytelling , paraphrasing , and dialogue generation aim at learning a certain correlation between text pairs that maps an arbitrary-length input to another arbitrary-length output . Traditional methods are mostly trained with " teacher forcing " and lead to an " exposure bias " problem ( Schmidt , 2019 ) . Incorporating the generation method with contrastive learning achieved impressive performance on tackling such issues , which takes an extra consideration of synthetic negative samples contrastively ( Lee et al . , 2021 Existing contrastive mechanisms are mainly focused on the instance level ( Lee et al . , 2021 ; Cai et al . , 2020 ) . However , word-level information is also of great importance . Take the case shown in the upper part of Figure 1 for example , the keyword covers the gist of the input text and determines the embedding space of the text . The text representation will be significantly affected if adding a slight perturbation on the keyword , i.e. , changing " cosmology " to " astrophysics " . In addition , as shown on the bottom part , under some circumstances , it is too easy for the model to do the classification since the semantic gap between contrastive pairs is huge . Thus , the model fails to distinguish the actual discrepancy , which causes a " contrast vanishing " problem at both instance-level and keyword-level . Based on the above motivation , in this paper , we propose a hierarchical contrastive learning method built on top of the classic CVAE structure . We choose CVAE due to its ability in modeling global properties such as syntactic , semantic , and discourse coherence ( Li et al . , 2015 ; Yu et al . , 2020 ) . We first learn different granularity representations through two independent contrast , i.e. , instancelevel and keyword-level . Specifically , we use the universal and classic TextRank ( Mihalcea and Tarau , 2004 ) method to extract keywords from each text , which contain the most important information and need to be highlighted . On the instancelevel , we treat the keyword in the input text as an additional condition for a better prior semantic distribution . Then , we utilize Kullback-Leibler divergence ( Kullback and Leibler , 1951 ) to reduce the distance between prior distribution and positive posterior distribution , and increase the distance with the negative posterior distribution . While on the keyword-level , we propose a keyword graph via contrastive correlations of positive-negative pairs to learn informative and accurate keyword representations . By treating the keyword in the output text as an anchor , the imposter keyword is produced by neighboring nodes of the anchor keyword and forms the keyword-level contrast , where the similarity between the imposter keyword and the anchor keyword is poorer than the positive keyword . To unify individual intra-contrasts and tackle the " contrast vanishing " problem in independent contrastive granularities , we leverage an inter-contrast , the Mahalanobis contrast , to investigate the contrastive enhancement based on the Mahalanobis distance ( De Maesschalck et al . , 2000 ) , a measure of the distance between a point and a distribution , between the instance distribution and the keyword representation . Concretely , we ensure the distance from the anchor instance distribution to the groundtruth keyword vector is closer than to the imposter keyword vector . The Mahalanobis contrast plays an intermediate role that joins the different granularities contrast via incorporating the distribution of instance with the representation of its crucial part , and makes up a more comprehensive keyworddriven hierarchical contrastive mechanism , so as to ameliorate the generated results . We empirically show that our model outperforms CVAE and other baselines significantly on three generation tasks : paraphrasing , dialogue genera-tion , and storytelling . Our contributions can be summarized as follows : • To our best knowledge , we are the first to propose an inter-level contrastive learning method , which unifies instance-level and keyword-level contrasts in the CVAE framework . • We propose three contrastive learning measurements : KL divergence for semantic distribution , cosine distance for points , and Mahalanobis distance for points with distribution . • We introduce a global keyword graph to obtain polished keyword representations and construct imposter keywords for contrastive learning . In this paper , we propose a hierarchical contrastive learning mechanism , which consists of intra-contrasts within instance-level and keywordlevel and inter-contrast with Mahalanobis contrast . The experimental results yield significant out-performance over baselines when applied in the CVAE framework . In the future , we aim to extend the contrastive learning mechanism to different basic models , and will explore contrastive learning methods based on external knowledge .

Existing works on contrastive learning for text generation focus only on instance-level while word-level information such as keywords is also of great importance.

They propose a CVAE-based hierarchical contrastive learning within instance and keyword-level using a keyword graph which iteratively polishes the keyword representations.

The proposed model outperforms CVAE and baselines on storytelling, paraphrasing, and dialogue generation tasks.

2020.emnlp-main.384

Word embedding models are typically able to capture the semantics of words via the distributional hypothesis , but fail to capture the numerical properties of numbers that appear in a text . This leads to problems with numerical reasoning involving tasks such as question answering . We propose a new methodology to assign and learn embeddings for numbers . Our approach creates Deterministic , Independentof-Corpus Embeddings ( referred to as DICE ) for numbers , such that their cosine similarity reflects the actual distance on the number line . DICE outperforms a wide range of pre-trained word embedding models across multiple examples of two tasks : ( i ) evaluating the ability to capture numeration and magnitude ; and ( ii ) to perform list maximum , decoding , and addition . We further explore the utility of these embeddings in downstream applications by initializing numbers with our approach for the task of magnitude prediction . We also introduce a regularization approach to learn model-based embeddings of numbers in a contextual setting . Word embeddings capture semantic relationships between words by operationalizing the distributional hypothesis ( Harris , 1954 ; Firth , 1957 ) . They can be learned either non-contextually ( Mikolov et al . , 2013b ; Pennington et al . , 2014 ; Bojanowski et al . , 2017 ) or contextually ( Devlin et al . , 2018 ; Peters et al . , 2018 ) . Non-contextual embeddings have worked well on various language understanding and semantic tasks ( Rumelhart et al . , 1988 ; Mikolov et al . , 2013a , b ) . More recently , they have also been used as pre-trained word embeddings to aid more sophisticated contextual models for solving rigorous natural language processing ( NLP ) problems , including translation , paraphrasing , and sentence-similarity tasks ( Kiros et al . , 2015 ; Wieting et al . , 2015 ) . While word embeddings effectively capture semantic relationships between words , they are less effective at capturing numeric properties associated with numbers . Though numbers represent a significant percentage of tokens in a corpus , they are often overlooked . In non-contextual word embedding models , they are treated like any other word , which leads to misinterpretation . For instance , they exhibit unintuitive similarities with other words and do not contain strong prior information about the magnitude of the number they encode . In sentence similarity and reasoning tasks , failure to handle numbers causes as much as 29 % of contradictions ( De Marneffe et al . , 2008 ) . In other data-intensive tasks where numbers are abundant , like neural machine translation , they are masked to hide the translation models inefficiency in dealing with them ( Mitchell and Lapata , 2009 ) . There are a variety of tests proposed to measure the efficiency of number embeddings . For instance , Naik et al . ( 2019 ) shows that GloVe ( Pennington et al . , 2014 ) , word2vec ( Mikolov et al . , 2013b ) , and fastText ( Joulin et al . , 2016 ; Bojanowski et al . , 2017 ) fail to capture numeration and magnitude properties of a number . Numeration is the property of associating numbers with their corresponding word representations ( " 3 " and " three " ) while magnitude represents a number 's actual value ( 3 < 4 ) . Further , Wallace et al . ( 2019 ) proposes several tests for analyzing numerical reasoning of number embeddings that include list maximum , decoding , and addition . In this paper , we experimentally demonstrate that if the cosine similarity between word embeddings of two numbers reflects their actual distance on the number line , the resultant word embeddings are useful in downstream tasks . We first demonstrate how Deterministic , Independent-of-Corpus Embeddings ( DICE ) can be constructed such that they almost perfectly capture properties of numera-tion and magnitude . These non-contextual embeddings also perform well on related tests for numeracy ( Wallace et al . , 2019 ) . To demonstrate the efficacy of DICE for downstream tasks , we explore its utility in two experiments . First , we design a DICE embedding initialized Bi-LSTM network to classify the magnitude of masked numbers in the 600 K dataset ( Chen et al . , 2019 ) . Second , given the popularity of modern contextual model-based embeddings , we devise a regularization procedure that emulates the hypothesis proposed by DICE and can be employed in any task-based fine-tuning process . We demonstrate that adding such regularization helps the model internalize notions of numeracy while learning task-based contextual embeddings for the numbers present in the text . We find promising results in a numerical reasoning task that involves numerical question answering based on a sub-split of the popular SQuAD dataset ( Rajpurkar et al . , 2016 ) . Our contribution can be summarized as follows : • We propose a deterministic technique to learn numerical embeddings . DICE embeddings are learned independently of corpus and effectively capture properties of numeracy . • We prove experimentally that the resultant embeddings learned using the above methods improve a model 's ability to reason about numbers in a variety of tasks , including numeration , magnitude , list maximum , decoding , and addition . • We also demonstrate that properties of DICE can be adapted to contextual models , like BERT ( Devlin et al . , 2018 ) , through a novel regularization technique for solving tasks involving numerical reasoning . In this work , we methodologically assign and learn embeddings for numbers to reflect their numerical properties . We validate our proposed approach with several experiments that test number embeddings . The tests that evaluate the numeral embeddings are fundamentally applicable to all real numbers . Finally , we introduced an approach to jointly learn embeddings of numbers and words that preserve numerical properties and evaluated them on a contextual word embedding based model . In our future work , we would like to extend this idea to unseen numbers in vocabulary as a function of seen ones .

Existing word embeddings treat numbers like words failing to capture numeration and magnitude properties of numbers which is problematic for tasks such as question answering.

They propose a deterministic technique to learn numerical embeddings where cosine similarity reflects the actual distance and a regularization approach for a contextual setting.

A Bi-LSTM network initialized with the proposed embedding shows the ability to capture numeration and magnitude and to perform list maximum, decoding, and addition.

P12-1103

We propose a novel approach to improve SMT via paraphrase rules which are automatically extracted from the bilingual training data . Without using extra paraphrase resources , we acquire the rules by comparing the source side of the parallel corpus with the target-to-source translations of the target side . Besides the word and phrase paraphrases , the acquired paraphrase rules mainly cover the structured paraphrases on the sentence level . These rules are employed to enrich the SMT inputs for translation quality improvement . The experimental results show that our proposed approach achieves significant improvements of 1.6~3.6 points of BLEU in the oral domain and 0.5~1 points in the news domain . This work was done when the first author was visiting Baidu . The translation quality of the SMT system is highly related to the coverage of translation models . However , no matter how much data is used for training , it is still impossible to completely cover the unlimited input sentences . This problem is more serious for online SMT systems in real-world applications . Naturally , a solution to the coverage problem is to bridge the gaps between the input sentences and the translation models , either from the input side , which targets on rewriting the input sentences to the MT-favored expressions , or from the side of translation models , which tries to enrich the translation models to cover more expressions . In recent years , paraphrasing has been proven useful for improving SMT quality . The proposed methods can be classified into two categories according to the paraphrase targets : ( 1 ) enrich translation models to cover more bilingual expressions ; ( 2 ) paraphrase the input sentences to reduce OOVs or generate multiple inputs . In the first category , He et al . ( 2011 ) , Bond et al . ( 2008 ) and Nakov ( 2008 ) enriched the SMT models via paraphrasing the training corpora . Kuhn et al . ( 2010 ) and Max ( 2010 ) used paraphrases to smooth translation models . For the second category , previous studies mainly focus on finding translations for unknown terms using phrasal paraphrases . Callison-Burch et al . ( 2006 ) and Marton et al . ( 2009 ) paraphrase unknown terms in the input sentences using phrasal paraphrases extracted from bilingual and monolingual corpora . Mirkin et al . ( 2009 ) rewrite OOVs with entailments and paraphrases acquired from WordNet . Onishi et al . ( 2010 ) and Du et al . ( 2010 ) use phrasal paraphrases to build a word lattice to get multiple input candidates . In the above methods , only word or phrasal paraphrases are used for input sentence rewriting . No structured paraphrases on the sentence level have been investigated . However , the information in the sentence level is very important for disambiguation . For example , we can only substitute play with drama in a context related to stage or theatre . Phrasal paraphrase substitutions can hardly solve such kind of problems . In this paper , we propose a method that rewrites the input sentences of the SMT system using automatically extracted paraphrase rules which can capture structures on sentence level in addition to paraphrases on the word or phrase level . Without extra paraphrase resources , a novel approach is proposed to acquire paraphrase rules from the bilingual training corpus based on the results of Forward-Translation and Back-Translation . The rules target on rewriting the input sentences to an MT-favored expression to ensure a better translation . The paraphrase rules cover all kinds of paraphrases on the word , phrase and sentence levels , enabling structure reordering , word or phrase insertion , deletion and substitution . The experimental results show that our proposed approach achieves significant improvements of 1.6~3.6 points of BLEU in the oral domain and 0.5~1 points in the news domain . The remainder of the paper is organized as follows : Section 2 makes a comparison between the Forward-Translation and Back-Translation . Section 3 introduces our methods that extract paraphrase rules from the bilingual corpus of SMT . Section 4 describes the strategies for constructing word lattice with paraphrase rules . The experimental results and some discussions are presented in Section 5 and Section 6 . Section 7 compares our work to the previous researches . Finally , Section 8 concludes the paper and suggests directions for future work . In this paper , we propose a novel method for extracting paraphrase rules by comparing the source side of bilingual corpus to the target-tosource translation of the target side . The acquired paraphrase rules are employed to enrich the SMT inputs , which target on rewriting the input sentences to an MT-favored form . The paraphrase rules cover all kinds of paraphrases on the word , phrase and sentence levels , enabling structure reordering , word or phrase insertion , deletion and substitution . Experimental results show that the paraphrase rules can improve SMT quality in both the oral and news domains . The manual investigation on oral translation results indicate that the paraphrase rules capture four kinds of MTfavored transformation to ensure translation quality improvement .

Incorporating paraphrases improves statistical machine translation however no works investigate sentence level paraphrases.

They propose to use bilingual training data to obtain paraphrase rules on word, phrase and sentence levels to rewrite inputs to be MT-favored.

The acquired paraphrase rules improve translation qualities in oral and news domains.

N09-1072

Automatically extracting social meaning and intention from spoken dialogue is an important task for dialogue systems and social computing . We describe a system for detecting elements of interactional style : whether a speaker is awkward , friendly , or flirtatious . We create and use a new spoken corpus of 991 4-minute speed-dates . Participants rated their interlocutors for these elements of style . Using rich dialogue , lexical , and prosodic features , we are able to detect flirtatious , awkward , and friendly styles in noisy natural conversational data with up to 75 % accuracy , compared to a 50 % baseline . We describe simple ways to extract relatively rich dialogue features , and analyze which features performed similarly for men and women and which were gender-specific . How can we extract social meaning from speech , deciding if a speaker is particularly engaged in the conversation , is uncomfortable or awkward , or is particularly friendly and flirtatious ? Understanding these meanings and how they are signaled in language is an important sociolinguistic task in itself . Extracting them automatically from dialogue speech and text is crucial for developing socially aware computing systems for tasks such as detection of interactional problems or matching conversational style , and will play an important role in creating more natural dialogue agents ( Pentland , 2005 ; Nass and Brave , 2005 ; Brave et al . , 2005 ) . Cues for social meaning permeate speech at every level of linguistic structure . Acoustic cues such as low and high F0 or energy and spectral tilt are important in detecting emotions such as annoyance , anger , sadness , or boredom ( Ang et al . , 2002 ; Lee and Narayanan , 2002 ; Liscombe et al . , 2003 ) , speaker characteristics such as charisma ( Rosenberg and Hirschberg , 2005 ) , or personality features like extroversion ( Mairesse et al . , 2007 ; Mairesse and Walker , 2008 ) . Lexical cues to social meaning abound . Speakers with links to depression or speakers who are under stress use more first person singular pronouns ( Rude et al . , 2004 ; Pennebaker and Lay , 2002 ; Cohn et al . , 2004 ) , positive emotion words are cues to agreeableness ( Mairesse et al . , 2007 ) , and negative emotion words are useful cues to deceptive speech ( Newman et al . , 2003 ) . The number of words in a sentence can be a useful feature for extroverted personality ( Mairesse et al . , 2007 ) . Finally , dialog features such as the presence of disfluencies can inform listeners about speakers ' problems in utterance planning or about confidence ( Brennan and Williams , 1995 ; Brennan and Schober , 2001 ) . Our goal is to see whether cues of this sort are useful in detecting particular elements of conversational style and social intention ; whether a speaker in a speed-dating conversation is judged by the interlocutor as friendly , awkward , or flirtatious . The results presented here should be regarded with some caution . The sample is not a random sample of English speakers or American adults , and speed dating is not a natural context for expressing every conversational style . Therefore , a wider array of studies across populations and genres would be required before a more general theory of conversational styles is established . On the other hand , the presented results may under-reflect the relations being captured . The quality of recordings and coarse granularity ( 1 second ) of the time-stamps likely cloud the relations , and as the data is cleaned and improved , we expect the associations to only grow stronger . Caveats aside , we believe the evidence indicates that the perception of several types of conversational style have relatively clear signals across genders , but with some additional gender contextualization . Both genders convey flirtation by laughing more , speaking faster , and using higher and more variable pitch . Both genders convey friendliness by laughing more , and using collaborative completions . However , we do find gender differences ; men asl more questions when ( labeled as ) flirting , women ask fewer . Men labeled as flirting are softer , but women labeled as flirting are louder . Women flirt-ing swear more , while men are more likely to use sexual vocabulary . Gender differences exist as well for the other variables . Men labeled as friendly use you while women labeled as friendly use I. Friendly women are very disfluent ; friendly men are not . While the features for friendly and flirtatious speech overlap , there are clear differences . Men speaker faster and with higher f0 ( min ) in flirtatious speech , but not faster and with lower f0 ( min ) in friendly speech . For men , flirtatious speech involves more questions and repair questions , while friendly speech does not . For women , friendly speech is more disfluent than flirtatious speech , and has more collaborative style ( completions , repair questions , appreciations ) . We also seem to see a model of collaborative conversational style ( probably related to the collaborative floor of Edelsky ( 1981 ) and Coates ( 1996 ) ) , cued by the use of more collaborative completions , repair questions and other questions , you , and laughter . These collaborative techniques were used by both women and men who were labeled as friendly , and occurred less with men labeled as awkward . Women themselves displayed more of this collaborative conversational style when they labeled the men as friendly . For women only , collaborative style included appreciations ; while for men only , collaborative style included overlaps . In addition to these implications for social science , our work has implications for the extraction of meaning in general . A key focus of our work was on ways to extract useful dialog act and disfluency features ( repair questions , backchannels , appreciations , restarts , dispreferreds ) with very shallow methods . These features were indeed extractable and proved to be useful features in classification . We are currently extending these results to predict date outcomes including ' liking ' , extending work such as Madan and Pentland ( 2006 ) .

Methods to extract social meanings such as engagement from speech remain unknown while it is important in sociolinguistics and to develop socially aware computing systems.

They create a spoken corpus from conversations in speed-dating and perform analysis using extracted dialogue features with a focus on genders.

They found several gender dependent and independent phenomena in conversations related to the speed of speaking, laughing or asking questions.

P18-1256

We introduce the task of predicting adverbial presupposition triggers such as also and again . Solving such a task requires detecting recurring or similar events in the discourse context , and has applications in natural language generation tasks such as summarization and dialogue systems . We create two new datasets for the task , derived from the Penn Treebank and the Annotated English Gigaword corpora , as well as a novel attention mechanism tailored to this task . Our attention mechanism augments a baseline recurrent neural network without the need for additional trainable parameters , minimizing the added computational cost of our mechanism . We demonstrate that our model statistically outperforms a number of baselines , including an LSTM-based language model . In pragmatics , presuppositions are assumptions or beliefs in the common ground between discourse participants when an utterance is made ( Frege , 1892 ; Strawson , 1950 ; Stalnaker , 1973 Stalnaker , , 1998 ) ) , and are ubiquitous in naturally occurring discourses ( Beaver and Geurts , 2014 ) . Presuppositions underly spoken statements and written sentences and understanding them facilitates smooth communication . We refer to expressions that indicate the presence of presuppositions as presupposition triggers . These include definite descriptions , factive verbs and certain adverbs , among others . For example , consider the following statements : ( 1 ) John is going to the restaurant again . * Authors ( listed in alphabetical order ) contributed equally . ( 2 ) John has been to the restaurant . ( 1 ) is only appropriate in the context where ( 2 ) is held to be true because of the presence of the presupposition trigger again . One distinguishing characteristic of presupposition is that it is unaffected by negation of the presupposing context , unlike other semantic phenomena such as entailment and implicature . The negation of ( 1 ) , John is not going to the restaurant again . , also presupposes ( 2 ) . Our focus in this paper is on adverbial presupposition triggers such as again , also and still . Adverbial presupposition triggers indicate the recurrence , continuation , or termination of an event in the discourse context , or the presence of a similar event . In one study of presuppositional triggers in English journalistic texts ( Khaleel , 2010 ) , adverbial triggers were found to be the most commonly occurring presupposition triggers after existential triggers . 1 Despite their frequency , there has been little work on these triggers in the computational literature from a statistical , corpus-driven perspective . As a first step towards language technology systems capable of understanding and using presuppositions , we propose to investigate the detection of contexts in which these triggers can be used . This task constitutes an interesting testing ground for pragmatic reasoning , because the cues that are indicative of contexts containing recurring or similar events are complex and often span more than one sentence , as illustrated in Sentences ( 1 ) and ( 2 ) . Moreover , such a task has immediate practical consequences . For example , in language generation applications such as summarization and dialogue systems , adding presuppositional triggers in contextually appropriate loca-1 Presupposition of existence are triggered by possessive constructions , names or definite noun phrases . tions can improve the readability and coherence of the generated output . We create two datasets based on the Penn Treebank corpus ( Marcus et al . , 1993 ) and the English Gigaword corpus ( Graff et al . , 2007 ) , extracting contexts that include presupposition triggers as well as other similar contexts that do not , in order to form a binary classification task . In creating our datasets , we consider a set of five target adverbs : too , again , also , still , and yet . We focus on these adverbs in our investigation because these triggers are well known in the existing linguistic literature and commonly triggering presuppositions . We control for a number of potential confounding factors , such as class balance , and the syntactic governor of the triggering adverb , so that models can not exploit these correlating factors without any actual understanding of the presuppositional properties of the context . We test a number of standard baseline classifiers on these datasets , including a logistic regression model and deep learning methods based on recurrent neural networks ( RNN ) and convolutional neural networks ( CNN ) . In addition , we investigate the potential of attention-based deep learning models for detecting adverbial triggers . Attention is a promising approach to this task because it allows a model to weigh information from multiple points in the previous context and infer long-range dependencies in the data ( Bahdanau et al . , 2015 ) . For example , the model could learn to detect multiple instances involving John and restaurants , which would be a good indication that again is appropriate in that context . Also , an attention-based RNN has achieved success in predicting article definiteness , which involves another class of presupposition triggers ( Kabbara et al . , 2016 ) . As another contribution , we introduce a new weighted pooling attention mechanism designed for predicting adverbial presupposition triggers . Our attention mechanism allows for a weighted averaging of our RNN hidden states where the weights are informed by the inputs , as opposed to a simple unweighted averaging . Our model uses a form of self-attention ( Paulus et al . , 2018 ; Vaswani et al . , 2017 ) , where the input sequence acts as both the attention mechanism 's query and key / value . Unlike other attention models , instead of simply averaging the scores to be weighted , our approach aggregates ( learned ) attention scores by learning a reweighting scheme of those scores through another level ( dimension ) of attention . Additionally , our mechanism does not introduce any new parameters when compared to our LSTM baseline , reducing its computational impact . We compare our model using the novel attention mechanism against the baseline classifiers in terms of prediction accuracy . Our model outperforms these baselines for most of the triggers on the two datasets , achieving 82.42 % accuracy on predicting the adverb " also " on the Gigaword dataset . The contributions of this work are as follows : 1 . We introduce the task of predicting adverbial presupposition triggers . 2 . We present new datasets for the task of detecting adverbial presupposition triggers , with a data extraction method that can be applied to other similar pre-processing tasks . 3 . We develop a new attention mechanism in an RNN architecture that is appropriate for the prediction of adverbial presupposition triggers , and show that its use results in better prediction performance over a number of baselines without introducing additional parameters . 2 Related Work In this work , we have investigated the task of predicting adverbial presupposition triggers and introduced several datasets for the task . Additionally , we have presented a novel weighted-pooling attention mechanism which is incorporated into a recurrent neural network model for predicting the presence of an adverbial presuppositional trigger . Our results show that the model outperforms the CNN and LSTM , and does not add any additional parameters over the standard LSTM model . This shows its promise in classification tasks involving capturing and combining relevant information from multiple points in the previous context . In future work , we would like to focus more on designing models that can deal with and be optimized for scenarios with severe data imbalance . We would like to also explore various applications of presupposition trigger prediction in language generation applications , as well as additional attention-based neural network architectures .

Adverbaial triggers indicate the event recurrence, continuation, or termination in the discourse context and are frequently found in English but there are few related works.

They introduce an adverbial presupposition trigger prediction task and datasets and propose an attention mechanism that augments a recurrent neural network without additional trainable parameters.

The proposed model outperforms baselines including an LSTM-based language model on most of the triggers on the two datasets.

P08-1116

This paper proposes a novel method that exploits multiple resources to improve statistical machine translation ( SMT ) based paraphrasing . In detail , a phrasal paraphrase table and a feature function are derived from each resource , which are then combined in a log-linear SMT model for sentence-level paraphrase generation . Experimental results show that the SMT-based paraphrasing model can be enhanced using multiple resources . The phrase-level and sentence-level precision of the generated paraphrases are above 60 % and 55 % , respectively . In addition , the contribution of each resource is evaluated , which indicates that all the exploited resources are useful for generating paraphrases of high quality . Paraphrases are alternative ways of conveying the same meaning . Paraphrases are important in many natural language processing ( NLP ) applications , such as machine translation ( MT ) , question answering ( QA ) , information extraction ( IE ) , multidocument summarization ( MDS ) , and natural language generation ( NLG ) . This paper addresses the problem of sentencelevel paraphrase generation , which aims at generating paraphrases for input sentences . An example of sentence-level paraphrases can be seen below : S1 : The table was set up in the carriage shed . S2 : The table was laid under the cart-shed . Paraphrase generation can be viewed as monolingual machine translation ( Quirk et al . , 2004 ) , which typically includes a translation model and a language model . The translation model can be trained using monolingual parallel corpora . However , acquiring such corpora is not easy . Hence , data sparseness is a key problem for the SMT-based paraphrasing . On the other hand , various methods have been presented to extract phrasal paraphrases from different resources , which include thesauri , monolingual corpora , bilingual corpora , and the web . However , little work has been focused on using the extracted phrasal paraphrases in sentence-level paraphrase generation . In this paper , we exploit multiple resources to improve the SMT-based paraphrase generation . In detail , six kinds of resources are utilized , including : ( 1 ) an automatically constructed thesaurus , ( 2 ) a monolingual parallel corpus from novels , ( 3 ) a monolingual comparable corpus from news articles , ( 4 ) a bilingual phrase table , ( 5 ) word definitions from Encarta dictionary , and ( 6 ) a corpus of similar user queries . Among the resources , ( 1 ) , ( 2 ) , ( 3 ) , and ( 4 ) have been investigated by other researchers , while ( 5 ) and ( 6 ) are first used in this paper . From those resources , six phrasal paraphrase tables are extracted , which are then used in a log-linear SMTbased paraphrasing model . Both phrase-level and sentence-level evaluations were carried out in the experiments . In the former one , phrase substitutes occurring in the paraphrase sentences were evaluated . While in the latter one , the acceptability of the paraphrase sentences was evaluated . Experimental results show that : ( 1 ) The SMT-based paraphrasing is enhanced using multiple resources . The phrase-level and sentence-level precision of the generated paraphrases exceed 60 % and 55 % , respectively . ( 2 ) Although the contributions of the resources differ a lot , all the resources are useful . ( 3 ) The performance of the method varies greatly on different test sets and it performs best on the test set of news sentences , which are from the same source as most of the training data . The rest of the paper is organized as follows : Section 2 reviews related work . Section 3 introduces the log-linear model for paraphrase generation . Section 4 describes the phrasal paraphrase extraction from different resources . Section 5 presents the parameter estimation method . Section 6 shows the experiments and results . Section 7 draws the conclusion . This paper proposes a method that improves the SMT-based sentence-level paraphrase generation using phrasal paraphrases automatically extracted from different resources . Our contribution is that we combine multiple resources in the framework of SMT for paraphrase generation , in which the dic-tionary definitions and similar user queries are first used as phrasal paraphrases . In addition , we analyze and compare the contributions of different resources . Experimental results indicate that although the contributions of the exploited resources differ a lot , they are all useful to sentence-level paraphrase generation . Especially , the dictionary definitions and similar user queries are effective for paraphrasing some certain types of phrases . In the future work , we will try to use syntactic and context constraints in paraphrase generation to enhance the acceptability of the paraphrases . In addition , we will extract paraphrase patterns that contain more structural variation and try to combine the SMT-based and pattern-based systems for sentencelevel paraphrase generation .

Paraphrase generation requires monolingual parallel corpora which is not easily obtainable, and few works focus on using the extracted phrasal paraphrases in sentence-level paraphrase generation.

They propose to exploit six paraphrase resources to extract phrasal paraphrase tables that are further used to build a log-linear statistical machine translation-based paraphrasing model.

They show that using multiple resources enhances paraphrase generation quality in precision on phrase and sentence level especially when they are similar to user queries.

P08-1027

There are many possible different semantic relationships between nominals . Classification of such relationships is an important and difficult task ( for example , the well known noun compound classification task is a special case of this problem ) . We propose a novel pattern clusters method for nominal relationship ( NR ) classification . Pattern clusters are discovered in a large corpus independently of any particular training set , in an unsupervised manner . Each of the extracted clusters corresponds to some unspecified semantic relationship . The pattern clusters are then used to construct features for training and classification of specific inter-nominal relationships . Our NR classification evaluation strictly follows the ACL SemEval-07 Task 4 datasets and protocol , obtaining an f-score of 70.6 , as opposed to 64.8 of the best previous work that did not use the manually provided WordNet sense disambiguation tags . Automatic extraction and classification of semantic relationships is a major field of activity , of both practical and theoretical interest . A prominent type of semantic relationships is that holding between nominals1 . For example , in noun compounds many different semantic relationships are encoded by the same simple form ( Girju et al . , 2005 ) : ' dog food ' denotes food consumed by dogs , while ' summer morn-ing ' denotes a morning that happens in the summer . These two relationships are completely different semantically but are similar syntactically , and distinguishing between them could be essential for NLP applications such as question answering and machine translation . Relation classification usually relies on a training set in the form of tagged data . To improve results , some systems utilize additional manually constructed semantic resources such as WordNet ( WN ) ( Beamer et al . , 2007 ) . However , in many domains and languages such resources are not available . Furthermore , usage of such resources frequently requires disambiguation and connection of the data to the resource ( word sense disambiguation in the case of WordNet ) . Manual disambiguation is unfeasible in many practical tasks , and an automatic one may introduce errors and greatly degrade performance . It thus makes sense to try to minimize the usage of such resources , and utilize only corpus contexts in which the relevant words appear . A leading method for utilizing context information for classification and extraction of relationships is that of patterns ( Hearst , 1992 ; Pantel and Pennacchiotti , 2006 ) . The standard classification process is to find in an auxiliary corpus a set of patterns in which a given training word pair co-appears , and use pattern-word pair co-appearance statistics as features for machine learning algorithms . In this paper we introduce a novel approach , based on utilizing pattern clusters that are prepared separately and independently of the training set . We do not utilize any manually constructed resource or any manual tagging of training data beyond the cor-rect classification , thus making our method applicable to fully automated tasks and less domain and language dependent . Moreover , our pattern clustering algorithm is fully unsupervised . Our method is based on the observation that while each lexical pattern can be highly ambiguous , several patterns in conjunction can reliably define and represent a lexical relationship . Accordingly , we construct pattern clusters from a large generic corpus , each such cluster potentially representing some important generic relationship . This step is done without accessing any training data , anticipating that most meaningful relationships , including those in a given classification problem , will be represented by some of the discovered clusters . We then use the training set to label some of the clusters , and the labeled clusters to assign classes to tested items . One of the advantages of our method is that it can be used not only for classification , but also for further analysis and retrieval of the observed relationships2 . The semantic relationships between the components of noun compounds and between nominals in general are not easy to categorize rigorously . Several different relationship hierarchies have been proposed ( Nastase and Szpakowicz , 2003 ; Moldovan et al . , 2004 ) . Some classes , like Container-Contained , Time-Event and Product-Producer , appear in several classification schemes , while classes like Tool-Object are more vaguely defined and are subdivided differently . Recently , SemEval-07 Task 4 ( Girju et al . , 2007 ) proposed a benchmark dataset that includes a subset of 7 widely accepted nominal relationship ( NR ) classes , allowing consistent evaluation of different NR classification algorithms . In the SemEval event , 14 research teams evaluated their algorithms using this benchmark . Some of the teams have used the manually annotated WN labels provided with the dataset , and some have not . We evaluated our algorithm on SemEval-07 Task 4 data , showing superior results over participating algorithms that did not utilize WordNet disambiguation tags . We also show how pattern clusters can be used for a completely unsupervised classification of the test set . Since in this case no training data is used , this allows the automated discovery of a potentially unbiased classification scheme . Section 2 discusses related work , Section 3 outlines the pattern clustering algorithm , Section 4 details three classification methods , and Sections 5 and 6 describe the evaluation protocol and results . Relationship classification is known to improve many practical tasks , e.g. , textual entailment ( Tatu and Moldovan , 2005 ) . We have presented a novel framework for relationship classification , based on pattern clusters prepared as a standalone resource independently of the training set . Our method outperforms current state-of-the-art algorithms that do not utilize WordNet tags on Task 4 of SemEval-07 . In practical situations , it would not be feasible to provide a large amount of such sense disambiguation tags manually . Our method also shows competitive performance compared to the majority of task participants that do utilize WN tags . Our method can produce labeled pattern clusters , which can be potentially useful for automatic discovery of additional instances for a given relationship . We intend to pursue this promising direction in future work .

Using annotated data or semantic resources such as WordNet for relation classification introduces errors and such data is not available in many domains and languages.

They propose an unsupervised pattern clustering method for nominal relation classification using a large generic corpus enabling scale in domain and language.

Experiments on the ACL SemEval-07 dataset show the proposed method performs better than existing methods that do not use disambiguation tags.

2021.emnlp-main.185

Learning sentence embeddings from dialogues has drawn increasing attention due to its low annotation cost and high domain adaptability . Conventional approaches employ the siamese-network for this task , which obtains the sentence embeddings through modeling the context-response semantic relevance by applying a feed-forward network on top of the sentence encoders . However , as the semantic textual similarity is commonly measured through the element-wise distance metrics ( e.g. cosine and L2 distance ) , such architecture yields a large gap between training and evaluating . In this paper , we propose DialogueCSE , a dialogue-based contrastive learning approach to tackle this issue . DialogueCSE first introduces a novel matching-guided embedding ( MGE ) mechanism , which generates a contextaware embedding for each candidate response embedding ( i.e. the context-free embedding ) according to the guidance of the multi-turn context-response matching matrices . Then it pairs each context-aware embedding with its corresponding context-free embedding and finally minimizes the contrastive loss across all pairs . We evaluate our model on three multi-turn dialogue datasets : the Microsoft Dialogue Corpus , the Jing Dong Dialogue Corpus , and the E-commerce Dialogue Corpus . Evaluation results show that our approach significantly outperforms the baselines across all three datasets in terms of MAP and Spearman 's correlation measures , demonstrating its effectiveness . Further quantitative experiments show that our approach achieves better performance when leveraging more dialogue context and remains robust when less training data is provided . Sentence embeddings are used with success for a variety of NLP applications ( Cer et al . , 2018 ) and many prior methods have been proposed with different learning schemes . Kiros et al . ( 2015 ) ; Logeswaran and Lee ( 2018 ) ; Hill et al . ( 2016 ) train sentence encoders in a self-supervised manner with web pages and books . Conneau et al . ( 2017 ) ; Cer et al . ( 2018 ) ; Reimers and Gurevych ( 2019 ) propose to learn sentence embeddings on the supervised datasets such as SNLI ( Bowman et al . , 2015 ) and MNLI ( Williams et al . , 2018 ) . Although the supervised-learning approaches achieve better performance , they suffer from high cost of annotation in building the training dataset , which makes them hard to adapt to other domains or languages . Recently , learning sentence embeddings from dialogues has begun to attract increasing attention . Dialogues provide strong semantic relationships among conversational utterances and are usually easy to collect in large amounts . Such advantages make the dialogue-based self-supervised learning methods promising to achieve competitive or even superior performance against the supervised-learning methods , especially under the low-resource conditions . While promising , the issue of how to effectively exploit the dialogues for this task has not been sufficiently explored . Yang et al . ( 2018 ) propose to train an input-response prediction model on Reddit dataset ( Al-Rfou et al . , 2016 ) . Since they build their architecture based on the single-turn dialogue , the multi-turn dialogue history is not fully exploited . Henderson et al . ( 2020 ) demonstrate that introducing the multi-turn dialogue context can improve the sentence embedding performance . However , they concatenate the multi-turn dialogue context into a long token sequence , failing to model intersentence semantic relationships among the utterances . Recently , more advanced methods such as ( Reimers and Gurevych , 2019 ) achieve better performance by employing BERT ( Devlin et al . , 2019 ) as the sentence encoder . These works have in common that they employ a feed-forward network with a non-linear activation on top of the sentence en-coders to model the context-response semantic relevance , thereby learning the sentence embeddings . However , such architecture presents two limitations : ( 1 ) It yields a large gap between training and evaluating , since the semantic textual similarity is commonly measured by the element-wise distance metrics such as cosine and L2 distance . ( 2 ) Concatenating all the utterances in the dialogue context inevitably introduces the noise as well as the redundant information , resulting in a poor result . In this paper , we propose DialogueCSE , a dialogue-based contrastive learning approach to tackle these issues . We hold that the semantic matching relationships between the context and the response can be implicitly modeled through contrastive learning , thus making it possible to eliminate the gap between training and evaluating . To this end , we introduce a novel matching-guided embedding ( MGE ) mechanism . Specifically , MGE first pairs each utterance in the context with the response and performs a token-level dot-product operation across all the utterance-response pairs to obtain the multi-turn matching matrices . Then the multi-turn matching matrices are used as guidance to generate a context-aware embedding for the response embedding ( i.e. the context-free embedding ) . Finally , the context-aware embedding and the context-free embedding are paired as a training sample , whose label is determined by whether the context and the response are originally from the same dialogue . Our motivation is that once the context semantically matches the response , it has the ability to distill the context-aware information from the context-free embedding , which is exactly the learning objective of the sentence encoder that aims to produce context-aware sentence embeddings . We train our model on three multi-turn dialogue datasets : the Microsoft Dialogue Corpus ( MDC ) ( Li et al . , 2018 ) , the Jing Dong Dialogue Corpus ( JDDC ) ( Chen et al . , 2020 ) , and the E-commerce Dialogue Corpus ( ECD ) ( Zhang et al . , 2018 ) . To evaluate our model , we introduce two types of tasks : the semantic retrieval ( SR ) task and the dialogue-based semantic textual similarity ( D-STS ) task . Here we do not adopt the standard semantic textual similarity ( STS ) task ( Cer et al . , 2017 ) for two reasons : ( 1 ) As revealed in ( Zhang et al . , 2020 ) , the sentence embedding performance varies greatly as the domain of the training data changes . As a dialogue dataset is always about several certain domains , evaluating on the STS benchmark may mis-lead the evaluation of the model . ( 2 ) The dialoguebased sentence embeddings focus on context-aware rather than context-free semantic meanings , which may not be suitable to be evaluated through the context-free benchmarks . Since previous dialoguebased works have not set up a uniform benchmark , we construct two evaluation datasets for each dialogue corpus . A total of 18,964 retrieval samples and 4,000 sentence pairs are annotated by seven native speakers through the crowd-sourcing platform1 . The evaluation results indicate that DialogueCSE significantly outperforms the baselines on the three datasets in terms of both MAP and Spearman 's correlation metrics , demonstrating its effectiveness . Further quantitative experiments show that Dia-logueCSE achieves better performance when leveraging more dialogue context and remains robust when less training data is provided . To sum up , our contributions are threefold : • We propose DialogueCSE , a dialogue-based contrastive learning approach with MGE mechanism for learning sentence embeddings from dialogues . As far as we know , this is the first attempt to apply contrastive learning in this area . • We construct the dialogue-based sentence embedding evaluation benchmarks for three dialogue corpus . All of the datasets will be released to facilitate the follow-up researches . • Extensive experiments show that Dia-logueCSE significantly outperforms the baselines , establishing the state-of-the-art results . 2 Related Work In this work , we propose DialogueCSE , a dialoguebased contrastive learning approach to learn sentence embeddings from dialogues . We also propose uniform evaluation benchmarks for evaluating the quality of the dialogue-based sentence embeddings . Evaluation results show that DialogueCSE achieves the best result over the baselines while adding no additional parameters . In the next step , we will study how to introduce more interaction information to learn the sentence embeddings and try to incorporate the contrast learning method into the pre-training stage .

Existing methods to learn representations from dialogues have a similarity-measurement gap between training and evaluation time and do not exploit the multi-turn structure of data.

They propose a dialogue-based contrastive learning approach to learn sentence embeddings from dialogues by modelling semantic matching relationships between the context and response implicitly.

The proposed approach outperforms baseline methods on two newly introduced tasks coupled with three multi-turn dialogue datasets in terms of MAP and Spearman's correlation measures.

P02-1051

Named entity phrases are some of the most difficult phrases to translate because new phrases can appear from nowhere , and because many are domain specific , not to be found in bilingual dictionaries . We present a novel algorithm for translating named entity phrases using easily obtainable monolingual and bilingual resources . We report on the application and evaluation of this algorithm in translating Arabic named entities to English . We also compare our results with the results obtained from human translations and a commercial system for the same task . Named entity phrases are being introduced in news stories on a daily basis in the form of personal names , organizations , locations , temporal phrases , and monetary expressions . While the identification of named entities in text has received significant attention ( e.g. , Mikheev et al . ( 1999 ) and Bikel et al . ( 1999 ) ) , translation of named entities has not . This translation problem is especially challenging because new phrases can appear from nowhere , and because many named-entities are domain specific , not to be found in bilingual dictionaries . A system that specializes in translating named entities such as the one we describe here would be an important tool for many NLP applications . Statisti-cal machine translation systems can use such a system as a component to handle phrase translation in order to improve overall translation quality . Cross-Lingual Information Retrieval ( CLIR ) systems could identify relevant documents based on translations of named entity phrases provided by such a system . Question Answering ( QA ) systems could benefit substantially from such a tool since the answer to many factoid questions involve named entities ( e.g. , answers to who questions usually involve Persons / Organizations , where questions involve Locations , and when questions involve Temporal Expressions ) . In this paper , we describe a system for Arabic-English named entity translation , though the technique is applicable to any language pair and does not require especially difficult-to-obtain resources . The rest of this paper is organized as follows . In Section 2 , we give an overview of our approach . In Section 3 , we describe how translation candidates are generated . In Section 4 , we show how monolingual clues are used to help re-rank the translation candidates list . In Section 5 , we describe how the candidates list can be extended using contextual information . We conclude this paper with the evaluation results of our translation algorithm on a test set . We also compare our system with human translators and a commercial system . We have presented a named entity translation algorithm that performs at near human translation accuracy when translating Arabic named entities to English . The algorithm uses very limited amount of hard-to-obtain bilingual resources and should be easily adaptable to other languages . We would like to apply to other languages such as Chinese and Japanese and to investigate whether the current algorithm would perform as well or whether new algorithms might be needed . Currently , our translation algorithm does not use any dictionary of named entities and they are translated on the fly . Translating a common name incorrectly has a significant effect on the translation accuracy . We would like to experiment with adding a small named entity translation dictionary for common names and see if this might improve the overall translation accuracy .

Translating named entities is challenging since they can appear from nowhere, and cannot be found in bilingual dictionaries because they are domain specific.

They propose an algorithm for Arabic-English named entity translation which uses easily obtainable monolingual and bilingual resources and a limited amount of hard-to-obtain bilingual resources.

The proposed algorithm is compared with human translators and a commercial system and it performs at near human translation.

E06-1014

Probabilistic Latent Semantic Analysis ( PLSA ) models have been shown to provide a better model for capturing polysemy and synonymy than Latent Semantic Analysis ( LSA ) . However , the parameters of a PLSA model are trained using the Expectation Maximization ( EM ) algorithm , and as a result , the trained model is dependent on the initialization values so that performance can be highly variable . In this paper we present a method for using LSA analysis to initialize a PLSA model . We also investigated the performance of our method for the tasks of text segmentation and retrieval on personal-size corpora , and present results demonstrating the efficacy of our proposed approach . In modeling a collection of documents for information access applications , the documents are often represented as a " bag of words " , i.e. , as term vectors composed of the terms and corresponding counts for each document . The term vectors for a document collection can be organized into a term by document co-occurrence matrix . When directly using these representations , synonyms and polysemous terms , that is , terms with multiple senses or meanings , are not handled well . Methods for smoothing the term distributions through the use of latent classes have been shown to improve the performance of a number of information access tasks , including retrieval over smaller collections ( Deerwester et al . , 1990 ) , text segmentation ( Brants et al . , 2002 ) , and text classification ( Wu and Gunopulos , 2002 ) . The Probabilistic Latent Semantic Analysis model ( PLSA ) ( Hofmann , 1999 ) provides a probabilistic framework that attempts to capture polysemy and synonymy in text for applications such as retrieval and segmentation . It uses a mixture decomposition to model the co-occurrence data , and the probabilities of words and documents are obtained by a convex combination of the aspects . The mixture approximation has a well defined probability distribution and the factors have a clear probabilistic meaning in terms of the mixture component distributions . The PLSA model computes the relevant probability distributions by selecting the model parameter values that maximize the probability of the observed data , i.e. , the likelihood function . The standard method for maximum likelihood estimation is the Expectation Maximization ( EM ) algorithm . For a given initialization , the likelihood function increases with EM iterations until a local maximum is reached , rather than a global maximum , so that the quality of the solution depends on the initialization of the model . Additionally , the likelihood values across different initializations are not comparable , as we will show . Thus , the likelihood function computed over the training data can not be used as a predictor of model performance across different models . Rather than trying to predict the best performing model from a set of models , in this paper we focus on finding a good way to initialize the PLSA model . We will present a framework for using Latent Semantic Analysis ( LSA ) ( Deerwester et al . , 1990 ) to better initialize the parameters of a corresponding PLSA model . The EM algorithm is then used to further refine the initial estimate . This combination of LSA and PLSA leverages the advantages of both . This paper is organized as follows : in section 2 , we review related work in the area . In section 3 , we summarize related work on LSA and its probabilistic interpretation . In section 4 we review the PLSA model and in section 5 we present our method for initializing a PLSA model using LSA model parameters . In section 6 , we evaluate the performance of our framework on a text segmentation task and several smaller information retrieval tasks . And in section 7 , we summarize our results and give directions for future work . We have presented LSA-PLSA , an approach for improving the performance of PLSA by leveraging the best features of PLSA and LSA . Our approach uses LSA to initialize a PLSA model , allowing for arbitrary weighting schemes to be incorporated into a PLSA model while leveraging the optimization used to improve the estimate of the PLSA parameters . We have evaluated the proposed framework on two tasks : personalsize information retrieval and text segmentation . The LSA-PLSA model outperformed PLSA on all tasks . And in all cases , combining PLSA-based models outperformed a single model . The best performance was obtained with combined models when one of the models was the LSA-PLSA model . When combining multiple PLSA models , the use of LSA-PLSA in combination with either two PLSA models or one PLSA and one LSA model improved performance while reducing the running time over the combination of four or more PLSA models as used by others . Future areas of investigation include quantifying the expected performance of the LSAinitialized PLSA model by comparing performance to that of the empirically best performing model and examining whether tempered EM could further improve performance .

EM algorithm-baed Probabilistic latent semantic analysis models provide high variance in performance and models with different initializations are not comparable.

They propose to use Latent Semantic Analysis to initialize probabilistic latent semantic analysis models, EM algorithm is further used to refine the initial estimate.

They show that the model initialized in the proposed method always outperforms existing methods.

2021.naacl-main.34

We rely on arguments in our daily lives to deliver our opinions and base them on evidence , making them more convincing in turn . However , finding and formulating arguments can be challenging . In this work , we present the Arg-CTRL-a language model for argument generation that can be controlled to generate sentence-level arguments for a given topic , stance , and aspect . We define argument aspect detection as a necessary method to allow this fine-granular control and crowdsource a dataset with 5,032 arguments annotated with aspects . Our evaluation shows that the Arg-CTRL is able to generate high-quality , aspectspecific arguments , applicable to automatic counter-argument generation . We publish the model weights and all datasets and code to train the Arg-CTRL . 1 Nuclear reactors produce radioactive waste ... Language models ( Bengio et al . , 2003 ) allow to generate text through learned distributions of a language and have been applied to a variety of areas like machine translation ( Bahdanau et al . , 2015 ) , summarization ( Paulus et al . , 2018 ) , or dialogue systems ( Wen et al . , 2017 ) . A rather new field for these models is the task of producing text with argumentative content ( Wang and Ling , 2016 ) . We believe this technology can support humans in the challenging task of finding and formulating arguments . A politician might use this to prepare for a debate with a political opponent or for a press conference . It may be used to support students in writing argumentative essays or to enrich one-sided discussions with counter-arguments . In contrast to retrieval methods , generation allows to combine and stylistically adapt text ( e.g. arguments ) based on a given input ( usually the beginning of a sentence ) . Current argument generation models , however , produce lengthy texts and allow the user little control over the aspect the argument should address ( Hua et al . , 2019 ; Hua and Wang , 2018 ) . We show that argument generation can be enhanced by allowing for a fine-grained control and limiting the argument to a single but concise sentence . Controllable language models like the CTRL ( Keskar et al . , 2019 ) allow to condition the model at training time to certain control codes . At inference , these can be used to direct the model 's output with regard to content or style . We build upon this architecture to control argument generation based solely on a given topic , stance , and argument aspect . For instance , to enforce focus on the aspect of cancer for the topic of nuclear energy , we input a control code " Nuclear Energy CON cancer " that creates a contra argument discussing this aspect , for instance : " Studies show that people living next to nuclear power plants have a higher risk of developing cancer . " . To obtain control codes from training data , we pre-define a set of topics to retrieve documents for and rely on an existing stance detection model to classify whether a sentence argues in favor ( pro ) or against ( con ) the given topic ( Stab et al . , 2018a ) . Regarding argument aspect detection , however , past work has two drawbacks : it either uses simple rule-based extraction of verb-and noun-phrases ( Fujii and Ishikawa , 2006 ) or the definition of aspects is based on target-concepts located within the same sentence ( Gemechu and Reed , 2019 ) . Aspects as we require and define them are not bound to any part-of-speech tag and ( 1 ) hold the core reason upon which the conclusion / evidence is built and ( 2 ) encode the stance towards a general but not necessarily explicitly mentioned topic the argument discusses . For instance : Topic : Nuclear Energy Argument : Running nuclear reactors is costly as it involves long-time disposal of radioactive waste . The evidence of this argument is based upon the two underlined aspects . While these aspects encode a negative stance towards the topic of " Nuclear Energy " , the topic itself is not mentioned explicitly in the argument . Our final controlled argument generation pipeline ( see Figure 1 ) works as follows : ( 1 ) We gather several million documents for eight different topics from two large data sources . All sentences are classified into pro- , con- , and non-arguments . We detect aspects of all arguments with a model trained on a novel dataset and concatenate arguments with the same topic , stance , and aspect into training documents . ( 2 ) We use the collected classified data to condition the Arg-CTRL on the topics , stances , and aspects of all gathered arguments . ( 3 ) At inference , passing the control code [ Topic ] [ Stance ] [ Aspect ] to the model will generate an argument that follows these commands . Our evaluation shows that the Arg-CTRL is able to produce aspect-specific , high-quality arguments , applicable to automatic counter-argument generation . The contributions are as follows : ( i ) We adapt and fine-tune the CTRL for aspect-controlled neural argument generation . ( ii ) We show that detecting argument aspects and conditioning the generation model on them are necessary steps to control the model 's training process and its perspective while generating . ( iii ) We propose several methods to analyze and evaluate the quality of ( controllable ) argument generation models . ( iv ) We develop a new scheme to annotate argument aspects and release a dataset with 5,032 samples . We apply the concept of controlled neural text generation to the domain of argument generation . Our Arg-CTRL is conditioned on topics , stances , and aspects and can reliably create arguments using these control codes . We show that arguments generated with our approach are genuine and of high argumentative and grammatical quality in general . Moreover , we show that our approach can be used to generate counter-arguments in a transparent and interpretable way . We fine-tune the Arg-CTRL on two different data sources and find that using mixed data from Common-Crawl results in a higher quality of generated arguments than using user discussions from Reddit-Comments . Further , we define argument aspect detection for controlled argument generation and introduce a novel annotation scheme to crowdsource argument aspect annotations , resulting in a high-quality dataset . We publish the model weights , data , and all code necessary to train the Arg-CTRL .

Argumentative content generation can support humans but current models produce lengthy texts and offer a little controllability on aspects of the argument for users.

They train a controllable language model on a corpus annotated with control codes provided by a stance detection model and introduce a dataset for evaluation.

The proposed model can generate arguments that are genuine and argumentative and grammatically correct and also counter-arguments in a transparent and interpretable way.

N16-1181

We describe a question answering model that applies to both images and structured knowledge bases . The model uses natural language strings to automatically assemble neural networks from a collection of composable modules . Parameters for these modules are learned jointly with network-assembly parameters via reinforcement learning , with only ( world , question , answer ) triples as supervision . Our approach , which we term a dynamic neural module network , achieves state-of-theart results on benchmark datasets in both visual and structured domains . This paper presents a compositional , attentional model for answering questions about a variety of world representations , including images and structured knowledge bases . The model translates from questions to dynamically assembled neural networks , then applies these networks to world representations ( images or knowledge bases ) to produce answers . We take advantage of two largely independent lines of work : on one hand , an extensive literature on answering questions by mapping from strings to logical representations of meaning ; on the other , a series of recent successes in deep neural models for image recognition and captioning . By constructing neural networks instead of logical forms , our model leverages the best aspects of both linguistic compositionality and continuous representations . Our model has two components , trained jointly : first , a collection of neural " modules " that can be freely composed ( Figure 1b ) ; second , a network layout predictor that assembles modules into complete deep networks tailored to each question ( Figure 1a ) . Previous work has used manually-specified modular structures for visual learning ( Andreas et al . , 2016 ) . Here we : • learn a network structure predictor jointly with module parameters themselves • extend visual primitives from previous work to reason over structured world representations Training data consists of ( world , question , answer ) triples : our approach requires no supervision of network layouts . We achieve state-of-the-art performance on two markedly different question answering tasks : one with questions about natural images , and another with more compositional questions about United States geography.1 We have introduced a new model , the dynamic neural module network , for answering queries about both structured and unstructured sources of information . Given only ( question , world , answer ) triples as training data , the model learns to assemble neural networks on the fly from an inventory of neural models , and simultaneously learns weights for these modules so that they can be composed into novel structures . Our approach achieves state-of-the-art results on two tasks . We believe that the success of this work derives from two factors : Continuous representations improve the expressiveness and learnability of semantic parsers : by replacing discrete predicates with differentiable neural network fragments , we bypass the challenging combinatorial optimization problem associated with induction of a semantic lexicon . In structured world representations , neural predicate representations allow the model to invent reusable attributes and relations not expressed in the schema . Perhaps more importantly , we can extend compositional questionanswering machinery to complex , continuous world representations like images . Semantic structure prediction improves generalization in deep networks : by replacing a fixed network topology with a dynamic one , we can tailor the computation performed to each problem instance , using deeper networks for more complex questions and representing combinatorially many queries with comparatively few parameters . In practice , this results in considerable gains in speed and sample efficiency , even with very little training data . These observations are not limited to the question answering domain , and we expect that they can be applied similarly to tasks like instruction following , game playing , and language generation .

Existing works on visual learning use manually-specified modular structures.

They propose a question-answering model trained jointly to translate from questions to dynamically assembled neural networks then produce answers with using images or knowledge bases.

The proposed model achieves state-of-the-arts on visual and structured domain datasets showing that coutinous representations improve the expressiveness and learnability of semantic parsers.

2020.aacl-main.88

Large pre-trained language models reach stateof-the-art results on many different NLP tasks when fine-tuned individually ; They also come with a significant memory and computational requirements , calling for methods to reduce model sizes ( green AI ) . We propose a twostage model-compression method to reduce a model 's inference time cost . We first decompose the matrices in the model into smaller matrices and then perform feature distillation on the internal representation to recover from the decomposition . This approach has the benefit of reducing the number of parameters while preserving much of the information within the model . We experimented on BERTbase model with the GLUE benchmark dataset and show that we can reduce the number of parameters by a factor of 0.4x , and increase inference speed by a factor of 1.45x , while maintaining a minimal loss in metric performance . Deep learning models have been demonstrated to achieve state-of-the-art results , but require large parameter storage and computation . It 's estimated that training a Transformer model with a neural architecture search has a CO 2 emissions equivalent to nearly five times the lifetime emissions of the average U.S. car , including its manufacturing ( Strubell et al . , 2019 ) . Alongside the increase in deep learning models complexity , in the NLP domain , there has been a shift in the NLP modeling paradigm from training a randomly initialized model to fine-tuning a large and computational heavy pre-trained language model ( Howard and Ruder , 2018 ; Peters et al . , 2018 ; Devlin et al . , 2018 ; Radford , 2018 ; Radford et al . , 2019 ; Dai et al . , 2019 ; Yang et al . , 2019 ; Lample and Conneau , 2019 ; Liu et al . , 2019b ; Raffel et al . , 2019 ; Lan et al . , 2019 ; Lewis et al . , 2019 ) . While re-using pre-trained models offsets the training costs , inference time costs of the finetuned models remain significant , and are showstoppers in many applications . The main challenge with pre-trained models is how can we reduce their size while saving the information contained within them . Recent work , approached this by keeping some of the layers while removing others ( Sanh et al . , 2019 ; Sun et al . , 2019 ; Xu et al . , 2020 ) . A main drawback of such approach is in its coarse-grained nature : removing entire layers might discard important information contained within the model , and working at the granularity of layers makes the trade-off between compression and accuracy of a model hard to control . Motivated by this , in this work we suggest a more finegrained approach which decomposes each matrix to two smaller matrices and then perform feature distillation on the internal representation to recover from the decomposition . This approach has the benefit of preserving much of the information while reducing the number of parameters . Alongside the advantage of preserving the information within each layer , there is also a memory flexibility advantage compared to removing entire layers ; As a result of decomposing each matrix to two smaller matrices , we can store each of the two matrices in two different memory blocks . This has the benefit of distributing the model matrices in many small memory blocks , which is useful when working in shared CPU-based environments . We evaluated our approach on the General Language Understanding Evaluation ( GLUE ) benchmark dataset ( Wang et al . , 2018 ) and show that our approach is superior or competitive in the different GLUE tasks to previous approaches which remove entire layers . Furthermore , we study the effects of different base models to decompose and show the superiority of decomposing a fine-tuned model compared to a pre-trained model or a ran-domly initialized model . Finally , we demonstrate the trade-off between compression and accuracy of a model . We presented a way to compress pre-trained large language models fine-tuned for specific tasks , while preserving much of the information contained within them , by using matrix decomposition to two small matrices . For future work it might be interesting to combine this approach with another approach such as pruning or quantization to achieve smaller models .

Existing coarse-grained approaches for reducing the inference time of pretraining models remove layers, posing a trade-off between compression and the accuracy of a model.

They propose a model-compression method which decompresses the matrix and performs feature distillation on the internal representations to recover from the decomposition.

The proposed method reduces the model size by 0.4x and increases inference speed by 1.45x while keeping the performance degradation minimum on the GLUE benchmark.

D16-1205

Several studies on sentence processing suggest that the mental lexicon keeps track of the mutual expectations between words . Current DSMs , however , represent context words as separate features , thereby loosing important information for word expectations , such as word interrelations . In this paper , we present a DSM that addresses this issue by defining verb contexts as joint syntactic dependencies . We test our representation in a verb similarity task on two datasets , showing that joint contexts achieve performances comparable to single dependencies or even better . Moreover , they are able to overcome the data sparsity problem of joint feature spaces , in spite of the limited size of our training corpus . Distributional Semantic Models ( DSMs ) rely on the Distributional Hypothesis ( Harris , 1954 ; Sahlgren , 2008 ) , stating that words occurring in similar contexts have similar meanings . On such theoretical grounds , word co-occurrences extracted from corpora are used to build semantic representations in the form of vectors , which have become very popular in the NLP community . Proximity between word vectors is taken as an index of meaning similarity , and vector cosine is generally adopted to measure such proximity , even though other measures have been proposed ( Weeds et al . , 2004 ; Santus et al . , 2016 ) . Most of DSMs adopt a bag-of-words approach , that is they turn a text span ( i.e. , a word window or a parsed sentence ) into a set of words and they register separately the co-occurrence of each word with a given target . The problem with this approach is that valuable information concerning word interrelations in a context gets lost , because words co-occurring with a target are treated as independent features . This is why works like Ruiz-Casado et al . ( 2005 ) , Agirre et al . ( 2009 ) and Melamud et al . ( 2014 ) proposed to introduce richer contexts in distributional spaces , by using entire word windows as features . These richer contexts proved to be helpful to semantically represent verbs , which are characterized by highly context-sensitive meanings , and complex argument structures . In fact , two verbs may share independent words as features despite being very dissimilar from the semantic point of view . For instance kill and heal share the same object nouns in The doctor healed the patient and the The poison killed the patient , but are highly different if we consider their joint dependencies as a single context . Nonetheless , richer contexts like these suffer from data sparsity , therefore requiring either larger corpora or complex smoothing processes . In this paper , we propose a syntactically savvy notion of joint contexts . To test our representation , we implement several DSMs and we evaluate them in a verb similarity task on two datasets . The results show that , even using a relatively small corpus , our syntactic joint contexts are robust with respect to data sparseness and perform similarly or better than single dependencies in a wider range of parameter settings . The paper is organized as follows . In Section 2 , we provide psycholinguistic and computational background for this research , describing recent models based on word windows . In Section 3 , we describe our reinterpretation of joint contexts with syntactic dependencies . Evaluation settings and results are presented in Section 4 . In this paper , we have presented our proposal for a new type of vector representation based on joint features , which should emulate more closely the general knowledge about event participants that seems to be the organizing principle of our mental lexicon . A core issue of previous studies was the data sparseness challenge , and we coped with it by means of a more abstract , syntactic notion of joint context . The models using joint dependencies were able at least to perform comparably to traditional , dependency-based DSMs . In our experiments , they even achieved the best correlation scores across several parameter settings , especially after the application of SVD . We want to emphasize that previous works such as Agirre et al . ( 2009 ) already showed that large word windows can have a higher discriminative power than indipendent features , but they did it by using a huge training corpus . In our study , joint context-based representations derived from a small corpus such as RCV1 are already showing competitive performances . This result strengthens our belief that dependencies are a possible solution for the data sparsity problem of joint feature spaces . We also believe that verb similarity might not be the best task to show the usefulness of joint contexts for semantic representation . The main goal of the present paper was to show that joint contexts are a viable option to exploit the full potential of distributional information . Our successful tests on verb similarity prove that syntactic joint contexts do not suffer of data sparsity and are also able to beat other types of representations based on independent word features . Moreover , syntactic joint contexts are much simpler and more competitive with respect to window-based ones . The good performance in the verb similarity task motivates us to further test syntactic joint contexts on a larger range of tasks , such as word sense disambiguation , textual entailment and classification of semantic relations , so that they can unleash their full potential . Moreover , our proposal opens interesting perspectives for computational psycholinguistics , especially for modeling those semantic phenomena that are inherently related to the activation of event knowledge ( e.g. thematic fit ) .

Providing richer contexts to Distributional Semantic Models improves by taking word interrelations into account but it would suffer from data sparsity.

They propose a Distributional Semantic Model that incorporates verb contexts as joint syntactic dependencies so that it emulates knowledge about event participants.

They show that representations obtained by the proposed model outperform more complex models on two verb similarity datasets with a limited training corpus.

2021.acl-long.57

In this paper , we propose Inverse Adversarial Training ( IAT ) algorithm for training neural dialogue systems to avoid generic responses and model dialogue history better . In contrast to standard adversarial training algorithms , IAT encourages the model to be sensitive to the perturbation in the dialogue history and therefore learning from perturbations . By giving higher rewards for responses whose output probability reduces more significantly when dialogue history is perturbed , the model is encouraged to generate more diverse and consistent responses . By penalizing the model when generating the same response given perturbed dialogue history , the model is forced to better capture dialogue history and generate more informative responses . Experimental results on two benchmark datasets show that our approach can better model dialogue history and generate more diverse and consistent responses . In addition , we point out a problem of the widely used maximum mutual information ( MMI ) based methods for improving the diversity of dialogue response generation models and demonstrate it empirically . In recent years , neural end-to-end dialogue response generation models ( Sordoni et al . , 2015 ; Serban et al . , 2016 ; Bordes et al . , 2016 ) has gained increasing popularity with the recent advancements of neural sequence-to-sequence ( seq2seq ) learning models ( Sutskever et al . , 2014 ; Vaswani et al . , 2017 ) . While neural dialogue models can generate seemingly fluent responses , due to the over-simplified maximum likelihood estimation ( MLE ) training objective and the high frequency of generic responses in training corpora , they tend to produce dull and generic responses such as " I do n't know " much more often than that humans generally do ( Li et al . , 2015 ) , which makes dialogue agents less engaging and ineffective . In addition , recent research on whether neural dialogue systems use dialogue history effectively ( Sankar et al . , 2019 ) shows that most neural dialogue agents fail to take the dialogue history into account when generating responses . This problem makes neural dialogue systems tend to generate responses irrelevant to the current topic of the conversation and are not consistent with the dialogue history . This problem may also intensify the generic response problem , as dull responses are generally off-topic and irrelevant to the dialogue history . To address the above issues , in this paper , we propose Inverse Adversarial Training ( IAT ) algorithm for training neural dialogue systems to avoid generic responses and model dialogue history better , thus generating diverse and informative responses . Conventional adversarial training methods generally generate label-preserving adversarial inputs with carefully designed methods and train the model to generate the same output to enhance the model 's robustness . In contrast , our approach perturbs in input dialogue history such that a good dialogue model should not generate the same output if the output is non-generic and relevant to the dialogue history . We name our proposed method as inverse adversarial training because it is related to conventional adversarial training methods which aim to improve the model 's adversarial robustness but our proposed objective is motivated in the opposite direction . Note that our work is not directly related to TextGANs as well as their applications on dialogue response generation . Specifically , the proposed inverse adversarial training assigns higher rewards to generated responses or ground-truth responses if their likeli-hood decreases more when the dialogue history is perturbed , and penalize the model when it generates responses whose likelihood is almost unchanged given either original or perturbed dialogue history as input . This encourages the model to generate more relevant and informative responses and capture dialogue history better . The proposed IAT algorithm can be used in both supervised and self-supervised fashion ( with / without reference response ) , which can be viewed as a form of reward-augmented maximum likelihood ( RAML ) method ( Norouzi et al . , 2016 ) that improves the original MLE objective or a rewarding scheme for RL-based text generation algorithms . The inverse adversarial learning framework is also conceptually related to self-adversarial learning ( Zhou et al . , 2020 ) where the the comparison is made between different checkpoints of the same model to provide reward for RL training of the NLG model . In addition , we identify a limitation of the widely-used maximum mutual information ( MMI ) based methods for improving the diversity of dialogue response generation models . This will be discussed in detail in section 2.1 and empirically demonstrated in section 4.2 . We conduct experiments on two dialogue datasets , OpenSubtitiles and DailyDialog , to demonstrate the effectiveness of the proposed approach . Experimental results show IAT helps neural dialogue systems model dialogue history better and generate more diverse and informative responses . In this work , we introduce inverse adversarial training ( IAT ) algorithm that is able to simultaneously reduce the dull response problem and help neural dialogue systems model dialogue history better . IAT measures the relevance and consistency of responses by the difference of their likelihood conditioning on either original and perturbed dialogue history . In this way , it is able to prevent the dialogue system from preferring generic responses , even they are often of high frequency in the training corpora . Our method also encourages the dialogue agent to model dialogue history better by penalizing the model when generating the same responses given perturbed dialogue history . Experimental results on two benchmark datasets show that the proposed inverse adversarial training algorithm helps dialogue models capture dialogue history better and generate more diverse and consistent responses . We also identify a limitation of the widely-used MMI based methods for improving the diversity of dialogue response generation models and empirically demonstrate the existence of this problem through our experimetns .

Neural end-to-end dialogue models generate fluent yet dull and generic responses without taking dialogue histories into account due to the over-simplified maximum likelihood estimation objective.

They propose an algorithm which encourages to be sensitive to perturbations in dialogue histories and generates more diverse and consistent responses by applying penalization.

The proposed approach can model dialogue history better and generate more diverse and consistent responses on OpenSubtitles and DailyDialog.

D09-1065

demonstrated that corpus-extracted models of semantic knowledge can predict neural activation patterns recorded using fMRI . This could be a very powerful technique for evaluating conceptual models extracted from corpora ; however , fMRI is expensive and imposes strong constraints on data collection . Following on experiments that demonstrated that EEG activation patterns encode enough information to discriminate broad conceptual categories , we show that corpus-based semantic representations can predict EEG activation patterns with significant accuracy , and we evaluate the relative performance of different corpus-models on this task . Models of semantic relatedness induced from corpus data have proven effective in a number of empirical tasks ( Sahlgren , 2006 ) and there is increasing interest in whether distributional information extracted from corpora correlates with aspects of speakers ' semantic knowledge : see Lund and Burgess ( 1996 ) , Landauer and Dumais ( 1997 ) , Almuhareb ( 2006 ) , Padó and Lapata ( 2007 ) , Schulte i m Walde ( 2008 ) , among many others . For this purpose , corpus models have been tested on datasets that are based on semantic judgements ( metalinguistic or meta-cognitive intuitions about synonymy , semantic distance , category-membership ) or behavioural experiments ( semantic priming , property generation , free association ) . While all these data are valuable , they are indirect reflections of semantic knowledge , and when the predictions they make diverge from those of corpora , interpretation is problematic : is the corpus model missing essential aspects of semantics , or are non-semantic factors biasing the data elicited from informants ? Reading semantic processes and representations directly from the brain would be an ideal way to get around these limitations . Until recently , analysis of linguistic quantities using neural data collected with EEG ( measurement at the scalp of voltages induced by neuronal firing ) or fMRI ( measurement of changes of oxygen concentrations in the brain tied to cognitive processes ) had neither the advantages of corpora ( scale ) nor of informants ( finer grained judgements ) . However , some clear patterns of differential activity have been found for broad semantic classes . Viewing images of natural ( typically animals and plants ) and non-natural ( typically artefacts like tools or vehicles ) objects elicits different loci of activity in fMRI ( Martin and Chao , 2001 ) and EEG ( Kiefer , 2001 ) , that persist across participants . Differences have also been found in response to auditorily or visually presented words of different lexical classes , such as abstract / concrete , and verb / noun ( Pulvermüller , 2002 ) . But interpretation of such group results remains somewhat difficult , as they may be consistent with more than one distinction : the natural / artefactual division just mentioned , may rather be between living / nonliving entities , dynamic / static entities , or be based on embodied experience ( e.g. manipulable or not ) . More recently , however , machine learning and other numerical techniques have been successfully applied to extract semantic information from neural data in a more discriminative fashion , down to the level of individual concepts . The work presented here builds on two strands of previous work : Murphy et al . ( 2008 ) use EEG data to perform semantic categorisation on single stimuli ; and Mitchell et al . ( 2008 ) introduce an fMRIbased method that detects word level distinctions by learning associations between features of neural activity and semantic features derived from a corpus . We combine these innovations by introducing a method that extracts featural representations from the EEG signal , and uses corpusbased models to predict word level distinctions in patterns of EEG activity . The proposed method achieves a performance level significantly above chance ( also when distinguishing between concepts from the same semantic category , e.g. , dog and cat ) , and approaching that achieved with fMRI . The paper proceeds as follows . The next section describes a simple behavioural experiment where Italian-speaking participants had to name photographic images of mammals and tools while their EEG activity was being recorded , and continues to detail how the rich and multidimensional signals collected were reduced to a small set of optimally informative features using a new method . Section 3 describes a series of corpus-based semantic models derived from both a raw-text web corpus , and from various parsings of a conventional corpus . In Section 4 we describe the training of a series of linear models , that each learn the associations between a set of corpus semantic features and an individual EEG activity feature . By combining these models it is possible to predict the EEG activity pattern for a single unseen word , and compare this to the observed pattern for the corresponding concept . Results ( Section 5 ) show that these predictions succeed at a level significantly above chance , both for coarser distinctions between words in different superordinate categories ( e.g. , differentiating between drill and gorilla ) , and , at least for the model based on the larger web corpus , for those within the same category ( e.g. , drill vs spanner , koala vs gorilla ) . Our results show that corpus-extracted conceptual models can be used to distinguish between the EEG activation levels associated with conceptual categories to a degree that is significantly above chance . Though category specific patterns are detectable in the EEG signal alone ( as illustrated by the PCA analysis in figure 3 ) , on that basis we can not be sure that semantics is being detected . Some other property of the stimuli that co-varies with the semantic classes of interest could be responsible , such as visual complexity , conceptual familiarity , lexical frequency , or phonological form . Only by cross-training with individual corpus features and showing that these hold a predictive relationship to neural activity have we been able to establish that EEG patterns encode semantics . Present evidence indicates that fMRI may provide richer data for training such models than EEG ( Mitchell and colleagues obtain an average accuracy of 77 % , and 65 % for the within category setting ) . However , fMRI has several clear disadvantages as a tool for language researchers . First of all , the fine spatial resolution it provides ( down to 2 - 3 mm ) , while of great interest to neuroscientists , is not in itself linguistically informative . Its coarse temporal resolution ( of the order of several seconds ) , makes it ill-suited to analysing on-line linguistic processes . EEG on the other hand , despite its low spatial resolution ( several centimetres ) , gives millisecond-level temporal resolution , Finally , there is no prospect of fMRI being miniaturised , while wearable EEG systems are already becoming commercially available , making experimentation in more ecological settings a possibility ( e.g. , playing with a child , meeting at a desk , walking around ) . In short , while EEG can be used to carry out systematic investigations of categorical distinctions , doing so with fMRI would be prohibitively expensive . Present results indicate that distinctions between categories are easier than distinctions between category elements ; and that selecting the conceptual features by hand gives better results than discovering them automatically . Both of these results however may be due to limitations of the current method . One limitation is that we have been using the same set of features for all concepts , which is likely to blur the distinctions between members of a category more than those between categories . A second limitation of our present methodology is that it is constrained to use very small numbers of semantic features , which limits its applicability . For example it is hard to conceive of a small set of verbs , or other parts-ofspeech , whose co-occurrence patterns could successfully characterise the full range of meaning found in the human lexicon . Even the more economical corpus-extracted conceptual models tend to run in the hundreds of features ( Almuhareb , 2006 ) . We are currently working on variations in the method that will address these shortcomings . The web-based model with manually picked features outperformed all la Repubblica-based models . However , the results attained with repubblica-window-svd are encouraging , especially considering that we are reporting results for an EEG feature configuration optimised for the web data ( see footnote 5 ) , and that la Repubblica is several orders of magnitude smaller than the web . That data sparseness might be the main issue with la Repubblica models is suggested by the fact that repubblica-window-svd is the least sparse of them , since it does not filter data by position or dependency path , and compresses information from many verbs via SVD . In future research , we plan to extract richer models from larger corpora . And as the discriminative accuracy of crosstraining techniques improves , further insights into the relative validity of corpus representations will be attainable . One research aim is to see if individual corpus semantic properties are encoded neurally , so providing strong evidence for a particular model . These techniques may also prove more objective and reliable in evaluating representations of abstract concepts , for which it is more difficult to collect reliable judgements from informants .

The expensive cost of using fMRI hinders studies on the relationship between corpus-extracted models of semantic knowledge and neural activation patterns.

They propose to use EEG activation patterns instead of fMRI to reduce the cost.

They show that using EEG signals with corpus-based models, they can predict word level distinctions significantly above chance.

D09-1085

This paper introduces a new parser evaluation corpus containing around 700 sentences annotated with unbounded dependencies , from seven different grammatical constructions . We run a series of off-theshelf parsers on the corpus to evaluate how well state-of-the-art parsing technology is able to recover such dependencies . The overall results range from 25 % accuracy to 59 % . These low scores call into question the validity of using Parseval scores as a general measure of parsing capability . We discuss the importance of parsers being able to recover unbounded dependencies , given their relatively low frequency in corpora . We also analyse the various errors made on these constructions by one of the more successful parsers . Statistical parsers are now obtaining Parseval scores of over 90 % on the WSJ section of the Penn Treebank ( Bod , 2003 ; Petrov and Klein , 2007 ; Huang , 2008 ; Carreras et al . , 2008 ) . McClosky et al . ( 2006 ) report an F-score of 92.1 % using selftraining applied to the reranker of Charniak and Johnson ( 2005 ) . Such scores , in isolation , may suggest that statistical parsing is close to becoming a solved problem , and that further incremental improvements will lead to parsers becoming as accurate as POS taggers . A single score in isolation can be misleading , however , for a number of reasons . First , the single score is an aggregate over a highly skewed distribution of all constituent types ; evaluations which look at individual constituent or dependency types show that the accuracies on some , semantically important , constructions , such as coordination and PP-attachment , are much lower ( Collins , 1999 ) . Second , it is well known that the accuracy of parsers trained on the Penn Treebank degrades when they are applied to different genres and domains ( Gildea , 2001 ) . Finally , some researchers have argued that the Parseval metrics ( Black et al . , 1991 ) are too forgiving with respect to certain errors and that an evaluation based on syntactic dependencies , for which scores are typically lower , is a better test of parser performance ( Lin , 1995 ; Carroll et al . , 1998 ) . In this paper we focus on the first issue , that the performance of parsers on some constructions is much lower than the overall score . The constructions that we focus on are various unbounded dependency constructions . These are interesting for parser evaluation for the following reasons : one , they provide a strong test of the parser 's knowledge of the grammar of the language , since many instances of unbounded dependencies are difficult to recover using shallow techniques in which the grammar is only superficially represented ; and two , recovering these dependencies is necessary to completely represent the underlying predicateargument structure of a sentence , useful for applications such as Question Answering and Information Extraction . To give an example of the sorts of constructions we are considering , and the ( in)ability of parsers to recover the corresponding unbounded dependencies , none of the parsers that we have tested were able to recover the dependencies shown in bold from the following sentences : We have also developed techniques for recognizing and locating underground nuclear tests through the waves in the ground which they generate . By Monday , they hope to have a sheaf of documents both sides can trust . In this paper we have demonstrated that current parsing technology is poor at recovering some of the unbounded dependencies which are crucial for fully representing the underlying predicateargument structure of a sentence . We have also argued that correct recovery of such dependencies will become more important as parsing technology improves , despite the relatively low frequency of occurrence of the corresponding grammatical constructions . We also see this more focused parser evaluation methodology -in this case construction-focused -as a way of improving parsing technology , as an alternative to the exclusive focus on incremental improvements in overall accuracy measures such as Parseval .

While recent statistical parsers perform well on Penn Treebank, the results can be misleading due to several reasons originating from evaluation and datasets.

They propose a new corpus with unbounded dependencies from difference grammatical constructions.

Their evaluation of existing parsers with the proposed corpus shows lower scores than reported in previous works indicating a poor ability to recover unbounded dependencies.

P12-1013

Learning entailment rules is fundamental in many semantic-inference applications and has been an active field of research in recent years . In this paper we address the problem of learning transitive graphs that describe entailment rules between predicates ( termed entailment graphs ) . We first identify that entailment graphs exhibit a " tree-like " property and are very similar to a novel type of graph termed forest-reducible graph . We utilize this property to develop an iterative efficient approximation algorithm for learning the graph edges , where each iteration takes linear time . We compare our approximation algorithm to a recently-proposed state-of-the-art exact algorithm and show that it is more efficient and scalable both theoretically and empirically , while its output quality is close to that given by the optimal solution of the exact algorithm . Performing textual inference is in the heart of many semantic inference applications such as Question Answering ( QA ) and Information Extraction ( IE ) . A prominent generic paradigm for textual inference is Textual Entailment ( TUE ) ( Dagan et al . , 2009 ) . In TUE , the goal is to recognize , given two text fragments termed text and hypothesis , whether the hypothesis can be inferred from the text . For example , the text " Cyprus was invaded by the Ottoman Empire in 1571 " implies the hypothesis " The Ottomans attacked Cyprus " . Semantic inference applications such as QA and IE crucially rely on entailment rules ( Ravichandran and Hovy , 2002 ; Shinyama and Sekine , 2006 ) or equivalently inference rules , that is , rules that describe a directional inference relation between two fragments of text . An important type of entailment rule specifies the entailment relation between natural language predicates , e.g. , the entailment rule ' X invade Y → X attack Y ' can be helpful in inferring the aforementioned hypothesis . Consequently , substantial effort has been made to learn such rules ( Lin and Pantel , 2001 ; Sekine , 2005 ; Szpektor and Dagan , 2008 ; Schoenmackers et al . , 2010 ) . Textual entailment is inherently a transitive relation , that is , the rules ' x → y ' and ' y → z ' imply the rule ' x → z ' . Accordingly , Berant et al . ( 2010 ) formulated the problem of learning entailment rules as a graph optimization problem , where nodes are predicates and edges represent entailment rules that respect transitivity . Since finding the optimal set of edges respecting transitivity is NP-hard , they employed Integer Linear Programming ( ILP ) to find the exact solution . Indeed , they showed that applying global transitivity constraints improves rule learning comparing to methods that ignore graph structure . More recently , Berant et al . ( Berant et al . , 2011 ) introduced a more efficient exact algorithm , which decomposes the graph into connected components and then applies an ILP solver over each component . Despite this progress , finding the exact solution remains NP-hard -the authors themselves report they were unable to solve some graphs of rather moderate size and that the coverage of their method is limited . Thus , scaling their algorithm to data sets with tens of thousands of predicates ( e.g. , the extractions of Fader et al . ( 2011 ) ) is unlikely . In this paper we present a novel method for learning the edges of entailment graphs . Our method computes much more efficiently an approximate solution that is empirically almost as good as the exact solution . To that end , we first ( Section 3 ) conjecture and empirically show that entailment graphs exhibit a " tree-like " property , i.e. , that they can be reduced into a structure similar to a directed forest . Then , we present in Section 4 our iterative approximation algorithm , where in each iteration a node is removed and re-attached back to the graph in a locally-optimal way . Combining this scheme with our conjecture about the graph structure enables a linear algorithm for node re-attachment . Section 5 shows empirically that this algorithm is by orders of magnitude faster than the state-of-the-art exact algorithm , and that though an optimal solution is not guaranteed , the area under the precision-recall curve drops by merely a point . To conclude , the contribution of this paper is twofold : First , we define a novel modeling assumption about the tree-like structure of entailment graphs and demonstrate its validity . Second , we exploit this assumption to develop a polynomial approximation algorithm for learning entailment graphs that can scale to much larger graphs than in the past . Finally , we note that learning entailment graphs bears strong similarities to related tasks such as Taxonomy Induction ( Snow et al . , 2006 ) and Ontology induction ( Poon and Domingos , 2010 ) , and thus our approach may improve scalability in these fields as well . Learning large and accurate resources of entailment rules is essential in many semantic inference applications . Employing transitivity has been shown to improve rule learning , but raises issues of efficiency and scalability . The first contribution of this paper is a novel modeling assumption that entailment graphs are very similar to FRGs , which is analyzed and validated empirically . The main contribution of the paper is an efficient polynomial approximation algorithm for learning entailment rules , which is based on this assumption . We demonstrate empirically that our method is by orders of magnitude faster than the state-of-the-art exact algorithm , but still produces an output that is almost as good as the optimal solution . We suggest our method as an important step towards scalable acquisition of precise entailment resources . In future work , we aim to evaluate TNF on large graphs that are automatically generated from huge corpora . This of course requires substantial efforts of pre-processing and test-set annotation . We also plan to examine the benefit of TNF in learning similar structures , e.g. , taxonomies or ontologies .

Current inefficient algorithms aim to obtain entailment rules for semantic inference hindering the use of large resources.

They propose an efficient polynomial approximation algorithm that exploits their observation, entailment graphs have a "tree-like" property.

Their iterative algorithm runs by orders of magnitude faster than current exact state-of-the-art solutions while keeping close quality.

D15-1054

Sponsored search is at the center of a multibillion dollar market established by search technology . Accurate ad click prediction is a key component for this market to function since the pricing mechanism heavily relies on the estimation of click probabilities . Lexical features derived from the text of both the query and ads play a significant role , complementing features based on historical click information . The purpose of this paper is to explore the use of word embedding techniques to generate effective text features that can capture not only lexical similarity between query and ads but also the latent user intents . We identify several potential weaknesses of the plain application of conventional word embedding methodologies for ad click prediction . These observations motivated us to propose a set of novel joint word embedding methods by leveraging implicit click feedback . We verify the effectiveness of these new word embedding models by adding features derived from the new models to the click prediction system of a commercial search engine . Our evaluation results clearly demonstrate the effectiveness of the proposed methods . To the best of our knowledge this work is the first successful application of word embedding techniques for the task of click prediction in sponsored search . Sponsored search is a multibillion dollar market ( Easley and Kleinberg , 2010 ) that makes most search engine revenue and is one of the most successful ways for advertisers to reach their intended audiences . When search engines deliver results to a user , sponsored advertisement impressions ( ads ) are shown alongside the organic search results ( Figure 1 ) . Typically the advertiser pays the search engine based on the pay-per-click model . In this model the advertiser pays only if the impression that accompanies the search results is clicked . The price is usually set by a generalized second-price ( GSP ) auction ( Edelman et al . , 2005 ) that encourages advertisers to bid truthfully . An advertiser wins if the expected revenue for this advertiser , which is the bid Figure 1 : Sponsored ads when " pizza " was searched at Yahoo ! ( www.yahoo.com ) . price times the expected click probability ( also know as click through rate , or CTR ) , is ranked the highest . The price the advertiser pays , known as cost-per-click ( CPC ) , is the bid price for the second ranked advertiser times the ratio of the expected CTR between the second and first ranked advertisers . From this discussion it should be clear that CTR plays a key role in deciding both the ranking and the pricing of the ads . Therefore it is very important to predict CTR accurately . The state of the art search engine typically uses a machine learning model to predict CTR by exploiting various features that have been found useful in practice . These include historical click performance features such as historical click probability for the query , the ad , the user , and a combination of these ; contextual features such as temporal and geographical information ; and text-based features such as query keywords or ad title and description . Among these , historical click performance features often have the most predictive power for queries , ads and users that have registered many impressions . For queries , ads and users that have not registered many impressions , however , historical CTR may have too high a variance to be useful . Hillard et al . ( 2011 ) observed that the number of impressions and clicks recorded on query-ad pairs have a very long tail : only 61 % of queries has greater than three clicks . They also reported a drastic drop in the accuracy of the click prediction model when fewer historical observations are available . Furthermore , fine-grained historical CTR information takes a huge amount of space , which makes it costly to maintain . On the other hand , text features are always readily available , and thus are particularly useful for those cases for which there is insufficient historical information . Multiple researchers , for example ( Richardson , 2007 ; Cheng and Cantú-Paz , 2010 ) , reported the usage of text features including simple lexical similarity scores between the query and ads , word or phrase overlaps and the number of overlapping words and characters . Such features rely on the assumption that query-ad overlap is correlated with perceived relevance . While this is true to a certain extent , the use of simple lexical similarity can not capture semantic information such as synonyms , entities of the same type and strong relationships between entities ( e.g. CEO-company , brandmodel , part-of ) . Recently a host of studies on word embedding have been conducted ; all map words into a vector space such that semantically relevant words are placed near each other in the space ( Mikolov et al . , 2013a ; Pennington et al . , 2014 ; Baroni et al . , 2014 ) . The use of continuous word vectors has been shown to be helpful for a wide range of NLP tasks by better capturing both syntactic and semantic information than simple lexical features ( Socher et al . , 2012a ) . No previous research on sponsored search has successfully used word embeddings to generate text features . In this paper , we explore the use of word embeddings for click prediction . However , it is clear that conventional word embeddings ( which solely rely on word co-occurrence in a context window ) can only offer limited discriminative power because queries and ad text are typically very short . In addition , conventional word embeddings can not capture user intents , preferences and desires . Wang et al . ( 2013 ) showed that specific frequently occurring lexical patterns , e.g. , x% off , guaranteed return in x days and official site , are effective in triggering users desires , and thus lead to significant differences in CTR . Conventional word embeddings can not capture these phenomena since they do not incorporate the implicit feedback users provide through clicks and non-clicks . These observations naturally lead us to leverage click feedback to infuse users ' intentions and desires into the vector space . The simplest way to harness click feedback is to train conventional word embedding models on a corpus that only includes clicked impressions , where each " sentence " is constructed by mixing the query and ad text . Having trained a word embedding model , we simply take the average of word vectors of the query and ads respectively to obtain sentence ( or paragraph ) vectors , which in turn are used to compute the similarity scores between the query and ads . Our experiments show that this method does improve click prediction performance . However , this method has several potential weaknesses . First , the use of only clicked impressions ignores the large amount of negative signals contained in the non-clicked ad impressions . Second , the use of indirect signals ( word co-occurrences ) can be noisy or even harmful to our ultimate goal ( accurate click prediction ) when it is combined with direct signals ( impressions with click feedback ) . Third , without explicit consideration about the averaging step in the training process of word embedding models , a simple averaging scheme across word vectors may be a suboptimal . We therefore propose several joint word embedding models ; all of these aim to put query vectors close to relevant ad vectors by explicitly utilizing both positive and negative click feedback . We evaluate all these models against a large sponsored search data set from a commercial search engine , and demonstrate that our proposed models significantly improve click prediction performance . The rest of this paper is organized as follows . In Section 2 we present a brief summary of related work . In Section 3 we give some background information on ad click prediction in sponsored search . In Section 4 we describe our methods . In Section 5 we discuss our experiments . We finish with some conclusions and future directions in Section 6 . In this paper we explored the use of word embedding techniques to overcome the shortcomings of traditional lexical features for ad click prediction in sponsored search . We identified several potential weaknesses of the plain application of conventional word embedding methodologies : the lack of the right machinery to harness both positive and negative click feedback , the limited utility of pure word co-occurrence signals , and no consideration of vector composition in the word embedding training process . We proposed a set of new implicit feedback-based joint word embedding methods to address those issues . We evaluated the new word embedding methods in the context of a very good baseline click prediction system , on a large scale data set collected from Yahoo ! search engine logs . Our experimental results clearly demonstrate the effectiveness of the proposed methods . We also presented several examples for qualitative analysis to advance our understanding on how each algorithm really contributes to the improved performance . To the best of our knowledge this work is the first successful application of word embedding techniques for the sponsored search task . There are multiple interesting research directions for future work . One of these directions is to extend the vocabulary by identifying significant phrases ( as well as words ) before training word vectors . Hillard et al . ( 2011 ) employed Conditional Random Fields to divide queries with multiple words into segments and collected historical CTR on the segment level . We also like to investigate more structured embedding methods such as RNNs ( probably for ad descriptions ) . In case the computational cost of such methods are too high to be practical for sponsored search , we can employ them only for a small fraction of ads filtered by faster methods . It may be possible to deal with the implicit negative feedback of unclicked ad impressions in a more principled way by adopting ranking-based loss functions . However , this is only possible with the extra cost of identifying and aggregating related ads into a single transaction . Though not directly related to NLP , yet another promising direction is to jointly embed not only text data but also a variety of user activities ( e.g. , organic search results , mobile app usages , other daily activities ) all together in the same vector space . Since many of the different sources contain their own unique information , we might be able to obtain a much better understanding about the user state and intent through this rich joint embedding space . Joint embedding with rich information can also help us to perform automatic clustering of users , eventually leading to powerful smoothing methods for personalized historical CTR statistics .

Conventional word embeddings with a simple integration of click feedback information and averaging to obtain sentence representations do not work well for ad click prediction.

They propose several joint word embedding methods to leverage positive and negative click feedback which put query vectors close to relevant ad vectors.

The use of features obtained from the new models improves on a large sponsored search data of commercial Yahoo! search engine.

D09-1072

We propose a new model for unsupervised POS tagging based on linguistic distinctions between open and closed-class items . Exploiting notions from current linguistic theory , the system uses far less information than previous systems , far simpler computational methods , and far sparser descriptions in learning contexts . By applying simple language acquisition techniques based on counting , the system is given the closed-class lexicon , acquires a large open-class lexicon and then acquires disambiguation rules for both . This system achieves a 20 % error reduction for POS tagging over state-of-the-art unsupervised systems tested under the same conditions , and achieves comparable accuracy when trained with much less prior information . All recent research on unsupervised tagging , as well as the majority of work on supervised taggers , views POS tagging as a sequential labeling problem and treats all POS tags , both closed-and open-class , as roughly equivalent . In this work we explore a different understanding of the tagging problem , viewing it as a process of first identifying functional syntactic contexts , which are flagged by closed-class items , and then using these functional contexts to determine the POS labels . This disambiguation model differs from most previous work in three ways : 1 ) it uses different encodings over two distinct domains ( roughly open-and closed-class words ) with complementary distribution ( and so decodes separately ) ; 2 ) it is deterministic and 3 ) it is non-lexicalized . By learning disambiguation models for open-and closed-classes separately , we found that the deterministic , rulebased model can be learned from unannotated data by a simple strategy of selecting a rule in each appropriate context with the highest count . In contrast to this , most previous work on unsupervised tagging ( especially for English ) concentrates on improving the parameter estimation techniques for training statistical disambiguation models from unannotated data . For example , ( Smith&Eisner , 2005 ) proposes contrastive estimation ( CE ) for log-linear models ( CRF ) , achieving the current state-of-the-art performance of 90.4 % ; ( Goldwater&Griffiths , 2007 ) applies a Bayesian approach to improve maximumlikelihood estimation ( MLE ) for training generative models ( HMM ) . In the main experiments of both of these papers , the disambiguation model is learned , but the algorithms assume a complete knowledge of the lexicon with all possible tags for each word . In this work , we propose making such a large lexicon unnecessary by learning the bulk of the lexicon along with learning a disambiguation model . Little previous work has been done on this natural and simple idea because the clusters found by previous induction schemes are not in line with the lexical categories that we care about . ( Chan , 2008 ) is perhaps the first with the intention of generating " a discrete set of clusters . " By applying similar techniques to ( Chan , 2008 ) , which we discuss later , we can generate clusters that closely approximate the central open-class lexical categories , a major advance , but we still require a closed-class lexicon specifying possible tags for these words . This asymmetry in our lexicon acquisition model conforms with our understanding of natural language as structured data over two distinct domains with complementary distribution : open-class ( lexical ) and closed-class ( functional ) . Provided with only a closed-class lexicon of 288 words , about 0.6 % of the full lexicon , the system acquires a large open-class lexicon and then acquires disambiguation rules for both closed-and open-class words , achieving a tagging accuracy of 90.6 % for a 24k dataset , as high as the current state-of-the-art ( 90.4 % ) achieved with a complete dictionary . In the test condition where both algorithms are provided with a full lexicon , and are trained and evaluated over the same 96k dataset , we reduce the tagging error by up to 20 % . In Section 2 we explain our understanding of the POS tagging problem in detail and define the notions of functional context and open-and closedclass elements . Then we will introduce our methods for acquiring the lexicon ( Section 3 ) and learning disambiguation models ( Section 4 , 5 and 6 ) step by step . Results are reported in Section 7 followed by Section 8 which discusses the linguistic motivation behind this work and the simplicity and efficiency of our model . In this work on unsupervised tagging , we combine lexicon acquisition with the learning of a POS disambiguation model . Moreover , the disambiguation model we used is deterministic , nonlexicalized and defined over two distinct domains with complementary distribution ( open-and closed-class ) . Building a lexicon based on induced clusters requires our morphological knowledge of three special endings in English : -ing , -ed and -s ; on the other hand , to reduce the feature space used for category induction , we utilize vectors of functional features only , exploiting our knowledge of the role of determiners and modal verbs . However , the above information is restricted to the lexicon acquisition model . Taking a lexicon as input , which either consists of a known closed-class lexicon together with an acquired open-class lexicon or is composed by automatic extraction from the Penn Treebank , we need NO language-specific knowledge for learning the disambiguation model . We would like to point the reader to ( Chan , 2008 ) for more discussion on Category induction14 ; and discussions below will concentrate on the proposed disambiguation model . Current Chomskian theory , developed in the Minimalist Program ( MP ) ( Chomsky , 2006 ) , argues ( very roughly speaking ) that the syntactic structure of a sentence is built around a scaffolding provided by a set of functional elements15 . Each of these provides a large tree fragment ( roughly corresponding to what Chomsky calls a phase ) that provide the piece parts for full utterances . Chomsky observes that when these fragments combine , only the very edge of the fragments can change and that the internal structure of these fragments is rigid ( he labels this observation the Phase Impenetrability Condition , PIC ) . With the belief in PIC , we propose the concept of functional context , in which category property can be determined ; also we notice the distinct distribution of the elements ( functional ) on the edge of phase and those ( lexical ) assembled within the phase . Instead of chasing the highest possible performance by using the strongest method possible , we wanted to explore how well a deterministic , nonlexicalized model , following certain linguistic intuitions , can approach the NLP problem . For the unsupervised tagging task , this simple model , with less than two hundred rules learned , even outperforms non-deterministic generative models with ten of thousands of parameters . Another motivation for our pursuit of this deterministic , non-lexicalized model is computational efficiency 16 . It takes less than 3 minutes total for our model to acquire the lexicon , learn the disambiguation model , tag raw data and evaluate the output for a 96k dataset on a small laptop17 . And a model using only counting and selecting is common in the research field of language acquisition and perhaps more compatible to the way humans process language . We are certainly aware that our work does not yet address two problems : 1 ) . How the system can be adapted to work for other languages and 2 ) How to automatically obtain the knowledge of functional elements . We believe that , given the proper understanding of functional elements , our system will be easily adapted to other languages , but we clearly need to test this hypothesis . Also , we are highly interested in completing our system by incorporating the acquisition of functional elements . ( Chan , 2008 ) presents an extensive discussion of his work on morphological induction and ( Mintz et al . , 2002 ) presents interesting psychological experiments we can build on to acquire closed-class words .

Current approaches tackle unsupervised POS tagging as a sequential labelling problem and require a complete knowledge of the lexicon.

They propose to first identify functional syntactic contexts and then use them to make predictions for POS tagging.

The proposed method achieves equivalent performance by using 0.6% of the lexical knowledge used in baseline models.

2021.naacl-main.458

Non-autoregressive Transformer is a promising text generation model . However , current non-autoregressive models still fall behind their autoregressive counterparts in translation quality . We attribute this accuracy gap to the lack of dependency modeling among decoder inputs . In this paper , we propose CNAT , which learns implicitly categorical codes as latent variables into the non-autoregressive decoding . The interaction among these categorical codes remedies the missing dependencies and improves the model capacity . Experiment results show that our model achieves comparable or better performance in machine translation tasks than several strong baselines . Non-autoregressive Transformer ( NAT , Gu et al . , 2018 ; Wang et al . , 2019 ; Lee et al . , 2018 ; Ghazvininejad et al . , 2019 ) is a promising text generation model for machine translation . It introduces the conditional independent assumption among the target language outputs and simultaneously generates the whole sentence , bringing in a remarkable efficiency improvement ( more than 10× speed-up ) versus the autoregressive model . However , the NAT models still lay behind the autoregressive models in terms of BLEU ( Papineni et al . , 2002 ) for machine translation . We attribute the low-quality of NAT models to the lack of dependencies modeling for the target outputs , making it harder to model the generation of the target side translation . A promising way is to model the dependencies of the target language by the latent variables . A line of research works ( Kaiser et al . , 2018 ; Roy et al . , 2018 ; Shu et al . , 2019 ; Ma et al . , 2019 ) introduce latent variable modeling to the non-autoregressive Transformer and improves translation quality . The latent variables could be regarded as the springboard to bridge the modeling gap , introducing more informative decoder inputs than the previ-ously copied inputs . More specifically , the latentvariable based model first predicts a latent variable sequence conditioned on the source representation , where each variable represents a chunk of words . The model then simultaneously could generate all the target tokens conditioning on the latent sequence and the source representation since the target dependencies have been modeled into the latent sequence . However , due to the modeling complexity of the chunks , the above approaches always rely on a large number ( more than 2 15 , Kaiser et al . , 2018 ; Roy et al . , 2018 ) of latent codes for discrete latent spaces , which may hurt the translation efficiencythe essential goal of non-autoregressive decoding . Akoury et al . ( 2019 ) introduce syntactic labels as a proxy to the learned discrete latent space and improve the NATs ' performance . The syntactic label greatly reduces the search space of latent codes , leading to a better performance in both quality and speed . However , it needs an external syntactic parser to produce the reference syntactic tree , which may only be effective in limited scenarios . Thus , it is still challenging to model the dependency between latent variables for non-autoregressive decoding efficiently . In this paper , we propose to learn a set of latent codes that can act like the syntactic label , which is learned without using the explicit syntactic trees . To learn these codes in an unsupervised way , we use each latent code to represent a fuzzy target category instead of a chunk as the previous research ( Akoury et al . , 2019 ) . More specifically , we first employ vector quantization ( Roy et al . , 2018 ) to discretize the target language to the latent space with a smaller number ( less than 128 ) of latent variables , which can serve as the fuzzy word-class information each target language word . We then model the latent variables with conditional random fields ( CRF , Lafferty et al . , 2001 ; Sun et al . , 2019 ) . To avoid the mismatch of the training and inference for latent variable modeling , we propose using a gated neural network to form the decoder inputs . Equipping it with scheduled sampling ( Bengio et al . , 2015 ) , the model works more robustly . Experiment results on WMT14 and IWSLT14 show that CNAT achieves the new state-of-theart performance without knowledge distillation . With the sequence-level knowledge distillation and reranking techniques , the CNAT is comparable to the current state-of-the-art iterative-based model while keeping a competitive decoding speedup . We propose CNAT , which implicitly models the categorical codes of the target language , narrowing the performance gap between the nonautoregressive decoding and autoregressive decoding . Specifically , CNAT builds upon the latent Transformer and models the target-side categorical information with vector quantization and conditional random fields ( CRF ) model . We further employ a gated neural network to form the decoder inputs . Equipped with the scheduled sampling , CNAT works more robust . As a result , the CNAT achieves a significant improvement and moves closer to the performance of the Transformer on machine translation . For each latent variable , we list the top 3 frequent pos tags and their corresponding percentages .

Non-autoregressive translation models fall behind their autoregressive counterparts in translation quality due to the lack of dependency modelling for the target outputs.

They propose a non-autoregressive transformer-based model which implicitly learns categorical codes as latent variables into the decoding to complement missing dependencies.

The proposed model achieves state-of-the-art without knowledge distillation and a competitive decoding speedup with an interactive-based model when coupled with knowledge distillation and reranking techniques.

2021.emnlp-main.765

The clustering-based unsupervised relation discovery method has gradually become one of the important methods of open relation extraction ( OpenRE ) . However , high-dimensional vectors can encode complex linguistic information which leads to the problem that the derived clusters can not explicitly align with the relational semantic classes . In this work , we propose a relationoriented clustering model and use it to identify the novel relations in the unlabeled data . Specifically , to enable the model to learn to cluster relational data , our method leverages the readily available labeled data of pre-defined relations to learn a relationoriented representation . We minimize distance between the instance with same relation by gathering the instances towards their corresponding relation centroids to form a cluster structure , so that the learned representation is cluster-friendly . To reduce the clustering bias on predefined classes , we optimize the model by minimizing a joint objective on both labeled and unlabeled data . Experimental results show that our method reduces the error rate by 29.2 % and 15.7 % , on two datasets respectively , compared with current SOTA methods . Relation extraction ( RE ) , a crucial basic task in the field of information extraction , is of the utmost practical interest to various fields including web search ( Xiong et al . , 2017 ) , knowledge base completion ( Bordes et al . , 2013 ) , and question answering ( Yu et al . , 2017 ) . However , conventional RE paradigms such as supervision and distant supervision are generally designed for pre-defined relations , which can not deal with new emerging relations in the real world . Under this background , open relation extraction ( OpenRE ) has been widely studied for its use Figure 1 : Although both instances S 2 and S 3 express founded relation while S 1 expresses CEO relation , the distance between S 1 and S 2 is still smaller than that between S 2 and S 3 . This is because there may be more similar surface information ( e.g. word overlapping ) or syntactic structure between S 1 and S 2 , thus the derived clusters can not explicitly align with relations . in extracting new emerging relational types from open-domain corpora . The approaches used to handle open relations roughly fall into one of two groups . The first group is open information extraction ( OpenIE ) ( Etzioni et al . , 2008 ; Yates et al . , 2007 ; Fader et al . , 2011 ) , which directly extracts related phrases as representations of different relational types . However , if not properly canonicalized , the extracted relational facts can be redundant and ambiguous . The second group is unsupervised relation discovery ( Yao et al . , 2011 ; Shinyama and Sekine , 2006 ; Simon et al . , 2019 ) . In this type of research , much attention has been focused on unsupervised clustering-based RE methods , which cluster and recognize relations from high-dimensional representations ( Elsahar et al . , 2017 ) . Recently , the self-supervised signals in pretrained language model are further exploited for clustering optimization ( Hu et al . , 2020 ) . However , many studies show that highdimensional embeddings can encode complex linguistic information such as morphological ( Peters et al . , 2018 ) , local syntactic ( Hewitt and Manning , 2019 ) , and longer range semantic information ( Jawahar et al . , 2019 ) . Consequently , the distance of representation is not completely consistent with relational semantic similarity . Although Hu et al . ( 2020 ) use self-supervised signals to optimize clustering , there is still no guarantee that the learned clusters will explicitly align with the desired relational semantic classes ( Xing et al . , 2002 ) . As shown in Figure 1 , we use the method proposed by Hu et al . ( 2020 ) to get the instance representations . Although both instances S 2 and S 3 express the founded relation , the euclidean distance between them is larger than that between S 1 and S 2 , which express different relation . Obviously , the clustering algorithm tends to group instances S 1 and S 2 together , rather than S 2 and S 3 which express the same relation . In this work , we propose a relation-oriented clustering method . To enable the model to learn to cluster relational data , pre-defined relations and their existing labeled instances are leveraged to optimize a non-linear mapping , which transforms high-dimensional entity pair representations into relation-oriented representations . Specifically , we minimize distance between the instances with same relation by gathering the instances representation towards their corresponding relation centroids to form the cluster structure , so that the learned representation is cluster-friendly . In order to reduce the clustering bias on the predefined classes , we iteratively train the entity pair representations by optimizing a joint objective function on the labeled and unlabeled subsets of the data , improving both the supervised classification of the labeled data , and the clustering of the unlabeled data . In addition , the proposed method can be easily extended to incremental learning by classifying the pre-defined and novel relations with a unified classifier , which is often desirable in real-world applications . Our experimental results show that our method outperforms current state-of-the-art methods for OpenRE . Our codes are publicly available at Github * . To summarize , the main contributions of our work are as follows : ( 1 ) we propose a novel relation-oriented clustering method RoCORE to enable model to learn to cluster relational data ; ( 2 ) the proposed method achieves the incremental learning of unlabeled novel relations , which is often desirable in real-world applications ; ( 3 ) experimental results show that our method reduces * https://github.com / Ac-Zyx / RoCORE . the error rate by 29.2 % and 15.7 % , on two realworld datasets respectively , compared with current state-of-the-art OpenRE methods . In this work , we introduce a relation-oriented clustering method that extends the current unsupervised clustering-based OpenRE method . The proposed method leverages the labeled data of pre-defined relations to learn a relation-oriented representation from which the derived clusters explicitly align with relational classes . Iterative joint training method effectively reduces the unwanted bias on labeled data . In addition , the proposed method can be easily extended to incremental learning of novel relations . Experimental results show that our method outperforms SOTA methods for OpenRE .

Even though high-dimensional vectors that can encode complex information used for relation extraction are not guaranteed to be consistent with relational semantic similarity.

They propose to use available relation labeled data to obtain relation-oriented representation by minimizing the distance between the same relation instances.

The proposed approach can reduce error rates significantly from the best models for open relation extraction.

P10-1077

Prior use of machine learning in genre classification used a list of labels as classification categories . However , genre classes are often organised into hierarchies , e.g. , covering the subgenres of fiction . In this paper we present a method of using the hierarchy of labels to improve the classification accuracy . As a testbed for this approach we use the Brown Corpus as well as a range of other corpora , including the BNC , HGC and Syracuse . The results are not encouraging : apart from the Brown corpus , the improvements of our structural classifier over the flat one are not statistically significant . We discuss the relation between structural learning performance and the visual and distributional balance of the label hierarchy , suggesting that only balanced hierarchies might profit from structural learning . Automatic genre identification ( AGI ) can be traced to the mid-1990s ( Karlgren and Cutting , 1994 ; Kessler et al . , 1997 ) , but this research became much more active in recent years , partly because of the explosive growth of the Web , and partly because of the importance of making genre distinctions in NLP applications . In Information Retrieval , given the large number of web pages on any given topic , it is often difficult for the users to find relevant pages that are in the right genre ( Vidulin et al . , 2007 ) . As for other applications , the accuracy of many tasks , such as machine translation , POS tagging ( Giesbrecht and Evert , 2009 ) or identification of discourse relations ( Webber , 2009 ) relies of defining the language model suitable for the genre of a given text . For example , the accuracy of POS tagging reaching 96.9 % on newspaper texts drops down to 85.7 % on forums ( Giesbrecht and Evert , 2009 ) , i.e. , every seventh word in forums is tagged incorrectly . This interest in genres resulted in a proliferation of studies on corpus development of web genres and comparison of methods for AGI . The two corpora commonly used for this task are KI-04 ( Meyer zu Eissen and Stein , 2004 ) and Santinis ( Santini , 2007 ) . The best results reported for these corpora ( with 10-fold cross-validation ) reach 84.1 % on KI-04 and 96.5 % accuracy on Santinis ( Kanaris and Stamatatos , 2009 ) . In our research ( Sharoff et al . , 2010 ) we produced even better results on these two benchmarks ( 85.8 % and 97.1 % , respectively ) . However , this impressive accuracy is not realistic in vivo , i.e. , in classifying web pages retrieved as a result of actual queries . One reason comes from the limited number of genres present in these two collections ( eight genres in KI-04 and seven in Santinis ) . As an example , only front pages of online newspapers are listed in Santinis , but not actual newspaper articles , so once an article is retrieved , it can not be assigned to any class at all . Another reason why the high accuracy is not useful concerns the limited number of sources in each collection , e.g. , all FAQs in Santinis come from either a website with FAQs on hurricanes or another one with tax advice . In the end , a classifier built for FAQs on this training data relies on a high topic-genre correlation in this particular collection and fails to spot any other FAQs . There are other corpora , which are more diverse in the range of their genres , such as the fifteen genres of the Brown Corpus ( Kučera and Francis , 1967 ) or the seventy genres of the BNC ( Lee , 2001 ) , but because of the number of genres in them and the diversity of documents within each genre , the accuracy of prior work on these collections is much less impressive . For example , Karlgren and Cutting ( 1994 ) using linear discriminant analysis achieve an accuracy of 52 % without us-ing cross-validation ( the entire Brown Corpus was used as both the test set and training set ) , with the accuracy improving to 65 % when the 15 genres are collapsed into 10 , and to 73 % with only 4 genres ( Figure 1 ) . This result suggests the importance of the hierarchy of genres . Firstly , making a decision on higher levels might be easier than on lower levels ( fiction or non-fiction rather than science fiction or mystery ) . Secondly , we might be able to improve the accuracy on lower levels , by taking into account the relevant position of each node in the hierarchy ( distinguishing between reportage or editorial becomes easier when we know they are safely under the category of press ) . This paper explores a way of using information on the hierarchy of labels for improving fine-grained genre classification . To the best of our knowledge , this is the first work presenting structural genre classification and distance measures for genres . In Section 2 we present a structural reformulation of Support Vector Machines ( SVMs ) that can take similarities between different genres into account . This formulation necessitates the development of distance measures between different genres in a hierarchy , of which we present three different types in Section 3 , along with possible estimation procedures for these distances . We present experiments with these novel structural SVMs and distance measures on three different corpora in Section 4 . Our experiments show that structural SVMs can outperform the non-structural standard . However , the improvement is only statistically significant on the Brown corpus . In Section 5 we investigate potential reasons for this , including the ( im)balance of different genre hierarchies and problems with our distance measures . In this paper , we have evaluated structural learning approaches to genre classification using sev-eral different genre distance measures . Although we were able to improve on non-structural approaches for the Brown corpus , we found it hard to improve over flat SVMs on other corpora . As potential reasons for this negative result , we suggest that current genre hierarchies are either not of sufficient depth or are visually or distributionally imbalanced . We think further investigation into the relationship between hierarchy balance and structural learning is warranted . Further investigation is also needed into the appropriateness of n-gram features for genre identification as well as good measures of genre distance . In the future , an important task would be the refinement or unsupervised generation of new hierarchies , using information theoretic or data-driven approaches . For a full assessment of hierarchical learning for genre classification , the field of genre studies needs a testbed similar to the Reuters or 20 Newsgroups datasets used in topic-based IR with a balanced genre hierarchy and a representative corpus of reliably annotated webpages . With regard to algorithms , we are also interested in other formulations for structural SVMs and their large-scale implementation as well as the combination of different distance measures , for example in ensemble learning .

Existing genre classification methods achieve high accuracy without regarding hierarchical structures exhibiting unrealistic experimental setups such as a limited number of genres and sources.

They propose a structural reformulation of the Support Vector Machine to take hierarchical information of genres into account by using similarities between different genres.

The proposed model outperforms non-hierarchical models on only one corpus and they discuss that it may be due to insufficient depth or inbalance of hierarchies.

2020.acl-main.282

The International Classification of Diseases ( ICD ) provides a standardized way for classifying diseases , which endows each disease with a unique code . ICD coding aims to assign proper ICD codes to a medical record . Since manual coding is very laborious and prone to errors , many methods have been proposed for the automatic ICD coding task . However , most of existing methods independently predict each code , ignoring two important characteristics : Code Hierarchy and Code Co-occurrence . In this paper , we propose a Hyperbolic and Co-graph Representation method ( HyperCore ) to address the above problem . Specifically , we propose a hyperbolic representation method to leverage the code hierarchy . Moreover , we propose a graph convolutional network to utilize the code co-occurrence . Experimental results on two widely used datasets demonstrate that our proposed model outperforms previous state-ofthe-art methods . The International Classification of Diseases ( ICD ) is a healthcare classification system supported by the World Health Organization , which provides a unique code for each disease , symptom , sign and so on . ICD codes have been widely used for analyzing clinical data and monitoring health issues ( Choi et al . , 2016 ; Avati et al . , 2018 ) . Due to the importance of ICD codes , ICD coding -which assigns proper ICD codes to a medical record -has drawn much attention . The task of ICD coding is usually undertaken by professional coders according to doctors ' diagnosis descriptions in the form of free texts . However , manual coding is very expensive , time-consuming and error-prone . The cost incurred by coding errors and the financial investment spent on improving coding quality are estimated to be $ 25 billion per year in the US ( Lang , 2007 ) . Two main reasons can account for this . First , only the people who have medical expert knowledge and specialized ICD coding skills can handle the task . However , it is hard to train such an eligible ICD coder . Second , it is difficult to correctly assign proper codes to the input document even for professional coders , because one document can be assigned multiple ICD codes and the number of codes in the taxonomy of ICD is large . For example , there are over 15,000 and 60,000 codes respectively in the ninth version ( ICD-9 ) and the tenth version ( ICD-10 ) of ICD taxonomies . To reduce human labor and coding errors , many methods have been carefully designed for automatic ICD coding ( Perotte et al . , 2013 ; Mullenbach et al . , 2018 ) . For example in Figure 1 , given the clinical text of a patient , the ICD coding model needs to automatically predict the corresponding ICD codes . The automatic ICD coding task can be modeled as a multi-label classification task since each clinical text is usually accompanied by mul- tiple codes . Most of the previous methods handle each code in isolation and convert the multi-label problem into a set of binary classification problems to predict whether each code of interest presents or not ( Mullenbach et al . , 2018 ; Rios and Kavuluru , 2018 ) . Though effective , they ignore two important characteristics : Code Hierarchy and Code Co-occurrence , which can be leveraged to improve coding accuracy . In the following , we will introduce the two characteristics and the reasons why they are critical for the automatic ICD coding . Code Hierarchy : Based on ICD taxonomy , ICD codes are organized under a tree-like hierarchical structure as shown in Figure 2 , which indicates the parent-child and sibling relations between codes . In the hierarchical structure , the upper level nodes represent more generic disease categories and the lower level nodes represent more specific diseases . The code hierarchy can capture the mutual exclusion of some codes . If code X and Y are both children of Z ( i.e. , X and Y are the siblings ) , it is unlikely to simultaneously assign X and Y to a patient in general ( Xie and Xing , 2018 ) . For example in Figure 2 , if code " 464.00 ( acute laryngitis without mention of obstruction ) " is assigned to a patient , it is unlikely to assign the code " 464.01 ( acute laryngitis with obstruction ) " to the patient at the same time . If automatic ICD coding models ignore such a characteristic , they are prone to giving inconsistent predictions . Thus , a challenging problem is how to model the code hierarchy and use it to capture the mutual exclusion of codes . Code Co-occurrence : Since some diseases are concurrent or have a causal relationship with each other , their codes usually co-occur in the clinical text , such as " 997.91 ( hypertension ) " and " 429.9 ( heart disease ) " . In this paper , we call such characteristic code co-occurrence which can capture the correlations of codes . The code co-occurrence can be utilized to correctly predict some codes which are difficult to predict by only using the clinical text itself . For example in Figure 1 , the code of " acute respiratory failure " can be easily inferred via capturing apparent clues ( i.e. , the green bold words ) from the text . Although there are also a few clues to infer the code of " acidosis " , they are very obscure , let alone predict the code of " acidosis " by only using these obscure clues . Fortunately , there is a strong association between these two diseases : one of the main causes of " acidosis " is " acute respiratory failure " . This prior knowledge can be captured via the fact that the codes of the two diseases usually co-occur in clinical texts . By considering the correlation , the automatic ICD coding model can better exploit obscure clues to predict the code of " acidosis " . Therefore , another problem is how to leverage code co-occurrence for ICD coding . In this paper , we propose a novel method termed as Hyperbolic and Co-graph Representation method ( HyperCore ) to address above problems . Since the tree-likeness properties of the hyperbolic space make it more suitable for representing symbolic data with hierarchical structures than the Euclidean space ( Nickel and Kiela , 2017 ) , we propose a hyperbolic representation learning method to learn the Code Hierarchy . Meanwhile , the graph has been proved effective in modeling data correlation and the graph convolutional network ( GCN ) enables to efficiently learn node representation ( Kipf and Welling , 2016 ) . Thus , we devise a code co-occurrence graph ( co-graph ) for capturing Code Co-occurrence and exploit the GCN to learn the code representation in the co-graph . The contributions of this paper are threefold . Firstly , to our best knowledge , this is the first work to propose a hyperbolic representation method to leverage the code hierarchy for automatic ICD coding . Secondly , this is also the first work to utilize a GCN to exploit code co-occurrence correlation for automatic ICD coding . Thirdly , experiments on two widely used automatic ICD coding datasets show that our proposed model outperforms previous state-of-the-art methods . In this paper , we propose a novel hyperbolic and cograph representation framework for the automatic ICD coding task , which can jointly exploit code hierarchy and code co-occurrence . We exploit the hyperbolic representation learning method to leverage the code hierarchy in the hyperbolic space . Moreover , we use the graph convolutional network to capture the co-occurrence correlation . Experimental results on two widely used datasets indicate that our proposed model outperforms previous state-ofthe-art methods . We believe our method can also be applied to other tasks that need to exploit hierarchical label structure and label co-occurrence , such as fine-grained entity typing and hierarchical multi-label classification .

Existing models that classify texts in medical records into the International Classification of Diseases reduce manual efforts however they ignore Code Hierarchy and Code Co-occurrence.

They propose a hyperbolic representation method to leverage the code hierarchy and a graph convolutional network to utilize the code-occurrence for automatic coding.

The proposed model outperforms state-of-the-art methods on two widely used datasets.

N12-1028

The important mass of textual documents is in perpetual growth and requires strong applications to automatically process information . Automatic titling is an essential task for several applications : ' No Subject ' e-mails titling , text generation , summarization , and so forth . This study presents an original approach consisting in titling journalistic articles by nominalizing . In particular , morphological and semantic processing are employed to obtain a nominalized form which has to respect titles characteristics ( in particular , relevance and catchiness ) . The evaluation of the approach , described in the paper , indicates that titles stemming from this method are informative and/or catchy . A title establishes a link between a reader and a text . It has two main functions . First of all , a title can be informative ( it conveys relevant information about the text content and aim ) , and second , it can be catchy or incentive ( Herrero Cecilia , 2007 ) . A heading is said to be catchy when it succeeds in capturing the reader 's attention on an aspect of the announced event , in a ingenious , metaphoric , enigmatic , or shocking way . From a syntactic point of view , a title can be a word , a phrase , an expression , a sentence , that designates a paper or one of its parts , by giving its subject . Titles are used within applications such as automatic generation of contents , or summarization . So , it is interesting to automate the process that produces relevant titles by extracting them from texts , and supplying other applications with such data , while avoiding any human intervention : Direct applications ( as automatic titling of " no object " e-mails ) are thus possible . The point is that several titles can be relevant for a same text : This constitutes the main difficulty of automatic titling . Some writers prefer informative titles , whereas others prefer catchy ones . Others juggle with both criteria according to the context and the type of the publication . So , evaluation of automatic titling is a complex step requiring a human intervention . Indeed , how can titles relevance be estimated ? How an automatic title can be compared to a human-written ( " real " ) title , knowing that both can have a very different morphosyntactic structure ? Automatic titling is a full process , possessing its own functions . It has to be sharply differentiated from summarization and indexation tasks . Its purpose is to propose title(s ) that have to be short , informative and/or catchy , and keep a coherent syntactic structure . NLP1 methods will be exploited in order to abide by language morphosyntactic and semantic constraints in titling . In this paper , we describe an approach of automatic titling relying on nominalization , i.e. rules transforming a verb phrase into a noun phrase ( e.g. " the president left " is nominalized into " President 's Departure " ) . This study raises two crucial questions : ( 1 ) Determining sentences and phrases containing relevant information ( 2 ) Nominalizing a chosen item and using it as a title . Example : From the following pair of sentences " The disappointing perfor-mance , on Sunday October 9th , of Ségolène Royal , amazed the French citizens . For months , they defended their candidate on the Web . " , containing the relevant information about an article in the French press in 2007 , the idea is to built the following title : " Ségolène Royal : Surprise of the French citizens " . In fact , other titles could apply such as " Ségolène Royal 's Disappointing Performance " or " Surprising the French Citizens " , but notice that both are less informative , since they drop a part of the information . This article is organized as such : The following section briefly positions automatic titling in its research environment and describes previous work ( section 2 ) . The next one describes NOMIT , our approach of automatic titling by nominalization , which consists in three successive steps : Extracting candidate headings from the document ( section 3.1 ) , processing them linguistically ( section 3.2 ) , and last , selecting one among the produced headings , which will play the role of the system heading suggestion ( section 3.3 ) . Finally , the results of NOMIT evaluation are presented and discussed ( section 4 ) . Automatic titling is a complex task because titles must be at once informative , catchy , and syntactically correct . Based on linguistic and semantic treatments , our approach determines titles among which approximately 80 % were evaluated as relevant and more than 60 % were qualified as catchy . Experiment and results discussion have pointed at the following liability : The value of Kappa , the inter-judges agreement coefficient , is very difficult to evaluate , mostly when catchiness is at stake . The main cause is that it depends on personal interests . It is thus necessary to ask the following question : Do we have to consider that a title is definitely catchy when at least one person judges it so ? Otherwise , how many people at least ? This is still an open question and needs to be further investigated . Also , some interesting extensions could be envisaged : The approach presented in this paper uses three rules of transformation based on the presence of an auxiliary followed by a past participle . The addition of new rules would enable a syntactic enrichment of the titles . So , it might be profitable to set up rules taking into account the presence of syntactical patterns ( others than " auxiliary + past participle " ) to allow more texts to be titled by NOMIT . Taking the punctuation of the end of sentences into account might also be a promising track . For example , " did it use an electric detonator ? " would become " Use of an electric detonator ? " . It is an interesting point because the presence of a punctuation at the end of a title ( in particular the exclamation or the interrogation ) constitutes a catchy criterion . Last , NOMIT is a method ( easily reproducible in other languages , English in particular ) that stepped out of preceding attempts in automatic headings generation ( POSTIT , CATIT ) . Exploring syntactic patterns , as it does , means that increasing the amount of linguistic information in the process might lead to a reliable heading method . One of the perspectives can be to track the optimum point between the richness of involved information and processes , and the cost of the method . The incremental methodology followed from POSTIT to NOMIT tends to enhance the belief that parameters ( i.e. length , shape , relevance , etc ... ) for an automatic heading procedure have to be studied and well defined , thus leading to a customized titling process .

Automatically titling documents is a complex task because of its subjectivity and titles must be informative, catchy and syntactically correct.

They propose to approach automatic titling by normalizing a verb phrase selected to be relevant into a noun phrase with morphological and semantic processing.

They show that the proposed normalizing process can produce informative and/or catchy titles but evaluations remain challenging due to its subjectivity.

E14-1026

We present a simple preordering approach for machine translation based on a featurerich logistic regression model to predict whether two children of the same node in the source-side parse tree should be swapped or not . Given the pair-wise children regression scores we conduct an efficient depth-first branch-and-bound search through the space of possible children permutations , avoiding using a cascade of classifiers or limiting the list of possible ordering outcomes . We report experiments in translating English to Japanese and Korean , demonstrating superior performance as ( a ) the number of crossing links drops by more than 10 % absolute with respect to other state-of-the-art preordering approaches , ( b ) BLEU scores improve on 2.2 points over the baseline with lexicalised reordering model , and ( c ) decoding can be carried out 80 times faster . Source-side preordering for translation is the task of rearranging the order of a given source sentence so that it best resembles the order of the target sentence . It is a divide-and-conquer strategy aiming to decouple long-range word movement from the core translation task . The main advantage is that translation becomes computationally cheaper as less word movement needs to be considered , which results in faster and better translations , if preordering is done well and efficiently . Preordering also can facilitate better estimation of alignment and translation models as the parallel data becomes more monotonically-aligned , and translation gains can be obtained for various system architectures , e.g. phrase-based , hierarchical phrase-based , etc . For these reasons , preordering has a clear research and commercial interest , as reflected by the extensive previous work on the subject ( see Section 2 ) . From these approaches , we are particularly interested in those that ( i ) involve little or no human intervention , ( ii ) require limited computational resources at runtime , and ( iii ) make use of available linguistic analysis tools . In this paper we propose a novel preordering approach based on a logistic regression model trained to predict whether to swap nodes in the source-side dependency tree . For each pair of sibling nodes in the tree , the model uses a feature-rich representation that includes lexical cues to make relative reordering predictions between them . Given these predictions , we conduct a depth-first branch-and-bound search through the space of possible permutations of all sibling nodes , using the regression scores to guide the search . This approach has multiple advantages . First , the search for permutations is efficient and does not require specific heuristics or hard limits for nodes with many children . Second , the inclusion of the regression prediction directly into the search allows for finer-grained global decisions as the predictions that the model is more confident about are preferred . Finally , the use of a single regression model to handle any number of child nodes avoids incurring sparsity issues , while allowing the integration of a vast number of features into the preordering model . We empirically contrast our proposed method against another preordering approach based on automatically-extracted rules when translating English into Japanese and Korean . We demonstrate a significant reduction in number of crossing links of more than 10 % absolute , as well as translation gains of over 2.2 BLEU points over the baseline . We also show it outperforms a multi-class classification approach and analyse why this is the case . We have presented a novel preordering approach that estimates a preference for swapping or not swapping pairs of children nodes in the sourceside dependency tree by training a feature-rich logistic regression model . Given the pair-wise scores , we efficiently search through the space of possible children permutations using depth-first branch-and-bound search . The approach is able to incorporate large numbers of features including lexical cues , is efficient at runtime even with a large number of children , and proves superior to other state-of-the-art preordering approaches both in terms of crossing score and translation performance .

Preordering methods for machine translation systems that involve little or no human assistance, run on limited computational resources and use linguistic analysis tools are required.

They propose a logistic regression-based model with lexical features which predicts whether two children of the same node in the parse tree should be swapped.

Experiments on translation tasks from English to Japanese and Korean show the proposed method outperforms baseline preordering methods and runs 80 times faster.

2020.acl-main.443

There is an increasing interest in studying natural language and computer code together , as large corpora of programming texts become readily available on the Internet . For example , StackOverflow currently has over 15 million programming related questions written by 8.5 million users . Meanwhile , there is still a lack of fundamental NLP techniques for identifying code tokens or software-related named entities that appear within natural language sentences . In this paper , we introduce a new named entity recognition ( NER ) corpus for the computer programming domain , consisting of 15,372 sentences annotated with 20 fine-grained entity types . We trained indomain BERT representations ( BERTOverflow ) on 152 million sentences from Stack-Overflow , which lead to an absolute increase of +10 F 1 score over off-the-shelf BERT . We also present the SoftNER model which achieves an overall 79.10 F 1 score for code and named entity recognition on StackOverflow data . Our SoftNER model incorporates a context-independent code token classifier with corpus-level features to improve the BERTbased tagging model . 1 Recently there has been significant interest in modeling human language together with computer code ( Quirk et al . , 2015 ; Iyer et al . , 2016 ; Yin and Neubig , 2018 ) , as more data becomes available on websites such as StackOverflow and GitHub . This is an ambitious yet promising direction for scaling up language understanding to richer domains . Access to domain-specific NLP tools could help a wide range of downstream applications . For example , extracting software knowledge bases from 1 Our code and data are available at : https:// github.com / jeniyat / StackOverflowNER/ text ( Movshovitz-Attias and Cohen , 2015 ) , developing better quality measurements of StackOverflow posts ( Ravi et al . , 2014 ) , finding similar questions ( Amirreza Shirani , 2019 ) and more . However , there is a lack of NLP resources and techniques for identifying software-related named entities ( e.g. , variable names or application names ) within natural language texts . In this paper , we present a comprehensive study that investigates the unique challenges of named entity recognition in the social computer programming domain . These named entities are often ambiguous and have implicit reliance on the accompanied code snippets . For example , the word ' list ' commonly refers to a data structure , but can also be used as a variable name ( Figure 1 ) . In order to recognize these entities , we propose a software-related named entity recognizer ( Soft-NER ) that utilizes an attention network to combine the local sentence-level context with corpuslevel information extracted from the code snippets . Using our newly annotated corpus of 15,372 sentences in StackOverflow , we rigorously test our proposed SoftNER model , which outperforms BiLSTM-CRF model and fine-tuned BERT model for identifying 20 types of software-related named entities . Our key contributions are the following : • A new StackOverflow NER corpus manually annotated with 20 types of named en-tities , including all in-line code within natural language sentences ( § 2 ) . We demonstrate that NER in the software domain is an ideal benchmark task for testing effectiveness of contextual word representations , such as ELMo ( Peters et al . , 2018 ) and BERT ( Devlin et al . , 2019 ) , due to its inherent polysemy and salient reliance on context . • An in-domain trained neural SoftNER tagger for StackOveflow ( § 3 ) that can recognize 20 fine-grained named entity types related to software developing . We also tested its performance on GitHub data of readme files and issue reports . • A code token recognizer ( § 3.1 ) that utilizes StackOveflow code snippets to capture the spelling patterns of code-related tokens , and consistently improves the NER tagger . • In-domain pretrained ELMo and BERT representations ( § 3.3 ) on 152 million sentences from StackOverflow that significantly outperforms off-the-shelf ELMo and leads to more than 21 points increase in F 1 score over offthe-shelf BERT . Overall , our named entity tagger ( SoftNER ) achieves a 79.10 % F 1 score on StackOverflow and 61.08 % F 1 score on GitHub data for extracting the 20 software related named entity types . We believe this performance is sufficiently strong to be practically useful . We have released our data and code , including the named entity tagger , our annotated corpus , annotation guideline , a specially designed tokenizer , and pre-trained StackOverflow BERT and ELMo embeddings . In this work , we investigated the task of named entity recognition in the social computer programming domain . We developed a new NER corpus of 15,372 sentences from StackOverflow and 6,510 sentences from GitHub annotated with 20 fine-grained named entities . We demonstrate that this new corpus is an ideal benchmark dataset for contextual word representations , as there are many challenging ambiguities that often require long-distance context to resolve . We also proposed a novel attention based model , named Soft-NER , that outperforms the state-of-the-art NER models on this dataset . Furthermore , we investigated the important sub-task of code recognition . Our code recognition model captures additional spelling information beyond then contextual word representations and consistently helps to improve the NER performance . We believe our corpus , StackOverflow-specific BERT embeddings and named entity tagger will be useful for various language-and-code tasks , such as code retrieval , software knowledge base extraction and automated question-answering . A Feature-Based CRF Baseline We implemented a CRF baseline model using CRFsuite 8 to extract the software entities . This model uses standard orthographic , contextual and gazetteer features . It also includes the code markdown tags ( § 3.1.3 ) and a set of regular expression features . The regular expressions are developed to recognize specific categories of coderelated entities . Feature ablation experiments on this CRF model are presented in Table 9

Resources and fundamental techniques are missing for identifying software-related named entities such as variable names or application names within natural language texts.

They introduce a manually annotated named entity corpus for the computer programming domain and an attention-based model which incorporates a context-independent code token classifier.

The proposed model outperforms BiLSTM-CRF and fine-tuned BERT models by achieving a 79.10 F1 score for code and named entity recognition on their dataset which

D14-1205

Populating Knowledge Base ( KB ) with new knowledge facts from reliable text resources usually consists of linking name mentions to KB entities and identifying relationship between entity pairs . However , the task often suffers from errors propagating from upstream entity linkers to downstream relation extractors . In this paper , we propose a novel joint inference framework to allow interactions between the two subtasks and find an optimal assignment by addressing the coherence among preliminary local predictions : whether the types of entities meet the expectations of relations explicitly or implicitly , and whether the local predictions are globally compatible . We further measure the confidence of the extracted triples by looking at the details of the complete extraction process . Experiments show that the proposed framework can significantly reduce the error propagations thus obtain more reliable facts , and outperforms competitive baselines with state-of-the-art relation extraction models . Recent advances in natural language processing have made it possible to construct structured KBs from online encyclopedia resources , at an unprecedented scale and much more efficiently than traditional manual edit . However , in those KBs , entities which are popular to the community usually contain more knowledge facts , e.g. , the basketball player LeBron James , the actor Nicholas Cage , etc . , while most other entities often have fewer facts . On the other hand , knowledge facts should be updated as the development of entities , such as changes in the cabinet , a marriage event , or an acquisition between two companies , etc . In order to address the above issues , we could consult populating existing KBs from reliable text resources , e.g. , newswire , which usually involves enriching KBs with new entities and populating KBs with new knowledge facts , in the form of < Entity , Relation , Entity > triple . In this paper , we will focus on the latter , identifying relationship between two existing KB entities . This task can be intuitively considered in a pipeline paradigm , that is , name mentions in the texts are first linked to entities in the KB ( entity linking , EL ) , and then the relationship between them are identified ( relation extraction , RE ) . It is worth mentioning that the first task EL is different from the task of named entity recognition ( NER ) in traditional information extraction ( IE ) tasks , where NER recognizes and classifies the entity mentions ( to several predefined types ) in the texts , but EL focuses on linking the mentions to their corresponding entities in the KB . Such pipeline systems often suffer from errors propagating from upstream to downstream , since only the local best results are selected to the next step . One idea to solve the problem is to allow interactions among the local predictions of both subtasks and jointly select an optimal assignment to eliminate possible errors in the pipeline . Let us first look at an example . Suppose we are extracting knowledge facts from two sentences in Figure 1 : in sentence [ 1 ] , if we are more confident to extract the relation fb : org.headquarters1 , we will be then prompted to select Bryant University , which indeed favors the RE prediction that requires an organization to be its subject . On the other side , if we are sure to link to Kobe Bryant in sentence [ 2 ] , we will probably select fb : pro athlete.teams , whose subject position expects an athlete , e.g. , an NBA player . It is not difficult to see that the argument type expectations of relations can encourage the two subtasks interact with each other and select coherent predictions for both of them . In KBs with well-defined schemas , such as Freebase , type requirements can be collected and utilized explicitly ( Yao et al . , 2010 ) . However , in other KBs with less reliable or even no schemas , it is more appropriate to implicitly capture the type expectations for a given relation ( Riedel et al . , 2013 ) . Furthermore , previous RE approaches usually process each triple individually , which ignores whether those local predictions are compatible with each other . For example , suppose the local predictions of the two sentences above are < Kobe Bryant , fb : org.headquarters , Smithfield , Rhode Is-land > and < Kobe Bryant , fb : pro athlete.teams , Los Angeles Lakers > , respectively , which , in fact , disagree with each other with respect to the KB , since , in most cases , these two relations can not share subjects . Now we can see that either the relation predictions or the EL results for " Bryant " are incorrect . Those disagreements provide us an effective way to remove the possible incorrect predictions that cause the incompatibilities . On the other hand , the automatically extracted knowledge facts inevitably contain errors , especially for those triples collected from open domain . Extractions with confidence scores will be more than useful for users to make proper decisions according to their requirements , such as trading recall for precision , or supporting approximate queries . In this paper , we propose a joint framework to populate an existing KB with new knowledge facts extracted from reliable text resources . The joint framework is designed to address the error propagation issue in a pipeline system , where subtasks are optimized in isolation and locally . We find an optimal configuration from top k results of both subtasks , which maximizes the scores of each step , fulfills the argument type expectations of relations , which can be captured explicitly or implicitly , in the KB , and avoids globally incoherent predictions . We formulate this optimization problem in an Integer Linear Program ( ILP ) framework , and further adopt a logistic regression model to measure the reliability of the whole process , and assign confidences to all extracted triples to facilitate further applications . The experiments on a real-world case study show that our framework can eliminate error propagations in the pipeline systems by taking relations ' argument type expectations and global compatibilities into account , thus outperforms the pipeline approaches based on state-ofthe-art relation extractors by a large margin . Furthermore , we investigate both explicit and implicit type clues for relations , and provide suggestions about which to choose according to the characteristics of existing KBs . Additionally , our proposed confidence estimations can help to achieve a precision of over 85 % for a considerable amount of high quality extractions . In the rest of the paper , we first review related work and then define the knowledge base population task that we will address in this paper . Next we detail the proposed framework and present our experiments and results . Finally , we conclude this paper with future directions . In this paper , we propose a joint framework for the task of populating KBs with new knowledge facts , which performs joint inference on two subtasks , maximizes their preliminary scores , fulfills the type expectations of relations and avoids global incompatibilities with respect to all local predictions to find an optimal assignment . Experimental results show that our framework can significantly eliminate the error propagations in pipeline systems and outperforms competitive pipeline systems with state-of-the-art RE models . Regarding the explicit argument type expectations and the implicit ones , the latter can result in a higher overall precision , while the former performs better in acquiring high quality knowledge facts with higher confidence control , indicating that if the KB has a well-defined schema we can use explicit type requirements for the KBP task , and if not , our model can still perform well by mining the implicit ones . Our framework can also generalize well with other preliminary RE models . Furthermore , we assign extraction confidences to all extracted facts to facilitate further applications . By setting a suitable threshold , our framework can populate high quality reliable knowledge facts to existing KBs . For future work , we will address the NIL issue of EL where we currently assume all entities should be linked to a KB . It would be also interesting to jointly model the two subtasks through structured learning , instead of joint inference only . Currently we only use the coherence of extracted triples and the KB to estimate confidences , which would be nice to directly model the issue in a joint model .

Existing pipeline approaches to populate Knowledge Base with new knowledge facts from texts suffer from error propagating from upstream entity linkers to downstream relation extractors.

They propose to formulate the problem in an Integer Linear Program to find an optimal configuration from the top k results of both tasks.

They show that the proposed framework can reduce error propagations and outperform competitive pipeline baselines with state-of-the-art relation extraction models.

N19-1233

Generative Adversarial Networks ( GANs ) are a promising approach for text generation that , unlike traditional language models ( LM ) , does not suffer from the problem of " exposure bias " . However , A major hurdle for understanding the potential of GANs for text generation is the lack of a clear evaluation metric . In this work , we propose to approximate the distribution of text generated by a GAN , which permits evaluating them with traditional probability-based LM metrics . We apply our approximation procedure on several GAN-based models and show that they currently perform substantially worse than stateof-the-art LMs . Our evaluation procedure promotes better understanding of the relation between GANs and LMs , and can accelerate progress in GAN-based text generation . Neural networks have revolutionized the field of text generation , in machine translation ( Sutskever et al . , 2014 ; Neubig , 2017 ; Luong et al . , 2015 ; Chen et al . , 2018 ) , summarization ( See et al . , 2017 ) , image captioning ( You et al . , 2016 ) and many other applications ( Goldberg , 2017 ) . Traditionally , text generation models are trained by going over a gold sequence of symbols ( characters or words ) from left-to-right , and maximizing the probability of the next symbol given the history , namely , a language modeling ( LM ) objective . A commonly discussed drawback of such LM-based text generation is exposure bias ( Ranzato et al . , 2015 ) : during training , the model predicts the next token conditioned on the ground truth history , while at test time prediction is based on predicted tokens , causing a train-test mismatch . Models trained in this manner often struggle to overcome previous prediction errors . Generative Adversarial Networks ( Goodfellow et al . , 2014 ) offer a solution for exposure bias . * The authors contributed equally Originally introduced for images , GANs leverage a discriminator , which is trained to discriminate between real images and generated images via an adversarial loss . In such a framework , the generator is not directly exposed to the ground truth data , but instead learns to imitate it using global feedback from the discriminator . This has led to several attempts to use GANs for text generation , with a generator using either a recurrent neural network ( RNN ) ( Yu et al . , 2017 ; Guo et al . , 2017 ; Press et al . , 2017 ; Rajeswar et al . , 2017 ) , or a Convolutional Neural Network ( CNN ) ( Gulrajani et al . , 2017 ; Rajeswar et al . , 2017 ) . However , evaluating GANs is more difficult than evaluating LMs . While in language modeling , evaluation is based on the log-probability of a model on held-out text , this can not be straightforwardly extended to GAN-based text generation , because the generator outputs discrete tokens , rather than a probability distribution . Currently , there is no single evaluation metric for GAN-based text generation , and existing metrics that are based on n-gram overlap are known to lack robustness and have low correlation with semantic coherence ( Semeniuta et al . , 2018 ) . In this paper , we propose a method for evaluating GANs with standard probability-based evaluation metrics . We show that the expected prediction of a GAN generator can be viewed as a LM , and suggest a simple Monte-Carlo method for approximating it . The approximated probability distribution can then be evaluated with standard LM metrics such as perplexity or Bits Per Character ( BPC ) . To empirically establish our claim , we implement our evaluation on several RNN-based GANs : ( Press et al . , 2017 ; Yu et al . , 2017 ; Guo et al . , 2017 ) . We find that all models have substantially lower BPC compared to state-of-the-art LMs . By directly comparing to LMs , we put in perspective the current performance of RNN-based GANs for text generation . Our results are also in line with recent concurrent work by Caccia et al . ( 2018 ) , who reached a similar conclusion by comparing the performance of textual GANs to that of LMs using metrics suggested for GAN evaluation . Our code is available at : http : //github.com / GuyTevet / SeqGAN-eval and http://github.com / GuyTevet/ rnn-gan-eval . We propose an evaluation procedure for text GANs that is based on approximating the GAN output distribution and using standard LM metrics . We provide a bound for the number of samples required for the approximation and empirically show in practice as few as 2000 samples per time-step suffice . We evaluate character-based GAN models using our procedure , and show their performance is substantially lower than state-of-the-art LM . We hope our simple evaluation method leads to progress in GAN-based text generation by shedding light on the quality of such models .

Generative Adversarial Networks-based text generation models do not suffer from the exposure bias problem, however; they cannot be evaluated as other language models with log-probability.

They propose a way to approximate distributions from GAN-based models' outputs so that they can be evaluated as standard language models.

When GAN-based models are compared using the same evaluation metric as proposed, they perform much worse than current best language models.

P14-1064

Statistical phrase-based translation learns translation rules from bilingual corpora , and has traditionally only used monolingual evidence to construct features that rescore existing translation candidates . In this work , we present a semi-supervised graph-based approach for generating new translation rules that leverages bilingual and monolingual data . The proposed technique first constructs phrase graphs using both source and target language monolingual corpora . Next , graph propagation identifies translations of phrases that were not observed in the bilingual corpus , assuming that similar phrases have similar translations . We report results on a large Arabic-English system and a medium-sized Urdu-English system . Our proposed approach significantly improves the performance of competitive phrasebased systems , leading to consistent improvements between 1 and 4 BLEU points on standard evaluation sets . Statistical approaches to machine translation ( SMT ) use sentence-aligned , parallel corpora to learn translation rules along with their probabilities . With large amounts of data , phrase-based translation systems ( Koehn et al . , 2003 ; Chiang , 2007 ) achieve state-of-the-art results in many typologically diverse language pairs ( Bojar et al . , 2013 ) . However , the limiting factor in the success of these techniques is parallel data availability . Even in resource-rich languages , learning reliable translations of multiword phrases is a challenge , and an adequate phrasal inventory is crucial ⇤ This work was done while the first author was interning at Microsoft Research for effective translation . This problem is exacerbated in the many language pairs for which parallel resources are either limited or nonexistent . While parallel data is generally scarce , monolingual resources exist in abundance and are being created at accelerating rates . Can we use monolingual data to augment the phrasal translations acquired from parallel data ? The challenge of learning translations from monolingual data is of long standing interest , and has been approached in several ways ( Rapp , 1995 ; Callison-Burch et al . , 2006 ; Haghighi et al . , 2008 ; Ravi and Knight , 2011 ) . Our work introduces a new take on the problem using graphbased semi-supervised learning to acquire translation rules and probabilities by leveraging both monolingual and parallel data resources . On the source side , labeled phrases ( those with known translations ) are extracted from bilingual corpora , and unlabeled phrases are extracted from monolingual corpora ; together they are embedded as nodes in a graph , with the monolingual data determining edge strengths between nodes ( § 2.2 ) . Unlike previous work ( Irvine and Callison-Burch , 2013a ; Razmara et al . , 2013 ) , we use higher order n-grams instead of restricting to unigrams , since our approach goes beyond OOV mitigation and can enrich the entire translation model by using evidence from monolingual text . This enhancement alone results in an improvement of almost 1.4 BLEU points . On the target side , phrases initially consisting of translations from the parallel data are selectively expanded with generated candidates ( § 2.1 ) , and are embedded in a target graph . We then limit the set of translation options for each unlabeled source phrase ( § 2.3 ) , and using a structured graph propagation algorithm , where translation information is propagated from labeled to unlabeled phrases proportional to both source and target phrase similarities , we estimate probability distributions over translations for Figure 1 : Example source and target graphs used in our approach . Labeled phrases on the source side are black ( with their corresponding translations on the target side also black ) ; unlabeled and generated ( § 2.1 ) phrases on the source and target sides respectively are white . Labeled phrases also have conditional probability distributions defined over target phrases , which are extracted from the parallel corpora . the unlabeled source phrases ( § 2.4 ) . The additional phrases are incorporated in the SMT system through a secondary phrase table ( § 2.5 ) . We evaluated the proposed approach on both Arabic-English and Urdu-English under a range of scenarios ( § 3 ) , varying the amount and type of monolingual corpora used , and obtained improvements between 1 and 4 BLEU points , even when using very large language models . In this work , we presented an approach that can expand a translation model extracted from a sentence-aligned , bilingual corpus using a large amount of unstructured , monolingual data in both source and target languages , which leads to improvements of 1.4 and 1.2 BLEU points over strong baselines on evaluation sets , and in some scenarios gains in excess of 4 BLEU points . In the future , we plan to estimate the graph structure through other learned , distributed representations .

The performance of statistical phrase-based translation is limited by the size of the available phrasal inventory both for resource rich and poor language pairs.

They propose a semi-supervised approach that produces new translation rules from monolingual data by phrase graph construction and graph propagation techniques.

Their method significantly improves over existing phrase-based methods on Arabic-English and Urdu-English systems when large language models are used.

N18-1114

We present a new approach to the design of deep networks for natural language processing ( NLP ) , based on the general technique of Tensor Product Representations ( TPRs ) for encoding and processing symbol structures in distributed neural networks . A network architecture -the Tensor Product Generation Network ( TPGN ) -is proposed which is capable in principle of carrying out TPR computation , but which uses unconstrained deep learning to design its internal representations . Instantiated in a model for image-caption generation , TPGN outperforms LSTM baselines when evaluated on the COCO dataset . The TPR-capable structure enables interpretation of internal representations and operations , which prove to contain considerable grammatical content . Our caption-generation model can be interpreted as generating sequences of grammatical categories and retrieving words by their categories from a plan encoded as a distributed representation . In this paper we introduce a new architecture for natural language processing ( NLP ) . On what type of principles can a computational architecture be founded ? It would seem a sound principle to require that the hypothesis space for learning which an architecture provides include network hypotheses that are independently known to be suitable for performing the target task . Our proposed architecture makes available to deep learning network configurations that perform natural language generation by use of Tensor Product Representations ( TPRs ) ( Smolensky and Legendre , 2006 ) . Whether learning will create TPRs is unknown in advance , but what we can say with certainty is that the hypothesis space being searched during learn-ing includes TPRs as one appropriate solution to the problem . TPRs are a general method for generating vector-space embeddings of complex symbol structures . Prior work has proved that TPRs enable powerful symbol processing to be carried out using neural network computation ( Smolensky , 2012 ) . This includes generating parse trees that conform to a grammar ( Cho et al . , 2017 ) , although incorporating such capabilities into deep learning networks such as those developed here remains for future work . The architecture presented here relies on simpler use of TPRs to generate sentences ; grammars are not explicitly encoded here . We test the proposed architecture by applying it to image-caption generation ( on the MS-COCO dataset , ( COCO , 2017 ) ) . The results improve upon a baseline deploying a state-of-the-art LSTM architecture ( Vinyals et al . , 2015 ) , and the TPR foundations of the architecture provide greater interpretability . Section 2 of the paper reviews TPR . Section 3 presents the proposed architecture , the Tensor Product Generation Network ( TPGN ) . Section 4 describes the particular model we study for image captioning , and Section 5 presents the experimental results . Importantly , what the model has learned is interpreted in Section 5.3 . Section 6 discusses the relation of the new model to previous work and Section 7 concludes . Tensor Product Representation ( TPR ) ( Smolensky , 1990 ) is a general technique for constructing vector embeddings of complex symbol structures in such a way that powerful symbolic functions can be computed using hand-designed neural network computation . Integrating TPR with deep learning is a largely open problem for which the work presented here proposes a general approach : design deep architectures that are TPRcapable -TPR computation is within the scope of the capabilities of the architecture in principle . For natural language generation , we proposed such an architecture , the Tensor Product Generation Network ( TPGN ): it embodies the TPR operation of unbinding which is used to extract particular symbols ( e.g. , words ) from complex structures ( e.g. , sentences ) . The architecture can be interpreted as containing a part that encodes a sentence and a part that selects one structural role at a time to extract from the sentence . We applied the approach to image-caption generation , developing a TPGN model that was evaluated on the COCO dataset , on which it outperformed LSTM baselines on a range of standard metrics . Unlike standard LSTMs , however , the TPGN model admits a level of interpretability : we can see which roles are being unbound by the unbinding vectors generated internally within the model . We find such roles contain considerable grammatical information , enabling POS tag prediction for the words they generate and displaying clustering by POS .

While Tensor Product Representation is a powerful model for obtaining vector embeddings for symbol structures, its application with deep learning models is still less investigated.

They propose a newly designed model that is based on Tensor Product Representations for encoding and processing words and sentences.

The Tensor Product Representation-based generative model outperforms LSTM models by evaluating on COCO image-caption dataset and also achieves high interpretability.

N15-1159

This paper describes a simple and principled approach to automatically construct sentiment lexicons using distant supervision . We induce the sentiment association scores for the lexicon items from a model trained on a weakly supervised corpora . Our empirical findings show that features extracted from such a machine-learned lexicon outperform models using manual or other automatically constructed sentiment lexicons . Finally , our system achieves the state-of-the-art in Twitter Sentiment Analysis tasks from Semeval-2013 and ranks 2nd best in Semeval-2014 according to the average rank . One of the early and rather successful models for sentiment analysis ( Pang and Lee , 2004 ; Pang and Lee , 2008 ) relied on manually constructed lexicons that map words to their sentiment , e.g. , positive , negative or neutral . The document-level polarity is then assigned by performing some form of averaging , e.g. , majority voting , of individual word polarities found in the document . These systems show an acceptable level of accuracy , they are easy to build and are highly computationally efficient as the only operation required to assign a polarity label are the word lookups and averaging . However , the information about word polarities in a document are best exploited when using machine learning models to train a sentiment classifier . In fact , most successful sentiment classification systems rely on supervised learning . Interestingly , a simple bag of words model using just unigrams and bigrams with an SVM has shown excellent results ( Wang and Manning , 2012 ) performing on par or beating more complicated models , e.g. , using neural networks ( Socher et al . , 2011 ) . Regarding Twitter sentiment analysis , the top performing system ( Mohammad et al . , 2013 ) from Semeval-2013 Twittter Sentiment Analysis task ( Nakov et al . , 2013 ) follows this recipe by training an SVM on various surface form , sentiment and semantic features . Perhaps , the most valuable finding is that sentiment lexicons appear to be the most useful source of features accounting for over 8 point gains in the F-measure on top of the standard feature sets . Sentiment lexicons are mappings from words to scores capturing the degree of the sentiment expressed by a given word . While several manually constructed lexicons are made available , e.g. , the MPQA ( Wilson et al . , 2005 ) , the Bing and Liu ( Hu and Liu , 2004 ) and NRC Emoticon ( Mohammad and Turney , 2013 ) lexicons , providing high quality word-sentiment associations compiled by humans , still their main drawback is low recall . For example , the largest NRC Emoticon lexicon contains only 14k items , whereas tweets with extremely sparse surface forms are known to form very large vocabularies . Hence , using larger lexicons with better recall has the potential of learning more accurate models . Extracting such lexicons automatically is a challenging and interesting problem ( Lau et al . , 2011 ; Bro and Ehrig , 2013 ; Liu et al . , 2013 ; Tai and Kao , 2013 ; Yang et al . , 2014 ; Huang et al . , 2014 ) . However , different from previous work our goal is not to extract human-interpretable lexicons but to use them as a source of features to improve the classifier accuracy . Following this idea , the authors in ( Mohammad et al . , 2013 ) use features derived from the lexicons to build a state-of-the-art sentiment classifier for Twitter . They construct automatic lexicons using noisy labels automatically inferred from emoticons and hashtags present in the tweets . The wordsentiment association scores are estimated using pointwise mutual information ( PMI ) computed between a word and a tweet label . While the idea to model statistical correlations between the words and tweet labels using PMI or any other metric is rather intuitive , we believe there is a more effective way to exploit noisy labels for estimating the word-sentiment association scores . Our method relies on the idea of distant supervision ( Marchetti-Bowick and Chambers , 2012 ) . We use a large distantly supervised Twitter corpus , which contains noisy opinion labels ( positive or negative ) to learn a supervised polarity classifier . We encode tweets using words and multi-word expressions as features ( which are also entries in our lexicon ) . The weights from the learned model are then used to define which lexicon items to keep , i.e. , items that constitute a good sentiment lexicon . The scores for the lexicon items can be then directly used to encode new tweets or used to derive more advanced features . Using machine learning to induce the scores for the lexicon items has an advantage of learning the scores that are directly optimized for the classification task , where lexicon items with higher discriminative power tend to receive higher weights . To assess the effectiveness of our approach , we reimplemented the state-of-the-art system ranking 1st in Semeval-2013 Twitter Sentiment Analysis challenge and used it as our baseline . We show that adding features from our machine-learned sentiment lexicon yields better results than any of the automatic PMI lexicons used in the baseline and all of them combined together . Our system obtains new state-of-the-art results on the SemEval-2013 message level task with an F-score of 71.32 -a 2 % of absolute improvement over the previous best system in SemEval-2013 . We also evaluate the utility of the ML lexicon on the five test sets from a recent Semeval-2014 task showing significant improvement over a strong baseline . Finally , our system shows high accuracy among the 42 systems participating in the Semeval-2014 challenge ranking 2nd best according to the average rank across all test sets . We demonstrated a simple and principled approach grounded in machine learning to construct sentiment lexicons . We show that using off-the-shelf machine learning tools to automatically extract lexicons greatly outperforms other automatically constructed lexicons that use pointwise mutual information to estimate sentiment scores for the lexicon items . We have shown that combining our machinelearned lexicon with the previous best system yields state-of-the-art results in Semeval-2013 gaining over 2 points in F-score and ranking our system 2nd according to the average rank over the five test sets of Semeval-2014 . Finally , our ML-based lexicon shows excellent results when added on top of the current state-of-the-art NRC system . While our experimental study is focused on Twitter , our method is general enough to be applied to sentiment classification tasks on other domains . In the future , we plan to experiment with constructing ML lexicons from larger Twitter corpora also using hashtags . Recently , deep convolutional neural networks for sentence modelling ( Kalchbrenner et al . , 2014 ; Kim , 2014 ) have shown promising results on several NLP tasks . In particular , ( Tang et al . , 2014 ) showed that learning sentiment-specific word embeddings and using them as features can boost the accuracy of existing sentiment classifiers . In the future work we plan to explore such approaches .

While sentiment lexicons are useful for building accurate sentiment classification systems, existing methods suffer from low recall or interpretability.

They propose to use Twitter's noisy opinion labels as distant supervision to learn a supervised polarity classifier and use it to obtain sentiment lexicons.

Using the obtained lexicon with an existing model achieves the state-of-the-art on the SemEval-13 message level task and outperforms baseline models in several other datasets.

D07-1036

Parallel corpus is an indispensable resource for translation model training in statistical machine translation ( SMT ) . Instead of collecting more and more parallel training corpora , this paper aims to improve SMT performance by exploiting full potential of the existing parallel corpora . Two kinds of methods are proposed : offline data optimization and online model optimization . The offline method adapts the training data by redistributing the weight of each training sentence pairs . The online method adapts the translation model by redistributing the weight of each predefined submodels . Information retrieval model is used for the weighting scheme in both methods . Experimental results show that without using any additional resource , both methods can improve SMT performance significantly . Statistical machine translation relies heavily on the available training data . Typically , the more data is used to estimate the parameters of the translation model , the better it can approximate the true translation probabilities , which will obviously lead to a higher translation performance . However , large corpora are not easily available . The collected corpora are usually from very different areas . For example , the parallel corpora provided by LDC come from quite different domains , such as Hongkong laws , Hangkong Hansards and Hongkong news . This results in the problem that a translation system trained on data from a particular domain(e.g . Hongkong Hansards ) will perform poorly when translating text from a different domain(e.g . news articles ) . Our experiments also show that simply putting all these domain specific corpora together will not always improve translation quality . From another aspect , larger amount of training data also requires larger computational resources . With the increasing of training data , the improvement of translation quality will become smaller and smaller . Therefore , while keeping collecting more and more parallel corpora , it is also important to seek effective ways of making better use of available parallel training data . There are two cases when we train a SMT system . In one case , we know the target test set or target test domain , for example , when building a specific domain SMT system or when participating the NIST MT evaluation1 . In the other case , we are unaware of any information of the testing data . This paper presents two methods to exploit full potential of the available parallel corpora in the two cases . For the first case , we try to optimize the training data offline to make it match the test data better in domain , topic and style , thus improving the translation performance . For the second case , we first divide the training data into several domains and train submodels for each domain . Then , in the translation process , we try to optimize the predefined models according to the online input source sentence . Information retrieval model is used for similar sentences retrieval in both methods . Our preliminary experiments show that both methods can improve SMT performance without using any additional data . The remainder of this paper is organized as follows : Section 2 describes the offline data selection and optimization method . Section 3 describes the online model optimization method . The evaluation and discussion are given in section 4 . Related work is introduced before concluding . This paper presents two new methods to improve statistical machine translation performance by making better use of the available parallel training corpora . The offline data selection method adapts the training corpora to the test domain by retrieving similar sentence pairs and redistributing their weight in the training data . Experimental results show that the selected small subset achieves comparable or even better performance than the baseline system with much less training data . The optimized training data can further improve translation performance without using any additional resource . The online model optimization method adapts the translation model to the online test sentence by redistributing the weight of each predefined submodels . Preliminary results show the effectiveness of the method . Our work also demonstrates that in addition to larger training data , more relevant training data is also important for SMT model training . In future work , we will improve our methods in several aspects . Currently , the similar sentence retrieval model and the weighting schemes are very simple . It might work better by trying other sophisticated similarity measure models or using some optimization algorithms to determine submodel 's weights . Introducing language model optimization into our system might further improve translation performance .

Statistical machine translation systems require corpora limited in domain and size, and a model trained on one domain does not perform well on other domains.

They propose offline and online methods to maximize the potential of available corpora by weighting training samples or submodules using an information retrieval model.

The proposed approaches improve translation quality without additional resources using even less data, further experiments with larger training data show that the methods can scale.

P03-1015

The paper describes two parsing schemes : a shallow approach based on machine learning and a cascaded finite-state parser with a hand-crafted grammar . It discusses several ways to combine them and presents evaluation results for the two individual approaches and their combination . An underspecification scheme for the output of the finite-state parser is introduced and shown to improve performance . In several areas of Natural Language Processing , a combination of different approaches has been found to give the best results . It is especially rewarding to combine deep and shallow systems , where the former guarantees interpretability and high precision and the latter provides robustness and high recall . This paper investigates such a combination consisting of an n-gram based shallow parser and a cascaded finite-state parser 1 with hand-crafted grammar and morphological checking . The respective strengths and weaknesses of these approaches are brought to light in an in-depth evaluation on a treebank of German newspaper texts ( Skut et al . , 1997 ) containing ca . 340,000 tokens in 19,546 sentences . The evaluation format chosen ( dependency tuples ) is used as the common denominator of the systems 1 Although not everyone would agree that finite-state parsers constitute a ' deep ' approach to parsing , they still are knowledge-based , require efforts of grammar-writing , a complex linguistic lexicon , manage without training data , etc . in building a hybrid parser with improved performance . An underspecification scheme allows the finite-state parser partially ambiguous output . It is shown that the other parser can in most cases successfully disambiguate such information . Section 2 discusses the evaluation format adopted ( dependency structures ) , its advantages , but also some of its controversial points . Section 3 formulates a classification problem on the basis of the evaluation format and applies a machine learner to it . Section 4 describes the architecture of the cascaded finite-state parser and its output in a novel underspecification format . Section 5 explores several combination strategies and tests them on several variants of the two base components . Section 6 provides an in-depth evaluation of the component systems and the hybrid parser . Section 7 concludes . The paper has presented two approaches to German parsing ( n-gram based machine learning and cascaded finite-state parsing ) , and evaluated them on the basis of a large amount of data . A new representation format has been introduced that allows underspecification of select types of syntactic ambiguity ( attachment and subcategorization ) even in the absence of a full-fledged chart . Several methods have been discussed for combining the two approaches . It has been shown that while combination with the shallow approach can only marginally improve performance of the cascaded parser if ideal disambiguation is assumed , a quite substantial rise is registered in situations closer to the real world where POS tagging is deficient and resolution of attachment and subcategorization ambiguities less than perfect . In ongoing work , we look at integrating a statistic context-free parser called BitPar , which was written by Helmut Schmid and achieves .816 F-score on NEGRA . Interestingly , the performance goes up to .9474 F-score when BitPar is combined with the FS parser ( upper bound ) and .9443 for the lower bound . So at least for German , combining parsers seems to be a pretty good idea . Thanks are due to Helmut Schmid and Prof. C. Rohrer for discussions , and to the reviewers for their detailed comments .

Combining different methods often achieves the best results especially combinations of shallow and deep can realize both interpretability and good results.

They propose several ways to combine a machine learning-based shallow method and a hand-crafted grammar-based cascaded method for parsers.

Evaluations on a treebank of German newspaper texts show that the proposed method achieves substantial gain when there are ambiguities.

N09-1062

Tree substitution grammars ( TSGs ) are a compelling alternative to context-free grammars for modelling syntax . However , many popular techniques for estimating weighted TSGs ( under the moniker of Data Oriented Parsing ) suffer from the problems of inconsistency and over-fitting . We present a theoretically principled model which solves these problems using a Bayesian non-parametric formulation . Our model learns compact and simple grammars , uncovering latent linguistic structures ( e.g. , verb subcategorisation ) , and in doing so far out-performs a standard PCFG . Many successful models of syntax are based on Probabilistic Context Free Grammars ( PCFGs ) ( e.g. , Collins ( 1999 ) ) . However , directly learning a PCFG from a treebank results in poor parsing performance , due largely to the unrealistic independence assumptions imposed by the context-free assumption . Considerable effort is required to coax good results from a PCFG , in the form of grammar engineering , feature selection and clever smoothing ( Collins , 1999 ; Charniak , 2000 ; Charniak and Johnson , 2005 ; Johnson , 1998 ) . This effort must be repeated when moving to different languages , grammar formalisms or treebanks . We propose that much of this hand-coded knowledge can be obtained automatically as an emergent property of the treebanked data , thereby reducing the need for human input in crafting the grammar . We present a model for automatically learning a Probabilistic Tree Substitution Grammar ( PTSG ) , an extension to the PCFG in which non-terminals can rewrite as entire tree fragments ( elementary trees ) , not just immediate children . These large fragments can be used to encode non-local context , such as head-lexicalisation and verb sub-categorisation . Since no annotated data is available providing TSG derivations we must induce the PTSG productions and their probabilities in an unsupervised way from an ordinary treebank . This is the same problem addressed by Data Oriented Parsing ( DOP , Bod et al . ( 2003 ) ) , a method which uses as productions all subtrees of the training corpus . However , many of the DOP estimation methods have serious shortcomings ( Johnson , 2002 ) , namely inconsistency for DOP1 ( Bod , 2003 ) and overfitting of the maximum likelihood estimate ( Prescher et al . , 2004 ) . In this paper we develop an alternative means of learning a PTSG from a treebanked corpus , with the twin objectives of a ) finding a grammar which accurately models the data and b ) keeping the grammar as simple as possible , with few , compact , elementary trees . This is achieved using a prior to encourage sparsity and simplicity in a Bayesian nonparametric formulation . The framework allows us to perform inference over an infinite space of grammar productions in an elegant and efficient manner . The net result is a grammar which only uses the increased context afforded by the TSG when necessary to model the data , and otherwise uses context-free rules . 1 That is , our model learns to use larger rules when the CFG 's independence assumptions do not hold . This contrasts with DOP , which seeks to use all elementary trees from the training set . While our model is able , in theory , to use all such trees , in practice the data does not justify such a large grammar . Grammars that are only about twice the size of a treebank PCFG provide large gains in accuracy . We obtain additional improvements with grammars that are somewhat larger , but still much smaller than the DOP all-subtrees grammar . The rules in these grammars are intuitive , potentially offering insights into grammatical structure which could be used in , e.g. , the development of syntactic ontologies and guidelines for future treebanking projects . In this work we have presented a non-parametric Bayesian model for inducing tree substitution grammars . By incorporating a structured prior over elementary rules our model is able to reason over the infinite space of all such rules , producing compact and simple grammars . In doing so our model learns local structures for latent linguistic phenomena , such as verb subcategorisation and lexical agreement . Our experimental results show that the induced grammars strongly out-perform standard PCFGs , and are comparable to a state-of-the-art parser on small data samples . While our results on the full treebank are well shy of the best available parsers , we have proposed a number of improvements to the model and the parsing algorithm that could lead to state-of-theart performance in the future .

Although Probabilistic Context Free Grammers-based models are currently successful, they suffer from inconsistency and over-fitting when learning from a treebank.

They propose a Probabilistic Tree Substitution Grammer model with a Bayesian-based algorithm for training to accurately model the data and to keep the grammar simple.

The proposed model learns local structures for latent linguistic phenomena outperforms standard methods and is comparable to state-of-the-art methods on small data.

2020.emnlp-main.505

News headline generation aims to produce a short sentence to attract readers to read the news . One news article often contains multiple keyphrases that are of interest to different users , which can naturally have multiple reasonable headlines . However , most existing methods focus on the single headline generation . In this paper , we propose generating multiple headlines with keyphrases of user interests , whose main idea is to generate multiple keyphrases of interest to users for the news first , and then generate multiple keyphrase-relevant headlines . We propose a multi-source Transformer decoder , which takes three sources as inputs : ( a ) keyphrase , ( b ) keyphrase-filtered article , and ( c ) original article to generate keyphrase-relevant , highquality , and diverse headlines . Furthermore , we propose a simple and effective method to mine the keyphrases of interest in the news article and build a first large-scale keyphraseaware news headline corpus , which contains over 180 K aligned triples of news article , headline , keyphrase . Extensive experimental comparisons on the real-world dataset show that the proposed method achieves state-of-theart results in terms of quality and diversity 1 . News Headline Generation is an under-explored subtask of text summarization ( See et al . , 2017 ; Gehrmann et al . , 2018 ; Zhong et al . , 2019 ) . Unlike text summaries that contain multiple contextrelated sentences to cover the main ideas of a document , news headlines often contain a single short sentence to encourage users to read the news . Since one news article typically contains multiple keyphrases or topics of interest to different users , it is useful to generate multiple headlines covering different keyphrases for the news article . Multiheadline generation aims to generate multiple independent headlines , which allows us to recommend news with different news headlines based on the interests of users . Besides , multi-headline generation can provide multiple hints for human news editors to assist them in writing news headlines . However , most existing methods ( Takase et al . , 2016 ; Ayana et al . , 2016 ; Murao et al . , 2019 ; Colmenares et al . , 2019 ; Zhang et al . , 2018 ) focus on single-headline generation . The headline generation process is treated as an one-to-one mapping ( the input is an article and the output is a headline ) , which trains and tests the models without any additional guiding information or constraints . We argue that this may lead to two problems . Firstly , since it is reasonable to generate multiple headlines for the news , training to generate the single ground-truth might result in a lack of more detailed guidance . Even worse , a single ground-truth without any constraint or guidance is often not enough to measure the quality of the generated headline for model testing . For example , even if a generated headline is considered reasonable by humans , it can get a low score in ROUGE ( Lin , 2004 ) , because it might focus on the keyphrases or aspects that are not consistent with the ground-truth . In this paper , we incorporate the keyphrase information into the headline generation as additional guidance . Unlike one-to-one mapping employed in previous works , we treat the headline generation process as a two-to-one mapping , where the inputs are news articles and keyphrases , and the output is a headline . We propose a keyphrase-aware news multi-headline generation method , which contains two modules : ( a ) Keyphrase Generation Model , which aims to generate multiple keyphrases of interest to users for the news article . ( b ) Keyphrase-Aware Multi-Headline Generation Model , which takes the news article and a keyphrase as input and generates a keyphrase-relevant news headline . For training models , we build a first large-scale news keyphrase-aware headline corpus that contains 180 K aligned triples of news article , headline , keyphrase . As in years past , a lot of the food trends of the year were based on creating perfectly photogenic dishes . An aesthetically pleasing dish , however , does n't mean it will stand the test of time . In fact , it 's not uncommon for food trends to be all the hype one year and die out the next . From broccoli coffee to " bowl food , " here are 10 food trends that you likely wo n't see in 2019 . ... [ 15 sentences with 307 words are abbreviated from here . ] In 2018 , restaurants all over the US decided it was a good idea to place gold foil on everything from ice cream to chicken wings to pizza resulting in an expensive food trend . For example , the Ainsworth in New York City sells $ 1,000 worth of gold covered chicken wings . It seems everyone can agree that this is a food trend that might soon disappear . In this paper , we demonstrate how to enable news headline generation systems to be aware of keyphrases such that the model can generate diverse news headlines in a controlled manner . We also build a first large-scale keyphrase-aware news headline corpus , which is based on mining the keyphrases of users ' interests in news articles with user queries . Moreover , we propose a keyphraseaware news multi-headline generation model that contains a multi-source Transformer decoder with three variants of attention-based fusing mechanisms . Extensive experiments on the real-world dataset show that our approach can generate highquality , keyphrase-relevant , and diverse news headlines , which outperforms many strong baselines .

Existing news headline generation models only focus on generating one output even though news articles often have multiple points.

They propose a multi-source transformer decoder and train it using a new large-scale keyphrase-aware news headline corpus built from a search engine.

Their model outperforms strong baselines on their new real-world keyphrase-aware headline generation dataset.

N16-1103

Universal schema builds a knowledge base ( KB ) of entities and relations by jointly embedding all relation types from input KBs as well as textual patterns observed in raw text . In most previous applications of universal schema , each textual pattern is represented as a single embedding , preventing generalization to unseen patterns . Recent work employs a neural network to capture patterns ' compositional semantics , providing generalization to all possible input text . In response , this paper introduces significant further improvements to the coverage and flexibility of universal schema relation extraction : predictions for entities unseen in training and multilingual transfer learning to domains with no annotation . We evaluate our model through extensive experiments on the English and Spanish TAC KBP benchmark , outperforming the top system from TAC 2013 slot-filling using no handwritten patterns or additional annotation . We also consider a multilingual setting in which English training data entities overlap with the seed KB , but Spanish text does not . Despite having no annotation for Spanish data , we train an accurate predictor , with additional improvements obtained by tying word embeddings across languages . Furthermore , we find that multilingual training improves English relation extraction accuracy . Our approach is thus suited to broad-coverage automated knowledge base construction in a variety of languages and domains . The goal of automatic knowledge base construction ( AKBC ) is to build a structured knowledge base ( KB ) of facts using a noisy corpus of raw text evidence , and perhaps an initial seed KB to be augmented ( Carlson et al . , 2010 ; Suchanek et al . , 2007 ; Bollacker et al . , 2008 ) . AKBC supports downstream reasoning at a high level about extracted entities and their relations , and thus has broad-reaching applications to a variety of domains . One challenge in AKBC is aligning knowledge from a structured KB with a text corpus in order to perform supervised learning through distant supervision . Universal schema ( Riedel et al . , 2013 ) along with its extensions ( Yao et al . , 2013 ; Gardner et al . , 2014 ; Neelakantan et al . , 2015 ; Rocktaschel et al . , 2015 ) , avoids alignment by jointly embedding KB relations , entities , and surface text patterns . This propagates information between KB annotation and corresponding textual evidence . The above applications of universal schema express each text relation as a distinct item to be embedded . This harms its ability to generalize to inputs not precisely seen at training time . Recently , Toutanova et al . ( 2015 ) addressed this issue by embedding text patterns using a deep sentence encoder , which captures the compositional semantics of textual relations and allows for prediction on inputs never seen before . This paper further expands the coverage abilities of universal schema relation extraction by introducing techniques for forming predictions for new entities unseen in training and even for new domains with no associated annotation . In the extreme example of domain adaptation to a completely new language , we may have limited linguistic resources or labeled data such as treebanks , and only rarely a KB with adequate coverage . Our method performs multilingual transfer learning , providing a predictive model for a language with no coverage in an existing KB , by leveraging common representations for shared entities across text corpora . As depicted in Figure 1 , we simply require that one language have an available KB of seed facts . We can further improve our models by tying a small set of word embeddings across languages using only simple knowledge about word-level translations , learning to embed semantically similar textual patterns from different languages into the same latent space . In extensive experiments on the TAC Knowledge Base Population ( KBP ) slot-filling benchmark we outperform the top 2013 system with an F1 score of 40.7 and perform relation extraction in Spanish with no labeled data or direct overlap between the Spanish training corpus and the training KB , demonstrating that our approach is wellsuited for broad-coverage AKBC in low-resource languages and domains . Interestingly , joint training with Spanish improves English accuracy . English Low-resource in KB not in KB Figure 1 : Splitting the entities in a multilingual AKBC training set into parts . We only require that entities in the two corpora overlap . Remarkably , we can train a model for the low-resource language even if entities in the lowresource language do not occur in the KB . By jointly embedding English and Spanish corpora along with a KB , we can train an accurate Spanish relation extraction model using no direct annotation for relations in the Spanish data . This approach has the added benefit of providing significant accuracy improvements for the English model , outperforming the top system on the 2013 TAC KBC slot filling task , without using the hand-coded rules or additional annotations of alternative systems . By using deep sentence encoders , we can perform prediction for arbitrary input text and for entities unseen in training . Sentence encoders also provides opportunities to improve cross-lingual transfer learning by sharing word embeddings across languages . In future work we will apply this model to many more languages and domains besides newswire text . We would also like to avoid the entity detection problem by using a deep architecture to both identify entity mentions and identify relations between them .

Existing approaches to incorporate universal schemas for automatic knowledge base construction has limitation in generalization to unseen inputs from training time.

They propose to combine universal schemas and neural network-based deep encoders to achieve generalization to an unseen language without additional annotations.

The proposed approach outperforms existing methods on benchmarks in English and Spanish while having no hand-coded rules or training data for Spanish.

E17-1022

We propose UDP , the first training-free parser for Universal Dependencies ( UD ) . Our algorithm is based on PageRank and a small set of head attachment rules . It features two-step decoding to guarantee that function words are attached as leaf nodes . The parser requires no training , and it is competitive with a delexicalized transfer system . UDP offers a linguistically sound unsupervised alternative to cross-lingual parsing for UD , which can be used as a baseline for such systems . The parser has very few parameters and is distinctly robust to domain change across languages . Grammar induction and unsupervised dependency parsing are active fields of research in natural language processing ( Klein and Manning , 2004 ; Gelling et al . , 2012 ) . However , many data-driven approaches struggle with learning relations that match the conventions of the test data , e.g. , Klein and Manning reported the tendency of their DMV parser to make determiners the heads of German nouns , which would not be an error if the test data used a DP analysis ( Abney , 1987 ) . Even supervised transfer approaches ( McDonald et al . , 2011 ) suffer from target adaptation problems when facing word order differences . The Universal Dependencies ( UD ) project ( Nivre et al . , 2015 ; Nivre et al . , 2016 ) offers a dependency formalism that aims at providing a consistent representation across languages , while enforcing a few hard constraints . The arrival of such treebanks , expanded and improved on a regular basis , provides a new milestone for crosslingual dependency parsing research ( McDonald et al . , 2013 ) . Furthermore , given that UD rests on a series of simple principles like the primacy of lexical heads , cf . Johannsen et al . ( 2015 ) for more details , we expect that such a formalism lends itself more naturally to a simple and linguistically sound rulebased approach to cross-lingual parsing . In this paper we present such an approach . Our system is a dependency parser that requires no training , and relies solely on explicit part-ofspeech ( POS ) constraints that UD imposes . In particular , UD prescribes that trees are single-rooted , and that function words like adpositions , auxiliaries , and determiners are always dependents of content words , while other formalisms might treat them as heads ( De Marneffe et al . , 2014 ) . We ascribe our work to the viewpoints of Bender ( 2009 ) about the incorporation of linguistic knowledge in language-independent systems . We have presented UDP , an unsupervised dependency parser for Universal Dependencies ( UD ) that makes use of personalized PageRank and a small set of head-dependent rules . The parser requires no training data and estimates adposition direction directly from the input . We achieve competitive performance on all but two UD languages , and even beat a multi-source delexicalized parser ( MSD ) on Hindi . We evaluated the parser on three POS setups and across domains . Our results show that UDP is less affected by deteriorating POS tags than MSD , and is more resilient to domain changes . Given how much of the overall dependency structure can be explained by this fairly system , we propose UDP as an additional UD parsing baseline . The parser , the in-house annotated test sets , and the domain data splits are made freely available . 6UD is a running project , and the guidelines are bound to evolve overtime . Indeed , the UD 2.0 guidelines have been recently released . UDP can be augmented with edge labeling for some deterministic labels like case or det . Some further constrains can be incorporated in UDP . Moreover , the parser makes no special treatment of multiword expression that would require a lexicon , coordinations or proper names . All these three kinds of structures have a flat tree where all words depend on the leftmost one . While coordination attachment is a classical problem in parsing and out of the scope of our work , a proper name sequence can be straightforwardly identified from the partof-speech tags , and it falls thus in the area of structures predictable using simple heuristics . Moreover , our use of PageRank could be expanded to directly score the potential dependency edges instead of words , e.g. , by means of edge reification .

For dependency parsing, unsupervised methods struggle with learning relations that match conventions of the test data and supervised counterparts suffer from word order target adaptation.

They propose an unsupervised approach based on PageRank and a set of head attachment rules that solely depend on explicit part-of-speech constraints from Universal Dependencies.

The proposed linguistically sound method performs competitively with a delexicalized transfer system while having few parameters and robustness to domain changes across languages.

P19-1081

We study a conversational reasoning model that strategically traverses through a largescale common fact knowledge graph ( KG ) to introduce engaging and contextually diverse entities and attributes . For this study , we collect a new Open-ended Dialog ↔ KG parallel corpus called OpenDialKG , where each utterance from 15 K human-to-human roleplaying dialogs is manually annotated with ground-truth reference to corresponding entities and paths from a large-scale KG with 1M+ facts . We then propose the DialKG Walker model that learns the symbolic transitions of dialog contexts as structured traversals over KG , and predicts natural entities to introduce given previous dialog contexts via a novel domain-agnostic , attention-based graph path decoder . Automatic and human evaluations show that our model can retrieve more natural and human-like responses than the state-ofthe-art baselines or rule-based models , in both in-domain and cross-domain tasks . The proposed model also generates a KG walk path for each entity retrieved , providing a natural way to explain conversational reasoning . The key element of an open-ended dialog system is its ability to understand conversational contexts and to respond naturally by introducing relevant entities and attributes , which often leads to increased engagement and coherent interactions ( Chen et al . , 2018 ) . While a large-scale knowledge graph ( KG ) includes vast knowledge of all the related entities connected via one or more factual connections from conversational contexts , the core challenge is in the domain-agnostic and scalable prediction of a small subset from those reachable entities that follows natural conceptual threads that can keep conversations engaging and meaningful . Hence , we study a data-driven reasoning model topical jumps across open-ended multi-turn dialogs are annotated and grounded with a large-scale commonfact KG . To generate a KG entity response at each dialog turn , the model learns walkable paths within KG that lead to engaging and natural topics or entities given dialog context , while pruning non-ideal ( albeit factually correct ) KG paths among 1M+ candidate facts . that map dialog transitions with KG paths , aimed at identifying a subset of ideal entities to mention as a response to previous dialog contexts . Figure 1 illustrates a motivating dialog example between two conversation participants , which spans multiple related KG entities from a starting seed entity The Catcher in the Rye . Specifically , we observe that there exists a small subset of walkable patterns within a KG or a preferred sequence of graph traversal steps which often leads to more engaging entities or attributes than others ( e.g. Literacy Realism , Nathaniel Hawthorne , etc . vs. Catch Me If You Can , 277 , etc . all connected via one-or multi-hop factual connections ) . Note also that the walkable degree of each entity varies by dialog contexts and domains , thus making conventional rule-based or entity-toentity learning approaches intractable or not scalable for open-ended dialogs with 1M+ candidate facts . Therefore , pruning the search space for entities based on dialog contexts and their relationbased walk paths is a crucial step in operating knowledge-augmented dialog systems at scale . To this end , we propose a new model called DialKG Walker that can learn natural knowledge paths among entities mentioned over dialog contexts , and reason grounded on a large commonsense KG . Specifically , we propose a novel graph decoder that attends on viable KG paths to predict the most relevant entities from a KG , by associating these paths with the given input contexts : dialog , sentence , and a set of starting KG entities mentioned in the previous turn . We then build a parallel zeroshot learning model that predicts entities in the KG embeddings space , and ranks candidate entities based on decoded graph path output . To train the DialKG Walker model with groundtruth reference to KG entities , we collect a new human-to-human multi-turn dialogs dataset ( 91 K utterances across 15 K dialog sessions ) using Par-lAI ( Miller et al . , 2017 ) , where conversation participants play a role either as a user or as an assistant , while annotating their mention of an entity in a large-scale common fact KG . This new dataset provides a new way for researchers to study how conversational topics could jump across many different entities within multi-turn dialogs , grounded on KG paths that thread all of them . To the best of our knowledge , our OpenDialKG is the first parallel Dialog ↔ KG corpus where each mention of a KG entity and its factual connection in an openended dialog is fully annotated , allowing for indepth study of symbolic reasoning and natural language conversations . Note that our approaches are distinct from the previous work on dialog systems in that we completely ground dialogs in a large-scale commonfact KG , allowing for domain-agnostic conversational reasoning in open-ended conversations across various domains and tasks ( e.g. chit-chat , recommendations , etc . ) We therefore perform extensive cross-domain and transfer learning evaluations to demonstrate its flexibility . See Section 5 for the detailed literature review . Our contributions are as follows : we propose ( 1 ) a novel attention-based graph decoder that walks an optimal path within a large commonsense KG ( 100 K entities , 1.1 M facts ) to effectively prune unlikely candidate entities , and ( 2 ) a zeroshot learning model that leverages previous sentence , dialog , and KG contexts to re-rank candidates from pruned decoder graph output based on their relevance and path scores , which allows for generalizable and robust classification with a large number of candidate classes . We present ( 3 ) a new parallel open-ended dialog ↔ KG corpus called OpenDialKG where each mention of an entity in dialog is manually linked with its corresponding ground-truth KG path . We show that the proposed approaches outperform baselines in both indomain and cross-domain evaluation , demonstrating that the model learns domain-agnostic walking patterns that are generalizable for unseen domains . We study conversational reasoning grounded on knowledge graphs , and formulate an approach in which the model learns to navigate a largescale , open-ended KG given conversational contexts . For this study , we collect a newly annotated Dialog ↔ KG parallel corpus of 15 K humanto-human dialogs which includes ground-truth annotation of each dialog turn to its reasoning reference in a large-scale common fact KG . Our proposed DialKG Walker model improves upon the state-of-the-art knowledge-augmented conversation models by 1 ) a novel attention-based graph decoder that penalizes decoding of unnatural paths which effectively prunes candidate entities and paths from a large search space ( 1.1 M facts ) , 2 ) a zeroshot learning model that predicts its relevance score in the KG embeddings space , combined score of which is used for candidate ranking . The empirical results from in-domain , cross-domain , and transfer learning evaluation demonstrate the efficacy of the proposed model in domain-agnostic conversational reasoning .

Using a large knowledge base for dialogue systems is intractable or not scalable which calls for methods that prune search space for entities.

They provide an open-ended dialogue corpus where each utterance is annotated with entities and paths and propose a model that works on this data structure.

The proposed model can produce more natural responses than state-of-the-art models on automatic and human evaluation, and generated knowledge graph paths provide explainability.

D12-1011

Existing techniques for disambiguating named entities in text mostly focus on Wikipedia as a target catalog of entities . Yet for many types of entities , such as restaurants and cult movies , relational databases exist that contain far more extensive information than Wikipedia . This paper introduces a new task , called Open-Database Named-Entity Disambiguation ( Open-DB NED ) , in which a system must be able to resolve named entities to symbols in an arbitrary database , without requiring labeled data for each new database . We introduce two techniques for Open-DB NED , one based on distant supervision and the other based on domain adaptation . In experiments on two domains , one with poor coverage by Wikipedia and the other with near-perfect coverage , our Open-DB NED strategies outperform a state-of-the-art Wikipedia NED system by over 25 % in accuracy . Named-entity disambiguation ( NED ) is the task of linking names mentioned in text with an established catalog of entities ( Bunescu and Pasca , 2006 ; Ratinov et al . , 2011 ) . It is a vital first step for semantic understanding of text , such as in grounded semantic parsing ( Kwiatkowski et al . , 2011 ) , as well as for information retrieval tasks like person name search ( Chen and Martin , 2007 ; Mann and Yarowsky , 2003 ) . NED requires a catalog of symbols , called referents , to which named-entities will be resolved . Most NED systems today use Wikipedia as the catalog of referents , but exclusive focus on Wikipedia as a target for NED systems has significant drawbacks : despite its breadth , Wikipedia still does not contain all or even most real-world entities mentioned in text . As one example , it has poor coverage of entities that are mostly important in a small geographical region , such as hotels and restaurants , which are widely discussed on the Web . 57 % of the named-entities in the Text Analysis Conference 's ( TAC ) 2009 entity linking task refer to an entity that does not appear in Wikipedia ( McNamee et al . , 2009 ) . Wikipedia is clearly a highly valuable resource , but it should not be thought of as the only one . Instead of relying solely on Wikipedia , we propose a novel approach to NED , which we refer to as Open-DB NED : the task is to resolve an entity to Wikipedia or to any relational database that meets mild conditions about the format of the data , described below . Leveraging structured , relational data should allow systems to achieve strong accuracy , as with domain-specific or database-specific NED techniques like Hoffart et al . 's NED system for YAGO ( Hoffart et al . , 2011 ) . And because of the availability of huge numbers of databases on the Web , many for specialized domains , a successful system for this task will cover entities that a Wikipedia NED or database-specific system can not . We investigate two complementary learning strategies for Open-DB NED , both of which significantly relax the assumptions of traditional NED systems . The first strategy , a distant supervision approach , uses the relational information in a given database and a large corpus of unlabeled text to learn a database-specific model . The second strat-egy , a domain adaptation approach , assumes a single source database that has accompanying labeled data . Classifiers in this setting must learn a model that transfers from the source database to any new database , without requiring new training data for the new database . Experiments show that both strategies outperform a state-of-the-art Wikipedia NED system by wide margins without requiring any labeled data from the test domain , highlighting the significant advantage of having domain-specific relational data . The next section contrasts Open-DB NED with previous work . Section 3 formalizes the task . Sections 4 and 5 present our distant supervision strategy and domain-adaptation strategy , respectively . Section 6 introduces a technique that is a hybrid of the two learning strategies . Section 7 describes our experiments , and Section 8 concludes . This paper introduces the task of Open-DB Named Entity Disambiguation , and presents two distinct strategies for solving this task . Experiments indicate that a mixture of the two strategies significantly outperforms a state-of-the-art Wikipedia NED system , on a dataset where Wikipedia has good coverage and on another dataset where Wikipedia has poor coverage . The results indicate that there is a significant benefit to leveraging other sources of knowledge in addition to Wikipedia , and that it is possible to leverage this knowledge without requiring labeled data for each new source . The initial success of these Open-DB NED approaches indicates that this task is a promising area for future research , including exciting extensions that link large numbers of domainspecific databases to text .

Existing approaches to disambiguate named entities solely use Wikipedia as a catalogue however Many kinds of named entities are missed in Wikipedia.

They propose a task where systems need to reference arbitrary databases for finding named entities not only Wikipedia, together with methods to achieve domain adaptation.

A mixture of two domain adaptation methods outperforms existing systems that only rely on Wikipedia for their new Open-DB Named Entity Disambiguation task.

2020.emnlp-main.308

Solving algebraic word problems has recently emerged as an important natural language processing task . To solve algebraic word problems , recent studies suggested neural models that generate solution equations by using ' Op ( operator / operand ) ' tokens as a unit of input / output . However , such a neural model suffered two issues : expression fragmentation and operand-context separation . To address each of these two issues , we propose a pure neural model , Expression-Pointer Transformer ( EPT ) , which uses ( 1 ) ' Expression ' token and ( 2 ) operand-context pointers when generating solution equations . The performance of the EPT model is tested on three datasets : ALG514 , DRAW-1 K , and MAWPS . Compared to the state-of-the-art ( SoTA ) models , the EPT model achieved a comparable performance accuracy in each of the three datasets ; 81.3 % on ALG514 , 59.5 % on DRAW-1 K , and 84.5 % on MAWPS . The contribution of this paper is two-fold ; ( 1 ) We propose a pure neural model , EPT , which can address the expression fragmentation and the operandcontext separation . ( 2 ) The fully automatic EPT model , which does not use hand-crafted features , yields comparable performance to existing models using hand-crafted features , and achieves better performance than existing pure neural models by at most 40 % . Solving algebraic word problems has recently become an important research task in that automatically generating solution equations requires understanding natural language . Table 1 shows a sample algebraic word problem , along with corresponding solution equations that are used to generate answers for the problem . To solve such problems with deep learning technology , researchers recently suggested neural models that generate solution equations automatically ( Huang Problem One number is eight more than twice another and their sum is 20 . What are their numbers ? Numbers 1('one ' ) , 8('eight ' ) , 2('twice ' ) , 20 . Equations x 0 -2x 1 = 8 , x 0 + x 1 = 20 Answers ( 16 , 4 ) et al . , 2018 ; Amini et al . , 2019 ; Chiang and Chen , 2019 ; Wang et al . , 2019 ) . However , suggested neural models showed a fairly large performance gap compared to existing state-of-the-art models based on hand-crafted features in popular algebraic word problem datasets , such as ALG514 ( 44.5 % for pure neural model vs. 83.0 % for using hand-crafted features ) ( Huang et al . , 2018 ; Upadhyay and Chang , 2016 ) . To address the large performance gap in this study , we propose a larger unit of input / output ( I / O ) token called " Expressions " for a pure neural model . Figure 1 illustrates conventionally used " Op ( operator / operands ) " versus our newly proposed " Expression " token . To improve the performance of pure neural models that can solve algebraic word problems , we identified two issues that can be addressed using Expression tokens , which are shown in Figure 1 : ( 1 ) expression fragmentation and ( 2 ) operandcontext separation . First , the expression fragmentation issue is a segmentation of an expression tree , which represents a computational structure of equations that are used to generate a solution . This issue arises when Op , rather than the whole expression tree , is used as an input / output unit of a problem-solving model . For example , as shown in Figure 1 ( a ) , using Op tokens as an input to a problem-solving model disassembles a tree structure into operators ( " × " ) and operands ( " x 1 " and " 2 " ) . Meanwhile , we propose using the Figure 1 : Illustration using the word problem in Table 1 for the ( a ) expression fragmentation issue , ( b ) operandcontext separation issue , and ( c ) our solution for these two issues . " Expression " ( ×(x 1 , 2 ) ) token , which can explicitly capture a tree structure as a whole , as shown in Figure 1 ( c ) . The second issue of operand-context separation is the disconnection between an operand and a number that is associated with the operand . This issue arises when a problem-solving model substitutes a number stated in an algebraic word problem into an abstract symbol for generalization . As shown in Figure 1 ( b ) , when using an Op token , the number 8 is changed into an abstract symbol ' N 1 ' . Meanwhile , when using an Expression token , the number 8 is not transformed into a symbol . Rather a pointer is made to the location where the number 8 occurred in an algebraic word problem . Therefore , using such an " operand-context pointer " enables a model to access contextual information about the number directly , as shown in Figure 1 ( c ) ; thus , the operand-context separation issue can be addressed . In this paper , we propose a pure neural model called Expression-Pointer Transformer ( EPT ) to address the two issues above . The contribution of this paper is two-fold ; 1 . We propose a pure neural model , Expression-Pointer Transformer ( EPT ) , which can address the expression fragmentation and operandcontext separation issues . 2 . The EPT model is the first pure neural model that showed comparable accuracy to the existing state-of-the-art models , which used handcrafted features . Compared to the state-ofthe-art pure neural models , the EPT achieves better performance by about 40 % . In the rest of the paper , we introduce existing approaches to solve algebraic word problems in Section 2 . Next , Section 3 introduces our proposed model , EPT , and Section 4 reports the experimental settings . Then in Section 5 , results of two studies are presented . Section 5.1 presents a performance comparison between EPT and existing SoTA models . Section 5.2 presents an ablation study examining the effects of Expression tokens and applying operand-context pointers . Finally , in Section 6 , a conclusion is presented with possible future directions for our work . In this study , we proposed a neural algebraic word problem solver , Expression-Pointer Transformer ( EPT ) , and examined its characteristics . We designed EPT to address two issues : expression fragmentation and operand-context separation . The EPT resolves the expression fragmentation issue by generating ' Expression ' tokens , which simultaneously generate an operator and required operands . In addition , the EPT resolves the operand-context separation issue by applying operand-context pointers . Our work is meaningful in that we demonstrated a possibility for alleviating the costly procedure of devising hand-crafted features in the domain of solving algebraic word problems . As future work , we plan to generalize the EPT to other datasets , including non-English word problems or non-algebraic domains in math , to extend our model .

Neural models largely underperform hand-crafted feature-based models on algebraic word datasets such as ALG514 because of two issues namely expression fragmentation and operand-context separation.

They propose a model which generates an operator and required operands and applies operand-context pointers to resolve the expression fragmentation and operand-context separation issues respectively.

The proposed model performs comparable results to the state-of-the-art models with hand-crafted features and outperforms neural models by 40% on three datasets.

P98-1104

In this paper I will report the result of a quantitative analysis of the dynamics of the constituent elements of Japanese terminology . In Japanese technical terms , the linguistic contribution of morphemes greatly differ according to their types of origin . To analyse this aspect , a quantitative method is applied , which can properly characterise the dynamic nature of morphemes in terminology on the basis of a small sample . In computational linguistics , the interest in terminological applications such as automatic term extraction is growing , and many studies use the quantitative information ( cf . Kageura & Umino , 1996 ) . However , the basic quantitative nature of terminological structure , which is essential for terminological theory and applications , has not yet been exploited . The static quantitative descriptions are not sufficient , as there are terms which do not appear in the sample . So it is crucial to establish some models , by which the terminological structure beyond the sample size can be properly described . In Japanese terminology , the roles of morphemes are different according to their types of origin , i.e. the morphemes borrowed mainly from Western languages ( borrowed morphemes ) and the native morphemes including Chineseorigined morphemes which are the majority . There are some quantitative studies ( Ishii , 1987 ; Nomura & Ishii , 1989 ) , but they only treat the static nature of the sample . Located in the intersection of these two backgrounds , the aim of the present study is twofold , i.e. ( 1 ) to introduce a quantitative framework in which the dynamic nature of terminology can be described , and to examine its theoretical validity , and ( 2 ) to describe the quantitative dynamics of morphemes as a ' mass ' in Japanese terminology , with reference to the types of origin . Using the LNRE models , we have succesfully analysed the dynamic nature of the morphemes in Japanese terminology . As the majority of the terminological data is located in the LNRE zone , it is important to use the statistical framework which allows for the LNRE characteristics . The LNRE models give the suitable means . We are currently extending our research to integrating the quantitative nature of morphological distributions to the qualitative mode ] of term formation , by taking into account the po-sitional and combinatorial nature of morphemes and the distributions of term length .

Static quantitative descriptions are not sufficient to analyse Japanese terminology because of the dynamic nature of samples calling for a method beyond the sample size.

They apply a quantitative method which can characterise the dynamic nature of morphemes using a small sample of Japanese terminology.

They show that the method can successfully analyze the dynamic nature of the morphemes in Japanese terminology with suitable means.

2021.acl-long.420

We study the problem of leveraging the syntactic structure of text to enhance pre-trained models such as BERT and RoBERTa . Existing methods utilize syntax of text either in the pre-training stage or in the fine-tuning stage , so that they suffer from discrepancy between the two stages . Such a problem would lead to the necessity of having human-annotated syntactic information , which limits the application of existing methods to broader scenarios . To address this , we present a model that utilizes the syntax of text in both pre-training and fine-tuning stages . Our model is based on Transformer with a syntax-aware attention layer that considers the dependency tree of the text . We further introduce a new pre-training task of predicting the syntactic distance among tokens in the dependency tree . We evaluate the model on three downstream tasks , including relation classification , entity typing , and question answering . Results show that our model achieves state-of-the-art performance on six public benchmark datasets . We have two major findings . First , we demonstrate that infusing automatically produced syntax of text improves pre-trained models . Second , global syntactic distances among tokens bring larger performance gains compared to local head relations between contiguous tokens . 1 * Work is done during internship at Microsoft . † For questions , please contact D. Tang and Z. Xu . Pre-trained models such as BERT ( Devlin et al . , 2019 ) , GPT ( Radford et al . , 2018 ) , and RoBERTa ( Liu et al . , 2019 ) have advanced the state-of-the-art performances of various natural language processing tasks . The successful recipe is that a model is first pre-trained on a huge volume of unsupervised data with self-supervised objectives , and then is fine-tuned on supervised data with the same data scheme . Dominant pre-trained models represent a text as a sequence of tokens2 . The merits are that such basic text representations are available from vast amounts of unsupervised data , and that models pre-trained and fine-tuned with the same paradigm usually achieve good accuracy in practice ( Guu et al . , 2020 ) . However , an evident limitation of these methods is that richer syntactic structure of text is ignored . In this paper , we seek to enhance pre-trained models with syntax of text . Related studies attempt to inject syntax information either only in the finetuning stage ( Nguyen et al . , 2020 ; Sachan et al . , 2020 ) , or only in the pre-training stage ( Wang et al . , 2020 ) , which results in discrepancies . When only fusing syntax information in the fine-tuning phase , Sachan et al . ( 2020 ) finds that there is no performance boost unless high quality human-annotated dependency parses are available . However , this requirement would limit the application of the model to broader scenarios where human-annotated dependency information is not available . To address this , we conduct a large-scale study on injecting automatically produced syntax of text in both the pre-training and fine-tuning stages . We construct a pre-training dataset by applying an offthe-shelf dependency parser ( Qi et al . , 2020 ) to one billion sentences from common crawl news . With these data , we introduce a syntax-aware pretraining task , called dependency distance prediction , which predicts the syntactic distance between tokens in the dependency structure . Compared with the pre-training task of dependency head prediction ( Wang et al . , 2020 ) that only captures local syntactic relations among words , dependency distance prediction leverages global syntax of the text . In addition , we developed a syntax-aware attention layer , which can be conveniently integrated into Transformer ( Vaswani et al . , 2017 ) to allow tokens to selectively attend to contextual tokens based on their syntactic distance in the dependency structure . We conduct experiments on entity typing , question answering and relation classification on six benchmark datasets . Experimental results show that our method achieves state-of-the-art performance on all six datasets . Further analysis shows that our model can indicate the importance of syntactic information on downstream tasks , and that the newly introduced dependency distance prediction task could capture the global syntax of the text , performs better than dependency head prediction . In addition , compared with experimental results of injecting syntax information in either the pre-training or fine-tuning stage , injecting syntax information in both stages achieves the best performance . In summary , the contribution of this paper is threefold . ( 1 ) We demonstrate that infusing automatically produced dependency structures into the pre-trained model shows superior performance over downstream tasks . ( 2 ) We propose a syntax-aware attention layer and a pre-training task for infusing syntactic information into the pre-trained model . ( 3 ) We find that the newly introduced dependency distance prediction task performs better than the dependency head prediction task . In this paper , we present SEPREM that leverage syntax information to enhance pre-trained models . To inject syntactic information , we introduce a syntax-aware attention layer and a newly designed pre-training task are proposed . Experimental results show that our method achieves state-of-theart performance over six datasets . Further analysis shows that the proposed dependency distance prediction task performs better than dependency head prediction task .

Existing ways of injecting syntactic knowledge into pretraining models cause discrepancies between pretraining and fine-tuning and require expensive annotation.

They propose to inject syntactic features obtained by an off-the-shelf parser into pretraining models coupled with a new syntax-aware attention layer.

The proposed model achieves state-of-the-art in relation classification, entity typing, and question answering tasks.

P16-1067

This paper proposes an unsupervised approach for segmenting a multiauthor document into authorial components . The key novelty is that we utilize the sequential patterns hidden among document elements when determining their authorships . For this purpose , we adopt Hidden Markov Model ( HMM ) and construct a sequential probabilistic model to capture the dependencies of sequential sentences and their authorships . An unsupervised learning method is developed to initialize the HMM parameters . Experimental results on benchmark datasets have demonstrated the significant benefit of our idea and our approach has outperformed the state-of-the-arts on all tests . As an example of its applications , the proposed approach is applied for attributing authorship of a document and has also shown promising results . Authorship analysis is a process of inspecting documents in order to extract authorial information about these documents . It is considered as a general concept that embraces several types of authorship subjects , including authorship verif ication , plagiarism detection and author attribution . Authorship verification ( Brocardo et al . , 2013 ; Potha and Stamatatos , 2014 ) decides whether a given document is written by a specific author . Plagiarism detection ( Stein et al . , 2011 ; Kestemont et al . , 2011 ) seeks to expose the similarity between two texts . However , it is un-able to determine if they are written by the same author . In author attribution ( Juola , 2006 ; Savoy , 2015 ) , a real author of an anonymous document is predicted using labeled documents of a set of candidate authors . Another significant subject in authorship analysis , which has received comparatively less attention from research community , is authorship-based document decomposition ( ABDD ) . This subject is to group the sentences of a multi-author document to different classes , of which each contains the sentences written by only one author . Many applications can take advantage of such a subject , especially those in forensic investigation , which aim to determine the authorship of sentences in a multi-author document . Furthermore , this kind of subject is beneficial for detecting plagiarism in a document and defining contributions of authors in a multi-author document for commercial purpose . ABDD can also be applied to identify which source ( regarded as an ' author ' in this paper ) a part of a document is copied from when the document is formed by taking contents from various sources . In despite of the benefits of ABDD , there has been little research reported on this subject . Koppel et al . ( 2011 ) are the first researchers who implemented an unsupervised approach for ABDD . However , their approach is restricted to Hebrew documents only . The authors of Akiva and Koppel ( 2013 ) addressed the drawbacks of the above approach by proposing a generic unsupervised approach for ABDD . Their approach utilized distance measurements to increase the precision and accuracy of clustering and classification phases , respectively . The accuracy of their approach is highly dependent on the number of au-thors . When the number of authors increases , the accuracy of the approach is significantly dropped . Giannella ( 2015 ) presented an improved approach for ABDD when the number of authors of the document is known or unknown . In his approach , a Bayesian segmentation algorithm is applied , which is followed by a segment clustering algorithm . However , the author tested his approach by using only documents with a few transitions among authors . Furthermore , the accuracy of the approach is very sensitive to the setting of its parameters . In Aldebei et al . ( 2015 ) , the authors presented an unsupervised approach ABDD by exploiting the differences in the posterior probabilities of a Naive-Bayesian model in order to increase the precision and the classification accuracy , and to be less dependent on the number of authors in comparing with the approach in Akiva and Koppel ( 2013 ) . Their work was tested on documents with up to 400 transitions among authors and the accuracy of their approach was not sensitive to the setting of parameters , in contrast with the approach in Giannella ( 2015 ) . However , the performance of their approach greatly depends on a threshold , of which the optimal value for an individual document is not easy to find . Some other works have focused on segmenting a document into components according to their topics . For applications where the topics of documents are unavailable , these topicbased solutions will fail . In this paper , the ABDD approach is independent of documents ' topics . All of the existing works have assumed that the observations ( i.e. , sentences ) are independent and identically distributed ( i.i.d . ) . No consideration has been given to the contextual information between the observations . However , in some cases , the i.i.d . assumption is deemed as a poor one ( Rogovschi et al . , 2010 ) . In this paper , we will relax this assumption and consider sentences of a document as a sequence of observations . We make use of the contextual information hidden between sentences in order to identify the authorship of each sentence in a document . In other words , the authorships of the " previous " and " subsequent " sentences have relationships with the authorship of the current sentence . There-fore , in this paper , a well-known sequential model , Hidden Markov Model ( HMM ) , is used for modelling the sequential patterns of the document in order to describe the authorship relationships . The contributions of this article are summarized as follows . 1 . We capture the dependencies between consecutive elements in a document to identify different authorial components and construct an HMM for classification . It is for the first time the sequential patterns hidden among document elements is considered for such a problem . 2 . To build and learn the HMM model , an unsupervised learning method is first proposed to estimate its initial parameters , and it does not require any information of authors or document 's context other than how many authors have contributed to write the document . 3 . Different from the approach in Aldebei et al . ( 2015 ) , the proposed unsupervised approach no longer relies on any predetermined threshold for ABDD . 4 . Comprehensive experiments are conducted to demonstrate the superior performance of our ideas on both widely-used artificial benchmark datasets and an authentic scientific document . As an example of its applications , the proposed approach is also applied for attributing authorship on a popular dataset . The proposed approach can not only correctly determine the author of a disputed document but also provide a way for measuring the confidence level of the authorship decision for the first time . The rest of this article is organised as follows . Section 2 reviews the HMM . Section 3 presents the details of our proposed approach , including the processes for initialization and learning of HMM parameters , and the Viterbi decoding process for classification . Experiments are conducted in Section 4 , followed by the conclusion in Section 5 . We have developed an unsupervised approach for decomposing a multi-author document based on authorship . Different from the stateof-the-art approaches , we have innovatively made use of the sequential information hidden among document elements . For this purpose , we have used HMM and constructed a sequential probabilistic model , which is used to find the best sequence of authors that represents the sentences of the document . An unsupervised learning method has also been developed to estimate the initial parameter values of HMM . Comparative experiments conducted on benchmark datasets have demonstrated the effectiveness of our ideas with superior perfor-mance achieved on both artificial and authentic documents . An application of the proposed approach on authorship attribution has also achieved perfect results of 100 % accuracies together with confidence measurement for the first time .

There is no method for multiauthor segmentation of a document into author components which can be applied to authorship verification, plagiarism detection and author attribution.

They propose a HMM-based sequential probabilistic model that captures the dependencies of sequential sentences and their authors coupled with an unsupervised initialization method.

Experiments with artificial and authentic scientific document datasets show that the proposed model outperforms existing methods and also be able to provide confidence scores.

N13-1083

We investigate two systems for automatic disfluency detection on English and Mandarin conversational speech data . The first system combines various lexical and prosodic features in a Conditional Random Field model for detecting edit disfluencies . The second system combines acoustic and language model scores for detecting filled pauses through constrained speech recognition . We compare the contributions of different knowledge sources to detection performance between these two languages . Speech disfluencies are common phenomena in spontaneous speech . They consist of spoken words and phrases that represent self-correction , hesitation , and floor-grabbing behaviors , but do not add semantic information ; removing them yields the intended , fluent utterance . The presence of disfluencies in conversational speech data can cause problems for both downstream processing ( parsing and other natural language processing tasks ) and human readability of speech transcripts . There has been much research effort on automatic disfluency detection in recent years ( Shriberg and Stolcke , 1997 ; Snover et al . , 2004 ; Liu et al . , 2006 ; Lin and Lee , 2009 ; Schuler et al . , 2010 ; Georgila et al . , 2010 ; Zwarts and Johnson , 2011 ) , particularly from the DARPA EARS ( Effective , Affordable , Reusable Speech-to-Text ) MDE ( MetaData Extraction ) ( DARPA Information Processing Technology Office , 2003 ) program , which focused on the automatic transcription of sizable amounts of speech data and rendering such transcripts in readable form , for both conversational telephone speech ( CTS ) and broadcast news ( BN ) . However , the EARS MDE effort was focused on English only , and there has n't been much research on the effectiveness of similar automatic disfluency detection approaches for multiple languages . This paper presents three main innovations . First , we extend the EARS MDE-style disfluency detection approach combining lexical and prosodic features using a Conditional Random Field ( CRF ) model , which was employed for detecting disfluency on English conversational speech data ( Liu et al . , 2005 ) , to Mandarin conversational speech , as presented in Section 2 . Second , we implement an automatic filled pause detection approach through constrained speech recognition , as presented in Section 3 . Third , for both disfluency detection systems , we compare side-by-side contributions of different knowledge sources to detection performance for two languages , English and Mandarin , as presented in Section 4 . Conclusions appear in Section 5 . In conclusion , we have presented two automatic disfluency detection systems , one combining various lexical and prosodic features , and the other combining LVCSR acoustic and language model knowledge sources . We observed significant improvements in combining lexical and prosodic features over just employing word n-gram features , for both languages . When combining AM and LM knowledge sources for FP detection in constrained speech recognition , we found increasing LM weight improved both false alarm and miss rates for Mandarin but degraded the miss rate for English .

Existing works on detecting speech disfluency which can be a problem for downstream processing and creating transcripts only focus on English.

They evaluate a Conditional Random Field-based edit disfluency detection model and a system which combines acoustic and language model that detects filled pauses in Mandarin.

Their system comparisons in English and Mandarin show that combining lexical and prosodic features achieves improvements in both languages.

P01-1026

We propose a method to generate large-scale encyclopedic knowledge , which is valuable for much NLP research , based on the Web . We first search the Web for pages containing a term in question . Then we use linguistic patterns and HTML structures to extract text fragments describing the term . Finally , we organize extracted term descriptions based on word senses and domains . In addition , we apply an automatically generated encyclopedia to a question answering system targeting the Japanese Information-Technology Engineers Examination . Reflecting the growth in utilization of the World Wide Web , a number of Web-based language processing methods have been proposed within the natural language processing ( NLP ) , information retrieval ( IR ) and artificial intelligence ( AI ) communities . A sample of these includes methods to extract linguistic resources ( Fujii and Ishikawa , 2000 ; Resnik , 1999 ; Soderland , 1997 ) , retrieve useful information in response to user queries ( Etzioni , 1997 ; McCallum et al . , 1999 ) and mine / discover knowledge latent in the Web ( Inokuchi et al . , 1999 ) . In this paper , mainly from an NLP point of view , we explore a method to produce linguistic resources . Specifically , we enhance the method proposed by Fujii and Ishikawa ( 2000 ) , which extracts encyclopedic knowledge ( i.e. , term descriptions ) from the Web . In brief , their method searches the Web for pages containing a term in question , and uses linguistic expressions and HTML layouts to extract fragments describing the term . They also use a language model to discard non-linguistic fragments . In addition , a clustering method is used to divide descriptions into a specific number of groups . On the one hand , their method is expected to enhance existing encyclopedias , where vocabulary size is relatively limited , and therefore the quantity problems has been resolved . On the other hand , encyclopedias extracted from the Web are not comparable with existing ones in terms of quality . In hand-crafted encyclopedias , term descriptions are carefully organized based on domains and word senses , which are especially effective for human usage . However , the output of Fujii 's method is simply a set of unorganized term descriptions . Although clustering is optionally performed , resultant clusters are not necessarily related to explicit criteria , such as word senses and domains . To sum up , our belief is that by combining extraction and organization methods , we can enhance both quantity and quality of Web-based encyclopedias . Motivated by this background , we introduce an organization model to Fujii 's method and reformalize the whole framework . In other words , our proposed method is not only extraction but generation of encyclopedic knowledge . Section 2 explains the overall design of our encyclopedia generation system , and Section 3 elaborates on our organization model . Section 4 then explores a method for applying our resultant encyclopedia to NLP research , specifically , question answering . Section 5 performs a number of experiments to evaluate our methods . The World Wide Web has been an unprecedentedly enormous information source , from which a number of language processing methods have been explored to extract , retrieve and discover various types of information . In this paper , we aimed at generating encyclopedic knowledge , which is valuable for many applications including human usage and natural language understanding . For this purpose , we reformalized an existing Web-based extraction method , and proposed a new statistical organization model to improve the quality of extracted data . Given a term for which encyclopedic knowledge ( i.e. , descriptions ) is to be generated , our method sequentially performs a ) retrieval of Web pages contain-ing the term , b ) extraction of page fragments describing the term , and c ) organizing extracted descriptions based on domains ( and consequently word senses ) . In addition , we proposed a question answering system , which answers interrogative questions associated with what , by using a Web-based encyclopedia as a knowledge base . For the purpose of evaluation , we used as test inputs technical terms collected from the Class II IT engineers examination , and found that the encyclopedia generated through our method was of operational quality and quantity . We also used test questions from the Class II examination , and evaluated the Web-based encyclopedia in terms of question answering . We found that our Webbased encyclopedia improved the system coverage obtained solely with an existing dictionary . In addition , when we used both resources , the performance was further improved . Future work would include generating information associated with more complex interrogations , such as ones related to how and why , so as to enhance Webbased natural language understanding .

Existing methods that extract encyclopedic knowledge from the Web output unorganized clusters of term descriptions not necessarily related to explicit criteria while clustering is performed.

They propose to use word senses and domains for organizing extracted term descriptions on questions extracted from the Web to improve the quality.

The generated encyclopedia is applied to a Japanese question and answering system and improves over a system which solely depends on a dictionary.

D15-1028

Research on modeling time series text corpora has typically focused on predicting what text will come next , but less well studied is predicting when the next text event will occur . In this paper we address the latter case , framed as modeling continuous inter-arrival times under a log-Gaussian Cox process , a form of inhomogeneous Poisson process which captures the varying rate at which the tweets arrive over time . In an application to rumour modeling of tweets surrounding the 2014 Ferguson riots , we show how interarrival times between tweets can be accurately predicted , and that incorporating textual features further improves predictions . Twitter is a popular micro-blogging service which provides real-time information on events happening across the world . Evolution of events over time can be monitored there with applications to disaster management , journalism etc . For example , Twitter has been used to detect the occurrence of earthquakes in Japan through user posts ( Sakaki et al . , 2010 ) . Modeling the temporal dynamics of tweets provides useful information about the evolution of events . Inter-arrival time prediction is a type of such modeling and has application in many settings featuring continuous time streaming text corpora , including journalism for event monitoring , real-time disaster monitoring and advertising on social media . For example , journalists track several rumours related to an event . Predicted arrival times of tweets can be applied for ranking rumours according to their activity and narrow the interest to investigate a rumour with a short interarrival time over that of a longer one . Modeling the inter-arrival time of tweets is a challenging task due to complex temporal patterns exhibited . Tweets associated with an event stream arrive at different rates at different points in time . For example , Figure 1a shows the arrival times ( denoted by black crosses ) of tweets associated with an example rumour around Ferguson riots in 2014 . Notice the existence of regions of both high and low density of arrival times over a one hour interval . We propose to address inter-arrival time prediction problem with log-Gaussian Cox process ( LGCP ) , an inhomogeneous Poisson process ( IPP ) which models tweets to be generated by an underlying intensity function which varies across time . Moreover , it assumes a non-parametric form for the intensity function allowing the model complexity to depend on the data set . We also provide an approach to consider textual content of tweets to model inter-arrival times . We evaluate the models using Twitter rumours from the 2014 Ferguson unrest , and demonstrate that they provide good predictions for inter-arrival times , beating the baselines e.g. homogeneous Poisson Process , Gaussian Process regression and univariate Hawkes Process . Even though the central application is rumours , one could apply the proposed approaches to model the arrival times of tweets corresponding to other types of memes , e.g. discussions about politics . This paper makes the following contributions : 1 . Introduces log-Gaussian Cox process to predict tweet arrival times . 2 . Demonstrates how incorporating text improves results of inter-arrival time prediction . This paper introduced the log-Gaussian Cox processes for the problem of predicting the interarrival times of tweets . We showed how text from posts helps to achieve significant improvements . Evaluation on a set of rumours from Ferguson riots showed efficacy of our methods comparing to baselines . The proposed approaches are generalizable to problems other than rumours , e.g. disaster management and advertisement campaigns .

Modeling the event inter-arrival time of tweets is challenging due to complex temporal patterns but few works aim to predict the next text event occurrence.

They propose to apply a log-Gaussian Cox process model which captures the varying arriving rate over time coupled with the textual contents of tweets.

The proposed model outperforms baseline models on an inter-arrival time prediction task around a riots rumour and shows that it improves with textual features.

P18-1222

Hypertext documents , such as web pages and academic papers , are of great importance in delivering information in our daily life . Although being effective on plain documents , conventional text embedding methods suffer from information loss if directly adapted to hyper-documents . In this paper , we propose a general embedding approach for hyper-documents , namely , hyperdoc2vec , along with four criteria characterizing necessary information that hyper-document embedding models should preserve . Systematic comparisons are conducted between hyperdoc2vec and several competitors on two tasks , i.e. , paper classification and citation recommendation , in the academic paper domain . Analyses and experiments both validate the superiority of hyperdoc2vec to other models w.r.t . the four criteria . The ubiquitous World Wide Web has boosted research interests on hypertext documents , e.g. , personal webpages ( Lu and Getoor , 2003 ) , Wikipedia pages ( Gabrilovich and Markovitch , 2007 ) , as well as academic papers ( Sugiyama and Kan , 2010 ) . Unlike independent plain documents , a hypertext document ( hyper-doc for short ) links to another hyper-doc by a hyperlink or citation mark in its textual content . Given this essential distinction , hyperlinks or citations are worth specific modeling in many tasks such as link-based classification ( Lu and Getoor , 2003 ) , web retrieval ( Page et al . , 1999 ) , entity linking ( Cucerzan , 2007 ) , and citation recommendation ( He et al . , 2010 ) . To model hypertext documents , various efforts ( Cohn and Hofmann , 2000 ; Kataria et al . , 2010 ; Perozzi et al . , 2014 ; Zwicklbauer et al . , 2016 ; Wang et al . , 2016 ) have been made to depict networks of hyper-docs as well as their content . Among potential techniques , distributed representation ( Mikolov et al . , 2013 ; Le and Mikolov , 2014 ) tends to be promising since its validity and effectiveness are proven for plain documents on many natural language processing ( NLP ) tasks . Conventional attempts on utilizing embedding techniques in hyper-doc-related tasks generally fall into two types . The first type ( Berger et al . , 2017 ; Zwicklbauer et al . , 2016 ) simply downcasts hyper-docs to plain documents and feeds them into word2vec ( Mikolov et al . , 2013 ) ( w2v for short ) or doc2vec ( Le and Mikolov , 2014 ) ( d2v for short ) . These approaches involve downgrading hyperlinks and inevitably omit certain information in hyper-docs . However , no previous work investigates the information loss , and how it affects the performance of such downcasting-based adaptations . The second type designs sophisticated embedding models to fulfill certain tasks , e.g. , citation recommendation ( Huang et al . , 2015b ) , paper classification ( Wang et al . , 2016 ) , and entity linking ( Yamada et al . , 2016 ) , etc . These models are limited to specific tasks , and it is yet unknown whether embeddings learned for those particular tasks can generalize to others . Based on the above facts , we are interested in two questions : • What information should hyper-doc embedding models preserve , and what nice property should they possess ? • Is there a general approach to learning taskindependent embeddings of hyper-docs ? To answer the two questions , we formalize the hyper-doc embedding task , and propose four criteria , i.e. , content awareness , context awareness , newcomer friendliness , and context intent aware-ness , to assess different models . Then we discuss simple downcasting-based adaptations of existing approaches w.r.t . the above criteria , and demonstrate that none of them satisfy all four . To this end , we propose hyperdoc2vec ( h-d2v for short ) , a general embedding approach for hyperdocs . Different from most existing approaches , h-d2v learns two vectors for each hyper-doc to characterize its roles of citing others and being cited . Owning to this , h-d2v is able to directly model hyperlinks or citations without downgrading them . To evaluate the learned embeddings , we employ two tasks in the academic paper domain1 , i.e. , paper classification and citation recommendation . Experimental results demonstrate the superiority of h-d2v . Comparative studies and controlled experiments also confirm that h-d2v benefits from satisfying the above four criteria . We summarize our contributions as follows : • We propose four criteria to assess different hyper-document embedding models . • We propose hyperdoc2vec , a general embedding approach for hyper-documents . • We systematically conduct comparisons with competing approaches , validating the superiority of h-d2v in terms of the four criteria . We focus on the hyper-doc embedding problem . We propose that hyper-doc embedding algorithms should be content aware , context aware , newcomer friendly , and context intent aware . To meet all four criteria , we propose a general approach , hyperdoc2vec , which assigns two vectors to each hyper-doc and models citations in a straightforward manner . In doing so , the learned embeddings satisfy all criteria , which no existing model is able to . For evaluation , paper classification and citation recommendation are conducted on three academic paper datasets . Results confirm the effectiveness of our approach . Further analyses also demonstrate that possessing the four properties helps h-d2v outperform other models .

Existing text embedding methods do not take structures of hyper-documents into account losing useful properties for downstream tasks.

They propose an embedding method for hyper-documents that learns citation information along with four criteria to assess the properties the models should preserve.

The proposed model satisfies all of the introduced criteria and performs two tasks in the academic domain better than existing models.

N18-1108

Recurrent neural networks ( RNNs ) have achieved impressive results in a variety of linguistic processing tasks , suggesting that they can induce non-trivial properties of language . We investigate here to what extent RNNs learn to track abstract hierarchical syntactic structure . We test whether RNNs trained with a generic language modeling objective in four languages ( Italian , English , Hebrew , Russian ) can predict long-distance number agreement in various constructions . We include in our evaluation nonsensical sentences where RNNs can not rely on semantic or lexical cues ( " The colorless green ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas ideas I ate with the chair sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep sleep furiously " ) , and , for Italian , we compare model performance to human intuitions . Our language-model-trained RNNs make reliable predictions about long-distance agreement , and do not lag much behind human performance . We thus bring support to the hypothesis that RNNs are not just shallowpattern extractors , but they also acquire deeper grammatical competence . Recurrent neural networks ( RNNs ; Elman , 1990 ) are general sequence processing devices that do not explicitly encode the hierarchical structure that is thought to be essential to natural language ( Everaert et al . , 2015 ) . Early work using artificial languages showed that they may nevertheless be able to approximate context-free languages ( Elman , 1991 ) . More recently , RNNs have achieved impressive results in large-scale tasks such as language modeling for speech recognition and machine translation , and are by now standard tools for sequential natural language tasks ( e.g. , Mikolov et al . , 2010 ; Graves , 2012 ; Wu et al . , 2016 ) . This suggests that RNNs may learn to track grammatical structure even when trained on noisier natural data . The conjecture is supported by the success of RNNs as feature extractors for syntactic parsing ( e.g. , Cross and Huang , 2016 ; Kiperwasser and Goldberg , 2016 ; Zhang et al . , 2017 ) . Linzen et al . ( 2016 ) directly evaluated the extent to which RNNs can approximate hierarchical structure in corpus-extracted natural language data . They tested whether RNNs can learn to predict English subject-verb agreement , a task thought to require hierarchical structure in the general case ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl the boys like . Their experiments confirmed that RNNs can , in principle , handle such constructions . However , in their study RNNs could only succeed when provided with explicit supervision on the target task . Linzen and colleagues argued that the unsupervised language modeling objective is not sufficient for RNNs to induce the syntactic knowledge necessary to cope with long-distance agreement . . . is is is is is is is is is is is is is is is is is The current paper reevaluates these conclusions . We strengthen the evaluation paradigm of Linzen and colleagues in several ways . Most importantly , their analysis did not rule out the possibility that RNNs might be relying on semantic or collocational / frequency-based information , rather than purely on syntactic structure . In " dogs dogs dogs dogs dogs about what typically barks ( dogs , not neighbourhoods ) , without relying on more abstract structural cues . In a follow-up study to Linzen and colleagues ' , Bernardy and Lappin ( 2017 ) observed that RNNs are better at long-distance agreement when they construct rich lexical representations of words , which suggests effects of this sort might indeed be at play . We introduce a method to probe the syntactic abilities of RNNs that abstracts away from potential lexical , semantic and frequency-based confounds . Inspired by Chomsky 's ( 1957 ) We extend the previous work in three additional ways . First , alongside English , which has few morphological cues to agreement , we examine Italian , Hebrew and Russian , which have richer morphological systems . Second , we go beyond subject-verb agreement and develop an automated method to harvest a variety of long-distance number agreement constructions from treebanks . Finally , for Italian , we collect human judgments for the tested sentences , providing an important comparison point for RNN performance . 1We focus on the more interesting unsupervised setup , where RNNs are trained to perform generic , large-scale language modeling ( LM ): they are not given explicit evidence , at training time , that they must focus on long-distance agreement , but they are rather required to track a multitude of cues that might help with word prediction in general . Our results are encouraging . RNNs trained with a LM objective solve the long-distance agreement problem well , even on nonce sentences . The pattern is consistent across languages , and , crucially , not far from human performance in Italian . Moreover , RNN performance on language modeling ( measured in terms of perplexity ) is a good predictor of long-distance agreement accuracy . This suggests that the ability to capture structural generalizations is an important aspect of what makes the best RNN architectures so good at language modeling . Since our positive results contradict , to some extent , those of Linzen et al . ( 2016 ) , we also replicate their relevant experiment using our best RNN ( an LSTM ) . We outperform their models , suggesting that a careful architecture / hyperparameter search is crucial to obtain RNNs that are not only good at language modeling , but able to extract syntactic generalizations . 2 Constructing a long-distance agreement benchmark Overview . We construct our number agreement test sets as follows . Original sentences are automatically extracted from a dependency treebank . They are then converted into nonce sentences by substituting all content words with random words with the same morphology , resulting in grammatical but nonsensical sequences . An LM is evaluated on its predictions for the target ( second ) word in the dependency , in both the original and nonce sentences . Long-distance agreement constructions . Agreement relations , such as subject-verb agreement in English , are an ideal test bed for the syntactic abilities of LMs , because the form of the second item ( the target ) is predictable from the first item ( the cue ) . Crucially , the cue and the target are linked by a structural relation , where linear order in the word sequence does not matter ( Everaert et al . , 2015 ) In all these cases , the number of the main verb " thinks " is determined by its subject ( " girl " ) , and this relation depends on the syntactic structure of the sentence , not on the linear sequence of words . As the last sentence shows , the word directly preceding the verb can even be a noun with the opposite number ( " friends " ) , but this does not influence the structurally-determined form of the verb . When the cue and the target are adjacent ( " the girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl girl thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks thinks . . . " ) , an LM can predict the target without access to syntactic structure : it can simply extract the relevant morphosyntactic features of words ( e.g. , number ) and record the co-occurrence frequencies of patterns such as N P lur V P lur ( Mikolov et al . , 2013 ) . Thus , we focus here on long-distance agreement , where an arbitrary num- ber of words can occur between the elements of the agreement relation . We limit ourselves to number agreement ( plural or singular ) , as it is the only overt agreement feature shared by all of the languages we study . Identifying candidate constructions . We started by collecting pairs of part-of-speech ( POS ) tags connected by a dependency arc . Independently of which element is the head of the relation , we refer to the first item as the cue and to the second as the target . We additionally refer to the POS sequence characterizing the entire pattern as a construction , and to the elements in the middle as context . For each candidate construction , we collected all of the contexts in the corpus that intervene between the cue and the target ( we define contexts as the sequence of POS tags of the top-level nodes in the dependency subtrees ) . For example , for the English subject-verb agreement construction shown in Fig . 1a , the context is defined by VERB ( head of the relative clause ) and ADV ( adverbial modifier of the target verb ) , which together dominate the sequence " the boys like often " . For the Russian adjective-noun agreement construction in Fig . 1b , the context is NOUN , because in the dependency grammar we use the noun " moment " is the head of the prepositional phrase " at that moment " , which modifies the adjective " deep " . The candidate agreement pair and the context form a construction , which is characterized by a sequence of POS tags , e.g. , NOUN VERB ADV VERB or VERB NOUN CCONJ VERB ( Fig . 1c ) . is is is is is is is is is is is is is is is is is " and " girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls girls who stayed at home were were were were were were were were were were were were were were were were were " . Conversely , standard syntactic structures might be split between different constructions , e.g. , relative clause contexts occur in both NOUN VERB VERB and NOUN VERB ADV VERB constructions ( the latter is illustrated by the English example in Fig . 1a ) . Construction contexts can contain a variable numbers of words . Since we are interested in challenging cases , we only considered cases in which at least three tokens intervened between the cue and the target . Excluding non-agreement constructions . In the next step , we excluded constructions in which the candidate cue and target did not agree in number in all of the instances of the construction in the treebank ( if both the cue and the target were morphologically annotated for number ) . This step retained English subject-verb constructions , for example , but excluded verb-object constructions , since any form of a verb can appear both with singular and plural objects . To focus on robust agreement patterns , we only kept constructions with at least 10 instances of both plural and singular agreement . When applied to the treebanks we used ( see Section 3 ) , this step resulted in between two ( English ) and 21 ( Russian ) constructions per lan-guage . English has the poorest morphology and consequently the lowest number of patterns with identifiable morphological agreement . Only the VP-conjunction construction ( Fig . 1c ) was identified in all four languages . Subject-verb agreement constructions were extracted in all languages but Russian ; Russian has relatively flexible word order and a noun dependent preceding a head verb is not necessarily its subject . The full list of extracted constructions in English and Italian is given in Tables 2 and 3 , respectively . For the other languages , see the Supplementary Material ( SM ) . 2 Original sentence test set . Our " original " sentence test set included all sentences from each construction where all words from the cue and up to and including the target occurred in the LM vocabulary ( Section 3 ) , and where the singular / plural counterpart of the target occurred in the treebank and in the language model vocabulary ( this is required by the evaluation procedure outlined below ) . The total counts of constructions and original sentences in our test sets are provided in Table 1 . The average number of context words separating the cue and the target ranged from 3.6 ( Hebrew ) to 4.5 ( Italian ) . Generating nonce sentences . We generated nine nonce variants of each original sentence as follows . Each content word ( noun , verb , adjective , proper noun , numeral , adverb ) in the sentence was substituted by another random content word from the treebank with matching POS and morphological features . To avoid forms that are ambiguous between several POS , which are particularly frequent in English ( e.g. , plural noun and singular verb forms ) , we excluded the forms that appeared with a different POS more than 10 % of the time in the treebank . Function words ( determiners , pronouns , adpositions , particles ) and punctuation were left intact . For example , we generated the nonce ( 1b ) from the original sentence ( 1a ): ( ( e.g. , " it stays the shuttle " in ( 1b ) ) . Evaluation procedure . For each sentence in our test set , we retrieved from our treebank the form that is identical to the agreement target in all morphological features except number ( e.g. , " finds " instead of " find " in ( 1b ) ) . Given a sentence with prefix p up to and excluding the target , we then compute the probabilities P ( t 1 |p ) and P ( t 2 |p ) for the singular and plural variants of the target , t 1 and t 2 , based on the language model . Following Linzen et al . ( 2016 ) , we say that the model identified the correct target if it assigned a higher probability to the form with the correct number . In ( 1b ) , for example , the model should assign a higher probability to " finds " than " find".3 3 Experimental setup Treebanks . We extracted our test sets from the Italian , English , Hebrew and Russian Universal Dependency treebanks ( UD , v2.0 , Nivre et al . , 2016 ) . The English and Hebrew treebanks were post-processed to obtain a richer morphological annotation at the word level ( see SM for details ) . We ran an extensive analysis of the abilities of RNNs trained on a generic language-modeling task to predict long-distance number agreement . Results were consistent across four languages and a number of constructions . They were above strong baselines even in the challenging case of nonsense sentences , and not far from human performance . We are not aware of other collections of human long-distance agreement judgments on nonsensical sentences , and we thus consider our publicly available data set an important contribution of our work , of interest to students of human language processing in general . The constructions we considered are quite infrequent ( according to a rough estimate based on the treebanks , the language in which they are most common is Hebrew , and even there they occur with average 0.8 % sentence frequency ) . Moreover , they vary in the contexts that separate the cue and the target . So , RNNs are not simply memorizing frequent morphosyntactic sequences ( which would already be impressive , for systems learning from raw text ) . We tentatively conclude that LMtrained RNNs can construct abstract grammatical representations of their input . This , in turn , suggests that the input itself contains enough information to trigger some form of syntactic learning in a system , such as an RNN , that does not contain an explicit prior bias in favour of syntactic structures . In future work , we would like to better understand what kind of syntactic information RNNs are encoding , and how . On the one hand , we plan to adapt methods to inspect information flow across RNN states ( e.g. , Hupkes et al . , 2017 ) . On the other , we would like to expand our empirical investigation by focusing on other long-distance phenomena , such as overt case assignment ( Blake , 2001 ) or parasitic gap licensing ( Culicover and Postal , 2001 ) . While it is more challenging to extract reliable examples of such phenomena from corpora , their study would probe more sophisticated syntactic capabilities , possibly even shedding light on the theoretical analysis of the underlying linguistic structures . Finally , it may be useful to complement the corpus-driven approach used in the current paper with constructed evaluation sentences that isolate particular syntactic phenomena , independent of their frequency in a natural corpus , as is common in psycholinguistics ( Enguehard et al . , 2017 ) .

Previous work only shows that RNNs can handle constructions that require hierarchical structure when explicit supervision on the target task is given.

They introduce a probing method for syntactic abilities to evaluate long-distance agreement on standard and nonsensical sentences in multiple languages with different morphological systems.

The RNNs trained on an LM objective can solve long-distance agreement problems well even on nonsensical sentences consistently across languages indicating their deeper grammatical competence.

D09-1115

Current system combination methods usually use confusion networks to find consensus translations among different systems . Requiring one-to-one mappings between the words in candidate translations , confusion networks have difficulty in handling more general situations in which several words are connected to another several words . Instead , we propose a lattice-based system combination model that allows for such phrase alignments and uses lattices to encode all candidate translations . Experiments show that our approach achieves significant improvements over the state-ofthe-art baseline system on Chinese-to-English translation test sets . System combination aims to find consensus translations among different machine translation systems . It has been proven that such consensus translations are usually better than the output of individual systems ( Frederking and Nirenburg , 1994 ) . In recent several years , the system combination methods based on confusion networks developed rapidly ( Bangalore et al . , 2001 ; Matusov et al . , 2006 ; Sim et al . , 2007 ; Rosti et al . , 2007a ; Rosti et al . , 2007b ; Rosti et al . , 2008 ; He et al . , 2008 ) , which show state-of-the-art performance in benchmarks . A confusion network consists of a sequence of sets of candidate words . Each candidate word is associated with a score . The optimal consensus translation can be obtained by selecting one word from each set to maximizing the overall score . To construct a confusion network , one first need to choose one of the hypotheses ( i.e. , candidate translations ) as the backbone ( also called " skeleton " in the literature ) and then decide the word alignments of other hypotheses to the backbone . Hypothesis alignment plays a crucial role in confusionnetwork-based system combination because it has a direct effect on selecting consensus translations . However , a confusion network is restricted in such a way that only 1-to-1 mappings are allowed in hypothesis alignment . This is not the fact even for word alignments between the same languages . It is more common that several words are connected to another several words . For example , " be capable of " and " be able to " have the same meaning . Although confusion-network-based approaches resort to inserting null words to alleviate this problem , they face the risk of producing degenerate translations such as " be capable to " and " be able of " . In this paper , we propose a new system combination method based on lattices . As a more general form of confusion network , a lattice is capable of describing arbitrary mappings in hypothesis alignment . In a lattice , each edge is associated with a sequence of words rather than a single word . Therefore , we select phrases instead of words in each candidate set and minimize the chance to produce unexpected translations such as " be capable to " . We compared our approach with the state-of-the-art confusion-network-based system ( He et al . , 2008 ) and achieved a significant absolute improvement of 1.23 BLEU points on the NIST 2005 Chinese-to-English test set and 0.93 BLEU point on the NIST 2008 Chinese-to-English test set . The alignment model plays an important role in system combination . Because of the expression limitation of confusion networks , only 1-to-1 mappings are employed in the confusion-network-based model . This paper proposes a lattice-based system combination model . As a general form of confusion networks , lattices can express n-to-n mappings . a confusion-network-based model processes words only . As a result , phrase pairs must be extracted before constructing a lattice . On NIST MT05 test set , the lattice-based system gave better results with an absolute improvement of 1.23 BLEU points over the confusionnetwork-based system ( He et al . , 2008 ) and 3.73 BLEU points over the best single system . On NIST MT08 test set , the lattice-based system outperformed the confusion-network-based system by 0.93 BLEU point and outperformed the best single system by 3.0 BLEU points .

System combination methods based on confusion networks only allow word level 1-to-1 mappings, and some workarounds cause another type of problem such as degeneration.

They propose to use lattices to combine systems that enable to process of a sequence of words rather than one word that can mitigate degeneration.

They show that their approach significantly outperforms the state-of-the-art confusion-network-based systems on Chinese-to-English translation tasks.

E17-1110

The growing demand for structured knowledge has led to great interest in relation extraction , especially in cases with limited supervision . However , existing distance supervision approaches only extract relations expressed in single sentences . In general , cross-sentence relation extraction is under-explored , even in the supervised-learning setting . In this paper , we propose the first approach for applying distant supervision to crosssentence relation extraction . At the core of our approach is a graph representation that can incorporate both standard dependencies and discourse relations , thus providing a unifying way to model relations within and across sentences . We extract features from multiple paths in this graph , increasing accuracy and robustness when confronted with linguistic variation and analysis error . Experiments on an important extraction task for precision medicine show that our approach can learn an accurate cross-sentence extractor , using only a small existing knowledge base and unlabeled text from biomedical research articles . Compared to the existing distant supervision paradigm , our approach extracted twice as many relations at similar precision , thus demonstrating the prevalence of cross-sentence relations and the promise of our approach . The accelerating pace in technological advance and scientific discovery has led to an explosive growth in knowledge . The ensuing information overload creates new urgency in assimilating frag-mented knowledge for integration and reasoning . A salient case in point is precision medicine ( Bahcall , 2015 ) . The cost of sequencing a person 's genome has fallen below $ 10001 , enabling individualized diagnosis and treatment of complex genetic diseases such as cancer . The availability of measurement for 20,000 human genes makes it imperative to integrate all knowledge about them , which grows rapidly and is scattered in millions of articles in PubMed2 . Traditional extraction approaches require annotated examples , which makes it difficult to scale to the explosion of extraction demands . Consequently , there has been increasing interest in indirect supervision ( Banko et al . , 2007 ; Poon and Domingos , 2009 ; Toutanova et al . , 2015 ) , with distant supervision ( Craven et al . , 1998 ; Mintz et al . , 2009 ) emerging as a particularly promising paradigm for augmenting existing knowledge bases from unlabeled text ( Poon et al . , 2015 ; Parikh et al . , 2015 ) . This progress is exciting , but distantsupervision approaches have so far been limited to single sentences , thus missing out on relations crossing the sentence boundary . Consider the following example:"The p56Lck inhibitor Dasatinib was shown to enhance apoptosis induction by dexamethasone in otherwise GC-resistant CLL cells . This finding concurs with the observation by Sade showing that Notch-mediated resistance of a mouse lymphoma cell line could be overcome by inhibiting p56Lck . " Together , the two sentences convey the fact that the drug Dasatinib could overcome resistance conferred by mutations to the Notch gene , which can not be inferred from either sentence alone . The impact of missed opportunities is especially pronounced in the long tail of knowledge . Such information is crucial for integrative reasoning as it includes the newest findings in specialized domains . In this paper , we present DISCREX , the first approach for distant supervision to relation extraction beyond the sentence boundary . The key idea is to adopt a document-level graph representation that augments conventional intra-sentential dependencies with new dependencies introduced for adjacent sentences and discourse relations . It provides a unifying way to derive features for classifying relations between entity pairs . As we augment this graph with new arcs , the number of possible paths between entities grow . We demonstrate that feature extraction along multiple paths leads to more robust extraction , allowing the learner to find structural patterns even when the language varies or the parser makes an error . The cross-sentence scenario presents a new challenge in candidate selection . This motivates our concept of minimal-span candidates in Section 3.2 . Excluding non-minimal candidates substantially improves classification accuracy . There is a long line of research on discourse phenomena , including coreference ( Haghighi and Klein , 2007 ; Poon and Domingos , 2008 ; Rahman and Ng , 2009 ; Raghunathan et al . , 2010 ) , narrative structures ( Chambers and Jurafsky , 2009 ; Cheung et al . , 2013 ) , and rhetorical relations ( Marcu , 2000 ) . For the most part , this work has not been connected to relation extraction . Our proposed extraction framework makes it easy to integrate such discourse relations . Our experiments evaluated the impact of coreference and discourse parsing , a preliminary step toward in-depth integration with discourse research . We conducted experiments on extracting druggene interactions from biomedical literature , an important task for precision medicine . By bootstrapping from a recently curated knowledge base ( KB ) with about 162 known interactions , our DIS-CREX system learned to extract inter-sentence drug-gene interactions at high precision . Crosssentence extraction doubled the yield compared to single-sentence extraction . Overall , by applying distant supervision , we extracted about 64,000 distinct interactions from about one million PubMed Central full-text articles , attaining two orders of magnitude increase compared to the original KB . We present the first approach for applying distant supervision to cross-sentence relation extraction , by adopting a document-level graph representation that incorporates both intra-sentential dependencies and inter-sentential relations such as adjacency and discourse relations . We conducted both automatic and manual evaluation on extracting drug-gene interactions from biomedical literature . With cross-sentence extraction , our DIS-CREX system doubled the yield of unique interactions , while maintaining the same accuracy . Using distant supervision , DISCREX improved the coverage of the Gene Drug Knowledge Database ( GDKD ) by two orders of magnitude , without requiring annotated examples . Future work includes : further exploration of features ; improved integration with coreference and discourse parsing ; combining distant supervision with active learning and crowd sourcing ; evaluate the impact of extractions to precision medicine ; applications to other domains .

Existing distance supervision methods for relation extraction cannot capture relations crossing the sentence boundary which is important in specialized domains with long-tail knowledge.

They propose a method for applying distance supervision to cross-sentence relation extraction by adopting a document-level graph representation that incorporates intra-sentential dependencies and inter-sentential relations.

Experiments on extracting drug-gene interactions from biomedical literature show that the proposed method doubles the performance of single-sentence extraction methods.

P07-1026

Convolution tree kernel has shown promising results in semantic role classification . However , it only carries out hard matching , which may lead to over-fitting and less accurate similarity measure . To remove the constraint , this paper proposes a grammardriven convolution tree kernel for semantic role classification by introducing more linguistic knowledge into the standard tree kernel . The proposed grammar-driven tree kernel displays two advantages over the previous one : 1 ) grammar-driven approximate substructure matching and 2 ) grammardriven approximate tree node matching . The two improvements enable the grammardriven tree kernel explore more linguistically motivated structure features than the previous one . Experiments on the CoNLL-2005 SRL shared task show that the grammardriven tree kernel significantly outperforms the previous non-grammar-driven one in SRL . Moreover , we present a composite kernel to integrate feature-based and tree kernel-based methods . Experimental results show that the composite kernel outperforms the previously best-reported methods . Given a sentence , the task of Semantic Role Labeling ( SRL ) consists of analyzing the logical forms expressed by some target verbs or nouns and some constituents of the sentence . In particular , for each predicate ( target verb or noun ) all the constituents in the sentence which fill semantic arguments ( roles ) of the predicate have to be recognized . Typical semantic roles include Agent , Patient , Instrument , etc . and also adjuncts such as Locative , Temporal , Manner , and Cause , etc . Generally , semantic role identification and classification are regarded as two key steps in semantic role labeling . Semantic role identification involves classifying each syntactic element in a sentence into either a semantic argument or a non-argument while semantic role classification involves classifying each semantic argument identified into a specific semantic role . This paper focuses on semantic role classification task with the assumption that the semantic arguments have been identified correctly . Both feature-based and kernel-based learning methods have been studied for semantic role classification ( Carreras and Màrquez , 2004 ; Carreras and Màrquez , 2005 ) . In feature-based methods , a flat feature vector is used to represent a predicateargument structure while , in kernel-based methods , a kernel function is used to measure directly the similarity between two predicate-argument structures . As we know , kernel methods are more effective in capturing structured features . Moschitti ( 2004 ) and Che et al . ( 2006 ) used a convolution tree kernel ( Collins and Duffy , 2001 ) for semantic role classification . The convolution tree kernel takes sub-tree as its feature and counts the number of common sub-trees as the similarity between two predicate-arguments . This kernel has shown very promising results in SRL . However , as a general learning algorithm , the tree kernel only carries out hard matching between any two sub-trees without considering any linguistic knowledge in kernel design . This makes the kernel fail to handle similar phrase structures ( e.g. , " buy a car " vs. " buy a red car " ) and near-synonymic grammar tags ( e.g. , the POS variations between " high / JJ degree / NN " 1 and " higher / JJR degree / NN " ) 2 . To some degree , it may lead to over-fitting and compromise performance . This paper reports our preliminary study in addressing the above issue by introducing more linguistic knowledge into the convolution tree kernel . To our knowledge , this is the first attempt in this research direction . In detail , we propose a grammar-driven convolution tree kernel for semantic role classification that can carry out more linguistically motivated substructure matching . Experimental results show that the proposed method significantly outperforms the standard convolution tree kernel on the data set of the CoNLL-2005 SRL shared task . The remainder of the paper is organized as follows : Section 2 reviews the previous work and Section 3 discusses our grammar-driven convolution tree kernel . Section 4 shows the experimental results . We conclude our work in Section 5 . In this paper , we propose a novel grammar-driven convolution tree kernel for semantic role classification . More linguistic knowledge is considered in the new kernel design . The experimental results verify that the grammar-driven kernel is more effective in capturing syntactic structure features than the previous convolution tree kernel because it allows grammar-driven approximate matching of substructures and node features . We also discuss the criteria to determine the optional nodes in a CFG rule in defining our grammar-driven convolution tree kernel . The extension of our work is to improve the performance of the entire semantic role labeling system using the grammar-driven tree kernel , including all four stages : pruning , semantic role identification , classification and post inference . In addition , a more interesting research topic is to study how to integrate linguistic knowledge and tree kernel methods to do feature selection for tree kernelbased NLP applications ( Suzuki et al . , 2004 ) . In detail , a linguistics and statistics-based theory that can suggest the effectiveness of different substructure features and whether they should be generated or not by the tree kernels would be worked out .

Despite its success in semantic role classification, convolution tree kernels based on the hard matching between two sub-trees suffer from over-fitting.

They propose to integrate a linguistically motivated grammar-baed convolution tree kernel into a standard tree kernel to achieve better substructure matching and tree node matching.

The new grammar-driven tree kernel significantly outperforms baseline kernels on the CoNLL-2005 task.

E09-1032

We explore the problem of resolving the second person English pronoun you in multi-party dialogue , using a combination of linguistic and visual features . First , we distinguish generic and referential uses , then we classify the referential uses as either plural or singular , and finally , for the latter cases , we identify the addressee . In our first set of experiments , the linguistic and visual features are derived from manual transcriptions and annotations , but in the second set , they are generated through entirely automatic means . Results show that a multimodal system is often preferable to a unimodal one . The English pronoun you is the second most frequent word in unrestricted conversation ( after I and right before it ) . 1 Despite this , with the exception of Gupta et al . ( 2007b ; 2007a ) , its resolution has received very little attention in the literature . This is perhaps not surprising since the vast amount of work on anaphora and reference resolution has focused on text or discourse -mediums where second-person deixis is perhaps not as prominent as it is in dialogue . For spoken dialogue pronoun resolution modules however , resolving you is an essential task that has an important impact on the capabilities of dialogue summarization systems . Besides being important for computational implementations , resolving you is also an interesting and challenging research problem . As for third person pronouns such as it , some uses of you are not strictly referential . These include discourse marker uses such as you know in example ( 1 ) , and generic uses like ( 2 ) , where you does not refer to the addressee as it does in ( 3 ) . ( 1 ) It 's not just , you know , noises like something hitting . ( 2 ) Often , you need to know specific button sequences to get certain functionalities done . ( 3 ) I think it 's good . You 've done a good review . However , unlike it , you is ambiguous between singular and plural interpretations -an issue that is particularly problematic in multi-party conversations . While you clearly has a plural referent in ( 4 ) , in ( 3 ) the number of its referent is ambiguous . 2(4 ) I do n't know if you guys have any questions . When an utterance contains a singular referential you , resolving the you amounts to identifying the individual to whom the utterance is addressed . This is trivial in two-person dialogue since the current listener is always the addressee , but in conversations with multiple participants , it is a complex problem where different kinds of linguistic and visual information play important roles ( Jovanovic , 2007 ) . One of the issues we investigate here is how this applies to the more concrete problem of resolving the second person pronoun you . We approach this issue as a three-step problem . Using the AMI Meeting Corpus ( McCowan et al . , 2005 ) of multi-party dialogues , we first discriminate between referential and generic uses of you . Then , within the referential uses , we distinguish between singular and plural , and finally , we resolve the singular referential instances by identifying the intended addressee . We use multimodal features : initially , we extract discourse features from manual transcriptions and use visual information derived from manual annotations , but then we move to a fully automatic approach , using 1-best transcriptions produced by an automatic speech recognizer ( ASR ) and visual features automatically extracted from raw video . In the next section of this paper , we give a brief overview of related work . We describe our data in Section 3 , and explain how we extract visual and linguistic features in Sections 4 and 5 respectively . Section 6 then presents our experiments with manual transcriptions and annotations , while Section 7 , those with automatically extracted information . We end with conclusions in Section 8 . We have investigated the automatic resolution of the second person English pronoun you in multi-party dialogue , using a combination of linguistic and visual features . We conducted a first set of experiments where our features were derived from manual transcriptions and annotations , and then a second set where they were generated by entirely automatic means . To our knowledge , this is the first attempt at tackling this problem using automatically extracted multimodal information . Our experiments showed that visual information can be highly predictive in resolving the addressee of singular referential uses of you . Visual features significantly improved the performance of both our manual and automatic systems , and the latter achieved an encouraging 75 % accuracy . We also found that our visual features had predictive power for distinguishing between generic and referential uses of you , and between referential singulars and plurals . Indeed , for the latter task , they significantly improved the manual system 's performance . The listeners ' gaze features were useful here : in our data set it was apparently the case that the speaker would often use the whiteboard / projector screen when addressing the group , thus drawing the listeners ' gaze in this direction . Future work will involve expanding our dataset , and investigating new potentially predictive features . In the slightly longer term , we plan to integrate the resulting system into a meeting assistant whose purpose is to automatically extract useful information from multi-party meetings .

Although the word "you" is frequently used and has several possible meanings, such as reference or generic, it is not well studied yet.

They first manually automatically separate the word "you" between generic and referential, then later use a multimodal system for automation.

They show that visual features can help distinguish the word "you" in multi-party conversations.

P10-1139

There is a growing research interest in opinion retrieval as on-line users ' opinions are becoming more and more popular in business , social networks , etc . Practically speaking , the goal of opinion retrieval is to retrieve documents , which entail opinions or comments , relevant to a target subject specified by the user 's query . A fundamental challenge in opinion retrieval is information representation . Existing research focuses on document-based approaches and documents are represented by bag-of-word . However , due to loss of contextual information , this representation fails to capture the associative information between an opinion and its corresponding target . It can not distinguish different degrees of a sentiment word when associated with different targets . This in turn seriously affects opinion retrieval performance . In this paper , we propose a sentence-based approach based on a new information representation , namely topic-sentiment word pair , to capture intra-sentence contextual information between an opinion and its target . Additionally , we consider inter-sentence information to capture the relationships among the opinions on the same topic . Finally , the two types of information are combined in a unified graph-based model , which can effectively rank the documents . Compared with existing approaches , experimental results on the COAE08 dataset showed that our graph-based model achieved significant improvement . In recent years , there is a growing interest in sharing personal opinions on the Web , such as product reviews , economic analysis , political polls , etc . These opinions can not only help independent users make decisions , but also obtain valuable feedbacks ( Pang et al . , 2008 ) . Opinion oriented research , including sentiment classifica-tion , opinion extraction , opinion question answering , and opinion summarization , etc . are receiving growing attention ( Wilson , et al . , 2005 ; Liu et al . , 2005 ; Oard et al . , 2006 ) . However , most existing works concentrate on analyzing opinions expressed in the documents , and none on how to represent the information needs required to retrieve opinionated documents . In this paper , we focus on opinion retrieval , whose goal is to find a set of documents containing not only the query keyword(s ) but also the relevant opinions . This requirement brings about the challenge on how to represent information needs for effective opinion retrieval . In order to solve the above problem , previous work adopts a 2-stage approach . In the first stage , relevant documents are determined and ranked by a score , i.e. tf-idf value . In the second stage , an opinion score is generated for each relevant document ( Macdonald and Ounis , 2007 ; Oard et al . , 2006 ) . The opinion score can be acquired by either machine learning-based sentiment classifiers , such as SVM ( Zhang and Yu , 2007 ) , or a sentiment lexicons with weighted scores from training documents ( Amati et al . , 2007 ; Hannah et al . , 2007 ; Na et al . , 2009 ) . Finally , an overall score combining the two is computed by using a score function , e.g. linear combination , to re-rank the retrieved documents . Retrieval in the 2-stage approach is based on document and document is represented by bag-of-word . This representation , however , can only ensure that there is at least one opinion in each relevant document , but it can not determine the relevance pairing of individual opinion to its target . In general , by simply representing a document in bag-of-word , contextual information i.e. the corresponding target of an opinion , is neglected . This may result in possible mismatch between an opinion and a target and in turn affects opinion retrieval performance . By the same token , the effect to documents consisting of mul-tiple topics , which is common in blogs and on-line reviews , is also significant . In this setting , even if a document is regarded opinionated , it can not ensure that all opinions in the document are indeed relevant to the target concerned . Therefore , we argue that existing information representation i.e. bag-of-word , can not satisfy the information needs for opinion retrieval . In this paper , we propose to handle opinion retrieval in the granularity of sentence . It is observed that a complete opinion is always expressed in one sentence , and the relevant target of the opinion is mostly the one found in it . Therefore , it is crucial to maintain the associative information between an opinion and its target within a sentence . We define the notion of a topic-sentiment word pair , which is composed of a topic term ( i.e. the target ) and a sentiment word ( i.e. opinion ) of a sentence . Word pairs can maintain intra-sentence contextual information to express the potential relevant opinions . In addition , inter-sentence contextual information is also captured by word pairs to represent the relationship among opinions on the same topic . In practice , the inter-sentence information reflects the degree of a word pair . Finally , we combine both intra-sentence and inter-sentence contextual information to construct a unified undirected graph to achieve effective opinion retrieval . The rest of the paper is organized as follows . In Section 2 , we describe the motivation of our approach . Section 3 presents a novel unified graph-based model for opinion retrieval . We evaluated our model and the results are presented in Section 4 . We review related works on opinion retrieval in Section 5 . Finally , in Section 6 , the paper is concluded and future work is suggested . In this work we focus on the problem of opinion retrieval . Different from existing approaches , which regard document relevance as the key indicator of opinion relevance , we propose to explore the relevance of individual opinion . To do that , opinion retrieval is performed in the granularity of sentence . We define the notion of word pair , which can not only maintain the association between the opinion and the corresponding target in the sentence , but it can also build up the relationship among sentences through the same word pair . Furthermore , we convert the relationships between word pairs and sentences into a unified graph , and use the HITS algorithm to achieve document ranking for opinion retrieval . Finally , we compare our approach with existing methods . Experimental results show that our proposed model performs well on COAE08 dataset . The novelty of our work lies in using word pairs to represent the information needs for opinion retrieval . On the one hand , word pairs can identify the relevant opinion according to intra-sentence contextual information . On the other hand , word pairs can measure the degree of a relevant opinion by taking inter-sentence contextual information into consideration . With the help of word pairs , the information needs for opinion retrieval can be represented appropriately . In the future , more research is required in the following directions : ( 1 ) Since word pairs can indicate relevant opinions effectively , it is worth further study on how they could be applied to other opinion oriented applications , e.g. opinion summarization , opinion prediction , etc . ( 2 ) The characteristics of blogs will be taken into consideration , i.e. , the post time , which could be helpful to create a more time sensitivity graph to filter out fake opinions . ( 3 ) Opinion holder is another important role of an opinion , and the identification of opinion holder is a main task in NTCIR . It would be interesting to study opinion holders , e.g. its seniority , for opinion retrieval .

Existing approaches to the opinion retrieval task represent documents using bag-of-words disregarding contextual information between an opinion and its corresponding text.

They propose a sentence-based approach which captures both inter and intra sentence contextual information combined with a unified undirected graph.

The proposed method outperforms existing approaches on the COAE08 dataset showing that word pairs can represent information for opinion retrieval well.

N03-1024

We describe a syntax-based algorithm that automatically builds Finite State Automata ( word lattices ) from semantically equivalent translation sets . These FSAs are good representations of paraphrases . They can be used to extract lexical and syntactic paraphrase pairs and to generate new , unseen sentences that express the same meaning as the sentences in the input sets . Our FSAs can also predict the correctness of alternative semantic renderings , which may be used to evaluate the quality of translations . In the past , paraphrases have come under the scrutiny of many research communities . Information retrieval researchers have used paraphrasing techniques for query reformulation in order to increase the recall of information retrieval engines ( Sparck Jones and Tait , 1984 ) . Natural language generation researchers have used paraphrasing to increase the expressive power of generation systems ( Iordanskaja et al . , 1991 ; Lenke , 1994 ; Stede , 1999 ) . And researchers in multi-document text summarization ( Barzilay et al . , 1999 ) , information extraction ( Shinyama et al . , 2002 ) , and question answering ( Lin and Pantel , 2001 ; Hermjakob et al . , 2002 ) have focused on identifying and exploiting paraphrases in the context of recognizing redundancies , alternative formulations of the same meaning , and improving the performance of question answering systems . In previous work ( Barzilay and McKeown , 2001 ; Lin and Pantel , 2001 ; Shinyama et al . , 2002 ) , paraphrases are represented as sets or pairs of semantically equivalent words , phrases , and patterns . Although this is adequate in the context of some applications , it is clearly too weak from a generative perspective . Assume , for example , that we know that text pairs ( stock market rose , stock market gained ) and ( stock market rose , stock prices rose ) have the same meaning . If we memorized only these two pairs , it would be impossible to infer that , in fact , consistent with our intuition , any of the following sets of phrases are also semantically equivalent : { stock market rose , stock market gained , stock prices rose , stock prices gained } and { stock market , stock prices } in the context of rose or gained ; { market rose } , { market gained } , { prices rose } and { prices gained } in the context of stock ; and so on . In this paper , we propose solutions for two problems : the problem of paraphrase representation and the problem of paraphrase induction . We propose a new , finite-statebased representation of paraphrases that enables one to encode compactly large numbers of paraphrases . We also propose algorithms that automatically derive such representations from inputs that are now routinely released in conjunction with large scale machine translation evaluations ( DARPA , 2002 ) : multiple English translations of many foreign language texts . For instance , when given as input the 11 semantically equivalent English translations in Figure 1 , our algorithm automatically induces the FSA in Figure 2 , which represents compactly 49 distinct renderings of the same semantic meaning . Our FSAs capture both lexical paraphrases , such as { fighting , bat-tle } , { died , were killed } and structural paraphrases such as { last week 's fighting , the battle of last week } . The contexts in which these are correct paraphrases are also conveniently captured in the representation . In previous work , Langkilde and Knight ( 1998 ) used word lattices for language generation , but their method involved hand-crafted rules . Bangalore et al . ( 2001 ) and Barzilay and Lee ( 2002 ) both applied the technique of multi-sequence alignment ( MSA ) to align parallel corpora and produced similar FSAs . For their purposes , they mainly need to ensure the correctness of consensus among different translations , so that different constituent orderings in input sentences do not pose a serious prob- lem . In contrast , we want to ensure the correctness of all paths represented by the FSAs , and direct application of MSA in the presence of different constituent orderings can be problematic . For example , when given as input the same sentences in Figure 1 , one instantiation of the MSA algorithm produces the FSA in Figure 3 , which contains many " bad " paths such as the battle of last week 's fighting took at least 12 people lost their people died in the fighting last week 's fighting ( See Section 4.2.2 for a more quantitative analysis . ) . It 's still possible to use MSA if , for example , the input is pre-clustered to have the same constituent ordering ( Barzilay and Lee ( 2003 ) ) . But we chose to approach this problem from another direction . As a result , we propose a new syntax-based algorithm to produce FSAs . In this paper , we first introduce the multiple translation corpus that we use in our experiments ( see Section 2 ) . We then present the algorithms that we developed to induce finite-state paraphrase representations from such data ( see Section 3 ) . An important part of the paper is dedicated to evaluating the quality of the finite-state representations that we derive ( see Section 4 ) . Since our representations encode thousands and sometimes millions of equivalent verbalizations of the same meaning , we use both manual and automatic evaluation techniques . Some of the automatic evaluations we perform are novel as well . In this paper , we presented a new syntax-based algorithm that learns paraphrases from a newly available dataset . The multiple translation corpus that we use in this paper is the first instance in a series of similar corpora that are built and made publicly available by LDC in the context of a series of DARPA-sponsored MT evaluations . The algorithm we proposed constructs finite state representations of paraphrases that are useful in many contexts : to induce large lists of lexical and structural paraphrases ; to generate semantically equivalent renderings of a given meaning ; and to estimate the quality of machine translation systems . More experiments need to be carried out in order to assess extrinsically whether the FSAs we produce can be used to yield higher agreement scores between human and automatic assessments of translation quality . In our future work , we wish to experiment with more flexible merging algorithms and to integrate better the top-down and bottom-up processes that are used to in-duce FSAs . We also wish to extract more abstract paraphrase patterns from the current representation . Such patterns are more likely to get reused -which would help us get reliable statistics for them in the extraction phase , and also have a better chance of being applicable to unseen data .

Existing approaches to represent paraphrases as sets or pairs of semantically equivalent words, phrases and patterns that are weak for text generation purposes.

They propose a syntax-based algorithm that builds Finite State Automata from translation sets which are good representations of paraphrases.

Manual and automatic evaluations show that the representations extracted by the proposed method can be used for automatic translation evaluations.

P18-1159

While sophisticated neural-based techniques have been developed in reading comprehension , most approaches model the answer in an independent manner , ignoring its relations with other answer candidates . This problem can be even worse in open-domain scenarios , where candidates from multiple passages should be combined to answer a single question . In this paper , we formulate reading comprehension as an extract-then-select twostage procedure . We first extract answer candidates from passages , then select the final answer by combining information from all the candidates . Furthermore , we regard candidate extraction as a latent variable and train the two-stage process jointly with reinforcement learning . As a result , our approach has improved the state-ofthe-art performance significantly on two challenging open-domain reading comprehension datasets . Further analysis demonstrates the effectiveness of our model components , especially the information fusion of all the candidates and the joint training of the extract-then-select procedure . Teaching machines to read and comprehend human languages is a long-standing objective in natural language processing . In order to evaluate this ability , reading comprehension ( RC ) is designed to answer questions through reading relevant passages . In recent years , RC has attracted intense interest . Various advanced neural models have been proposed along with newly released datasets ( Hermann et al . , 2015 ; Rajpurkar et al . , 2016 ; Dunn et al . , 2017 ; Dhingra et al . , 2017b ; He et al . , 2017 ) . In this paper , we formulate the problem of RC as a two-stage process , which first generates candidates with an extraction model , then selects the final answer by combining the information from all the candidates . Furthermore , we treat candidate extraction as a latent variable and jointly train these two stages with RL . Experiments on public open-domain RC datasets Quasar-T and SearchQA show the necessity of introducing the selection model and the effectiveness of fusing candidates information when modeling . Moreover , our joint training strategy leads to significant improvements in performance .

Existing models for reading comprehension do not consider multiple answer candidates which can be problematic when they need to fuse information from multiple passages.

They propose to approach reading comprehension with an extract-then-select procedure, where a model learns two tasks jointly using latent variables and reinforcement learning.

The proposed model can fuse answer candidates from multiple candidates and significantly outperform existing models on two open-domain reading comprehension tasks.

W06-1672

We present two discriminative methods for name transliteration . The methods correspond to local and global modeling approaches in modeling structured output spaces . Both methods do not require alignment of names in different languages -their features are computed directly from the names themselves . We perform an experimental evaluation of the methods for name transliteration from three languages ( Arabic , Korean , and Russian ) into English , and compare the methods experimentally to a state-of-theart joint probabilistic modeling approach . We find that the discriminative methods outperform probabilistic modeling , with the global discriminative modeling approach achieving the best performance in all languages . Name transliteration is an important task of transcribing a name from alphabet to another . For example , an Arabic ‫,"وﻟﻴﺎم"‬ Korean " 윌리엄 " , and Russian " Вильям " all correspond to English " William " . We address the problem of transliteration in the general setting : it involves trying to recover original English names from their transcription in a foreign language , as well as finding an acceptable spelling of a foreign name in English . We apply name transliteration in the context of cross-lingual information extraction . Name extractors are currently available in multiple languages . Our goal is to make the extracted names understandable to monolingual English speakers by transliterating the names into English . The extraction context of the transliteration application imposes additional complexity constraints on the task . In particular , we aim for the transliteration speed to be comparable to that of extraction speed . Since most current extraction systems are fairly fast ( > 1 Gb of text per hour ) , the complexity requirement reduces the range of techniques applicable to the transliteration . More precisely , we can not use WWW and the web count information to hone in on the right transliteration candidate . Instead , all relevant transliteration information has to be represented within a compact and self-contained transliteration model . We present two methods for creating and applying transliteration models . In contrast to most previous transliteration approaches , our models are discriminative . Using an existing transliteration dictionary D ( a set of name pairs { ( f , e ) } ) , we learn a function that directly maps a name f from one language into a name e in another language . We do not estimate either direct conditional p(e|f ) or reverse conditional p(f|e ) or joint p(e , f ) probability models . Furthermore , we do away with the notion of alignment : our transliteration model does not require and is not defined of in terms of aligned e and f. Instead , all features used by the model are computed directly from the names f and e without any need for their alignment . The two discriminative methods that we present correspond to local and global modeling paradigms for solving complex learning problems with structured output spaces . In the local setting , we learn linear classifiers that predict a letter e i from the previously predicted letters e 1 … e i-1 and the original name f. In the global setting , we learn a function W mapping a pair ( f , e ) into a score W(f , e)∈ R. The function W is linear in features computed from the pair ( f , e ) . We describe the pertinent feature spaces as well as pre-sent both training and decoding algorithms for the local and global settings . We perform an experimental evaluation for three language pairs ( transliteration from Arabic , Korean , and Russian into English ) comparing our methods to a joint probabilistic modeling approach to transliteration , which was shown to deliver superior performance . We show experimentally that both discriminative methods outperform the probabilistic approach , with global discriminative modeling achieving the best performance in all languages . We presented two novel discriminative approaches to name transliteration that do not employ the notion of alignment . We showed experimentally that the approaches lead to superior experimental results in all languages , with the global discriminative modeling approach achieving the best performance . The results are somewhat surprising , for the notion of alignment seems very intuitive and useful for transliteration . We will investigate whether similar alignment-free methodology can be extended to full-text translation . It will also be interesting to study the relationship between our discriminative alignment-free methods and recently proposed discriminative alignment-based methods for transliteration and translation ( Taskar et al . 2005a ; Moore 2005 ) . We also showed that for name transliteration , global discriminative modeling is superior to local classifier-based discriminative modeling . This may have resulted from poor calibration of scores and probabilities produced by individual classifiers . We plan to further investigate the relationship between the local and global approaches to complex learning problems in natural language .

The name transliteration task aims to transcribe extracted names into English, and since current extraction systems are fairly fast, applicable techniques for transliteration are limited.

They present two discriminative methods that learn a map function from one language into another using a dictionary without the notion of alignment.

The proposed methods outperform state-of-the-art probabilistic models on name transliteration from Arabic, Korean, and Russian to English, and the global discriminative modelling performs the best.

2022.acl-long.393

Motivated by the success of T5 ( Text-To-Text Transfer Transformer ) in pre-trained natural language processing models , we propose a unified-modal SpeechT5 framework that explores the encoder-decoder pre-training for self-supervised speech / text representation learning . The SpeechT5 framework consists of a shared encoder-decoder network and six modal-specific ( speech / text ) pre / post-nets . After preprocessing the input speech / text through the pre-nets , the shared encoder-decoder network models the sequence-to-sequence transformation , and then the post-nets generate the output in the speech / text modality based on the output of the decoder . Leveraging large-scale unlabeled speech and text data , we pre-train SpeechT5 to learn a unified-modal representation , hoping to improve the modeling capability for both speech and text . To align the textual and speech information into this unified semantic space , we propose a cross-modal vector quantization approach that randomly mixes up speech / text states with latent units as the interface between encoder and decoder . Extensive evaluations show the superiority of the proposed SpeechT5 framework on a wide variety of spoken language processing tasks , including automatic speech recognition , speech synthesis , speech translation , voice conversion , speech enhancement , and speaker identification . We release our code and model at https://github.com / microsoft/ SpeechT5 . Starting with ELMo ( Peters et al . , 2018 ) and BERT ( Devlin et al . , 2019 ) , substantial work has shown that pre-trained models can significantly improve in various natural language processing ( NLP ) tasks Figure 1 : An illustration of the SpeechT5 framework , which treats spoken language processing tasks as a speech / text to speech / text format , including automatic speech recognition ( ASR ) , speech translation ( ST ) , speech identification ( SID ) , text to speech ( TTS ) , voice conversion ( VC ) , and speech enhancement ( SE ) . ( Radford et al . , 2019 ; CONNEAU and Lample , 2019 ; Yang et al . , 2019 ; Dong et al . , 2019 ; Lewis et al . , 2020 ) . Following the pre-training techniques in NLP , self-supervised speech representation learning has also been investigated and shown promising results , benefiting from richly learned representations ( Chung and Glass , 2018 ; Chuang et al . , 2020 ; Song et al . , 2019 ; Baevski et al . , 2020 ; Wang et al . , 2021 ; Hsu et al . , 2021 ; Chung et al . , 2021a ) , such as wav2vec 2.0 ( Baevski et al . , 2020 ) and HuBERT ( Hsu et al . , 2021 ) . However , previous speech pre-training work suffers from two problems : ( 1 ) most of them learn the speech representation with only unlabeled speech data but ignore the importance of textual data to spoken language tasks ( e.g. , automatic speech recognition ) which require the modality transformation ; ( 2 ) most of these models solely rely on a pre-trained speech encoder for various downstream tasks , leaving the decoder not pre-trained for the sequence-to-sequence generation tasks . How to design a unified encoder-decoder model that can take advantage of both unlabeled speech and text data to improve various spoken language processing tasks is not well explored . Inspired by the T5 method ( Raffel et al . , 2020 ) , we attempt to formulate each spoken language processing task as a speech / text to speech / text problem via an encoder-decoder framework , which enables us to use the same pre-trained model with bimodal data across diverse tasks , as shown in Figure 1 . To achieve this , we propose a unifiedmodal pre-training framework , SpeechT5 , containing an encoder-decoder backbone network and modal-specific pre / post-nets . With the pre-nets , the input speech / text is embedded in a shared space , and the encoder-decoder backbone network models the sequence-to-sequence conversion , from which the model-specific post-nets generate the speech / text output . Particularly , SpeechT5 is mainly pre-trained with a denoising sequence-tosequence method by leveraging large-scale unlabeled text and speech corpus . To align the textual and acoustic information into a unified semantic space , the proposed SpeechT5 model ( 1 ) maps text and speech representations into a shared vector quantization space , and ( 2 ) randomly mixes up the quantized latent representations and the contextual states , which can better guide the quantizer to learn the cross-modal features . We fine-tune SpeechT5 on a wide variety of downstream spoken language processing tasks , including automatic speech recognition ( ASR ) , textto-speech ( TTS ) , speech translation ( ST ) , voice conversion ( VC ) , speech enhancement ( SE ) , and speaker identification ( SID ) . Massive experiments show that the proposed SpeechT5 model achieves a significant improvement on these spoken language processing tasks compared with the state-of-theart baselines . Specifically , the proposed SpeechT5 outperforms wav2vec 2.0 ( Baevski et al . , 2020 ) and HuBERT ( Hsu et al . , 2021 ) with the BASE model on the ASR task and also performs better than the state-of-the-art voice Transformer network ( Huang et al . , 2021 ) on the VC task . Besides , SpeechT5 is significantly superior to SpeechNet ( Chen et al . , 2021b ) and pre-trained models from SUPERB ( Yang et al . , 2021 ) and achieves the stateof-the-art performance ( i.e. , 96.49 % ) on the SID task . We further provide an empirical comparison of the pre-training tasks and modules , and the ablation study demonstrates the effectiveness of the proposed joint speech-text pre-training method . The contributions of this paper are summarized as follows . • To the best of our knowledge , this is the first work to investigate a unified encoder-decoder framework for various spoken language processing tasks . • We propose a cross-modal vector quantization approach , which learns the implicit alignment between acoustic and textual representation with large-scale unlabeled speech and text data . • Extensive experiments on spoken language processing tasks demonstrate the effectiveness and superiority of the proposed SpeechT5 model . In this paper , we have proposed SpeechT5 as a pretrained encoder-decoder model for various spoken language processing tasks . We convert all spoken language processing tasks into a speech / text to speech / text format and propose a novel joint pretraining method to utilize cross-modal information by leveraging the unlabeled speech and text data . The proposed unified encoder-decoder model can support generation tasks such as speech translation and voice conversion . Massive experiments show that SpeechT5 significantly outperforms all baselines in several spoken language processing tasks . In the future , we are going to pre-train the SpeechT5 with a larger model and more unlabeled data . We are also interested in extending the proposed SpeechT5 framework to address multilingual spoken language processing tasks for future work .

Existing speech pre-training methods ignore the importance of textual data and solely depend on encoders leaving the decoder out of pre-training for generation tasks.

They propose a unified-modal encoder-decoder framework with shared and modal-specific networks for self-supervised speech and text representation learning by using unlabeled text and speech corpus.

The fine-tuned proposed model is evaluated on a variety of spoken processing tasks and outperforms state-of-the-art models on voice conversion speaker identification tasks.

D13-1158

Recent studies on extractive text summarization formulate it as a combinatorial optimization problem such as a Knapsack Problem , a Maximum Coverage Problem or a Budgeted Median Problem . These methods successfully improved summarization quality , but they did not consider the rhetorical relations between the textual units of a source document . Thus , summaries generated by these methods may lack logical coherence . This paper proposes a single document summarization method based on the trimming of a discourse tree . This is a two-fold process . First , we propose rules for transforming a rhetorical structure theorybased discourse tree into a dependency-based discourse tree , which allows us to take a treetrimming approach to summarization . Second , we formulate the problem of trimming a dependency-based discourse tree as a Tree Knapsack Problem , then solve it with integer linear programming ( ILP ) . Evaluation results showed that our method improved ROUGE scores . State-of-the-art extractive text summarization methods regard a document ( or a document set ) as a set of textual units ( e.g. sentences , clauses , phrases ) and formulate summarization as a combinatorial optimization problem , i.e. selecting a subset of the set of textual units that maximizes an objective without violating a length constraint . For example , Mc-Donald ( 2007 ) formulated text summarization as a Knapsack Problem , where he selects a set of textual units that maximize the sum of significance scores of each unit . Filatova et al . ( 2004 ) proposed a summarization method based on a Maximum Coverage Problem , in which they select a set of textual units that maximizes the weighted sum of the conceptual units ( e.g. unigrams ) contained in the set . Although , their greedy solution is only an approximation , Takamura et al . ( 2009a ) extended it to obtain the exact solution . More recently , Takamura et al . ( 2009b ) regarded summarization as a Budgeted Median Problem and obtain exact solutions with integer linear programming . These methods successfully improved ROUGE ( Lin , 2004 ) scores , but they still have one critical shortcoming . Since these methods are based on subset selection , the summaries they generate can not preserve the rhetorical structure of the textual units of a source document . Thus , the resulting summary may lack coherence and may not include significant textual units from a source document . One powerful and potential way to overcome the problem is to include discourse tree constraints in the summarization procedure . Marcu ( 1998 ) regarded a document as a Rhetorical Structure Theory ( RST ) ( William Charles , Mann and Sandra Annear , Thompson , 1988 ) -based discourse tree ( RST-DT ) and selected textual units according to a preference ranking derived from the tree structure to make a summary . Daumé et al . ( 2002 ) proposed a document compression method that directly models the probability of a summary given an RST-DT by using a noisy-channel model . These methods generate well-organized summaries , however , since they do not formulate summarizations as combinatorial op-Figure 1 : Example RST-DT from ( Marcu , 1998 ) . timization problems , the optimality of the generated summaries is not guaranteed . In this paper , we propose a single document summarization method based on the trimming of a discourse tree based on the Tree Knapsack Problem . If a discourse tree explicitly represents parent-child relationships between textual units , we can apply the well-known tree-trimming approach to a discourse tree and reap the benefit of combinatorial optimization methods . In other words , to apply the treetrimming approach , we need a tree whose all nodes represent textual units . Unfortunately , the RST-DT does not allow it , because textual units in the RST-DT are located only on leaf nodes and parent-child relationship between textual units are represented implicitly at higher positions in a tree . Therefore , we first propose rules that transform an RST-DT into a dependency-based discourse tree ( DEP-DT ) that explicitly defines the parent-child relationships . Second , we treat it as a rooted subtree selection , in other words , a Tree Knapsack Problem and formulate the problem as an ILP . 2 From RST-DT to DEP-DT This paper proposed rules for transforming an RST-DT to a DEP-DT to obtain the parent-child relationships between EDUs . We treated a single document summarization method as a Tree Knapsack Problem , i.e. the summarizer selects the best rooted subtree from a DEP-DT . To demonstrate the effectiveness of our method , we conducted an experimental evaluation using 30 documents selected from the RST Discourse Treebank Corpus . The results showed that our method achieved the highest ROUGE-1,2 scores .

Existing optimization-based methods for extractive summarization do not consider the rhetorical relations between textual units leading to generating uncoherent summaries or missing significant textual units.

They propose to first transform a rhetorical discourse tree into a dependency-based tree and then trim it as a Tree Knapsack Problem.

The proposed method achieves the highest ROUGE-1,2 scores on 30 documents selected from the RST Discourse Treebank Corpus.

D19-1098

Pre-training Transformer from large-scale raw texts and fine-tuning on the desired task have achieved state-of-the-art results on diverse NLP tasks . However , it is unclear what the learned attention captures . The attention computed by attention heads seems not to match human intuitions about hierarchical structures . This paper proposes Tree Transformer , which adds an extra constraint to attention heads of the bidirectional Transformer encoder in order to encourage the attention heads to follow tree structures . The tree structures can be automatically induced from raw texts by our proposed " Constituent Attention " module , which is simply implemented by self-attention between two adjacent words . With the same training procedure identical to BERT , the experiments demonstrate the effectiveness of Tree Transformer in terms of inducing tree structures , better language modeling , and further learning more explainable attention scores 1 . Human languages exhibit a rich hierarchical structure which is currently not exploited nor mirrored by the self-attention mechanism that is the core of the now popular Transformer architecture . Prior work that integrated hierarchical structure into neural networks either used recursive neural networks ( Tree-RNNs ) ( C. Goller and A.Kuchler , 1996 ; Socher et al . , 2011 ; Tai et al . , 2015 ) or simultaneously generated a syntax tree and language in RNN ( Dyer et al . , 2016 ) , which have shown beneficial for many downstream tasks ( Aharoni and Goldberg , 2017 ; Eriguchi et al . , 2017 ; Strubell et al . , 2018 ; Zaremoodi and Haffari , 2018 ) . Considering the requirement of the annotated parse trees and the 1 The source code is publicly available at https:// github.com / yaushian / Tree-Transformer . costly annotation effort , most prior work relied on the supervised syntactic parser . However , a supervised parser may be unavailable when the language is low-resourced or the target data has different distribution from the source domain . Therefore , the task of learning latent tree structures without human-annotated data , called grammar induction ( Carroll and Charniak , 1992 ; Klein and Manning , 2002 ; Smith and Eisner , 2005 ) , has become an important problem and attractd more attention from researchers recently . Prior work mainly focused on inducing tree structures from recurrent neural networks ( Shen et al . , 2018a , b ) or recursive neural networks ( Yogatama et al . , 2017 ; Drozdov et al . , 2019 ) , while integrating tree structures into Transformer remains an unexplored direction . Pre-training Transformer from large-scale raw texts successfully learns high-quality language representations . By further fine-tuning pre-trained Transformer on desired tasks , wide range of NLP tasks obtain the state-of-the-art results ( Radford et al . , 2019 ; Devlin et al . , 2018 ; Dong et al . , 2019 ) . However , what pre-trained Transformer self-attention heads capture remains unknown . Although an attention can be easily explained by observing how words attend to each other , only some distinct patterns such as attending previous words or named entities can be found informative ( Vig , 2019 ) . The attention matrices do not match our intuitions about hierarchical structures . In order to make the attention learned by Transformer more interpretable and allow Transformer to comprehend language hierarchically , we propose Tree Transformer , which integrates tree structures into bidirectional Transformer encoder . At each layer , words are constrained to attend to other words in the same constituents . This constraint has been proven to be effective in prior work ( Wu et al . , 2018 ) . Different from the prior work that required a supervised parser , in Tree Transformer , the constituency tree structures is automatically induced from raw texts by our proposed " Constituent Attention " module , which is simply implemented by self-attention . Motivated by Tree-RNNs , which compose each phrase and the sentence representation from its constituent sub-phrases , Tree Transformer gradually attaches several smaller constituents into larger ones from lower layers to higher layers . The contributions of this paper are 3-fold : • Our proposed Tree Transformer is easy to implement , which simply inserts an additional " Constituent Attention " module implemented by self-attention to the original Transformer encoder , and achieves good performance on the unsupervised parsing task . • As the induced tree structures guide words to compose the meaning of longer phrases hierarchically , Tree Transformer improves the perplexity on masked language modeling compared to the original Transformer . • The behavior of attention heads learned by Tree Transformer expresses better interpretability , because they are constrained to follow the induced tree structures . By visualizing the self-attention matrices , our model provides the information that better matchs the human intuition about hierarchical structures than the original Transformer . This paper proposes Tree Transformer , a first attempt of integrating tree structures into Transformer by constraining the attention heads to attend within constituents . The tree structures are automatically induced from the raw texts by our proposed Constituent Attention module , which attaches the constituents to each other by selfattention . The performance on unsupervised parsing demonstrates the effectiveness of our model in terms of inducing tree structures coherent to human expert annotations . We believe that incorporating tree structures into Transformer is an important and worth exploring direction , because it allows Transformer to learn more interpretable attention heads and achieve better language modeling . The interpretable attention can better explain how the model processes the natural language and guide the future improvement .

This is unclear what attention heads in pre-training transformers models capture and it seems not to match human intuitions about hierarchical structures.

They propose to add an extra constraint to attention heads of the bidirectional Transformer encoder and a module induces tree structures from raw texts.

The proposed model achieves better unsupervised tree structure induction, language modelling, and more explainable attention scores which are coherent to human expert annotations.

D09-1131

This paper employs morphological structures and relations between sentence segments for opinion analysis on words and sentences . Chinese words are classified into eight morphological types by two proposed classifiers , CRF classifier and SVM classifier . Experiments show that the injection of morphological information improves the performance of the word polarity detection . To utilize syntactic structures , we annotate structural trios to represent relations between sentence segments . Experiments show that considering structural trios is useful for sentence opinion analysis . The best f-score achieves 0.77 for opinion word extraction , 0.62 for opinion word polarity detection , 0.80 for opinion sentence extraction , and 0.54 for opinion sentence polarity detection . Sentiment analysis has attracted much attention in recent years because a large scale of subjective information is disseminated through various platforms on the web . Sentiment information can be applied to a wide variety of fields , including product recommendation , review summarization , public polling , and so on . Opinion dictionaries are important resources for identifying subjective information . Several approaches were proposed to collect such resources . Wiebe ( 2000 ) learned subjective adjectives from corpora . Takamura et al . ( 2005 ) extracted semantic orientations of words . Ku et al . ( 2007 ) measured sentiment degrees of Chinese words by averaging the sentiment scores of the composing characters . When the opinion words are available , the polarities of sentences and documents can be determined by them . Riloff and Wiebe ( 2003 ) learned the extraction patterns for subjective expressions . Kim and Hovy ( 2004 ) found the polarity of subjective expressions . Pang et al . ( 2002 ) and Dave et al . ( 2003 ) explored various techniques at document level . Morphological information has been widely used in classifying words , telling the meanings , and doing other in-depth analysis ( Tzeng and Chen , 2002 ) . However , morphological information was seldom applied either in Chinese opinion extraction , or in solving the coverage problem of opinion dictionary . Instead of bag-ofcharacters approach ( Ku et al . , 2007 ) , this paper employs morphological structures of words to extract opinion words . Relations between sentence segments are also defined by linguistics in the Chinese language . These are similar to morphological structures between Chinese characters . Based on parsing trees of sentences , we identify these relations and utilize them for opinion analysis on sentences . As the experimental corpus , some researchers managed to generate annotated materials and gold standards under many constraints . Ku set a standard for generating final answers from annotations of multiple annotators ( Ku et al . , 2007 ) , and Somasundaran annotated discourse information from meeting dialogs to train a sentiment model ( Somasundaran et al . , 2007 ) . For multilingual issues , researchers concerned mainly about the applicability of corpus and algorithms from the native language to foreign languages ( Banea et al . , 2008 ; Bautin et al . , 2008 ) . Several opinion analysis systems have been developed so far . OASYS ( Cesarano et al . , 2007 ) and CopeOpi ( Ku et al . , 2007 ) allow users input their queries and select preferred data sources , and then track opinions in a time zone . For both systems , extracting opinions is the main focus , while holders and targets are identified implicitly when retrieving relevant documents . Carenini 's team proposed a graphical user interface for evaluative texts ( 2006 ) , in which color blocks were used to present the evaluations for components of products . Fair News Reader , a Japanese news Web system , incorporates sentiment information insensibly in an interesting way ( Kawai et al . , 2007 ) . It provides readers " balanced " reports by analyzing the sentiment in news articles which readers have read , and suggests them new articles according to the analysis results . It leads the application of opinion analysis to the direction of personalization . This paper considers morphological and syntactic structures in analyzing Chinese opinion words and sentences . For morphological structures , eight Chinese morphological types are defined . CRF classifier and SVM classifier for morphological type classification are proposed . Experiments show that CRF classifier achieves the best accuracy 0.70 in type classification , which is 8 % better than SVM classifier . We further show that word morphological structures benefit the opinion word extraction significantly . With the help of the sentiment dictionary NTUSD , the f-score of opinion word extraction achieves 0.77 and the f-score of the word polarity detection achieves 0.62 when the word morphological types are provided by the SVM classifier . They are comparably better than bag-of-character approach and the dictionary based approach . We defined structural trios to represent the relations between sentence segments and also extract these relations using CRF algorithm . Results show that considering structural trios benefits the opinion analysis on sentences . An fscore 0.80 for opinion extraction and an f-score 0.54 for polarity detection are achieved , which is a great improvement . The opinion scoring functions for morphological types and structural trios are critical for polarity detection , and scoring functions for words determine the scoring functions for sentences . Now we define these functions intuitively based on linguistic rules , but learning methods like regression will be investigated in the future . Examining the interaction of cues from word and sentence levels on the opinion sentence extraction and the opinion polarity detection is our next goal .

There is not much work on applying morphological information in opinion extraction in Chinese.

They propose to utilize morphological and syntactic features for Chinese opinion analysis on word and sentence levels.

They show that using morphological structures helps opinion analysis in Chinese, outperforming the existing bat-of-character approach and the dictionary-based approach.

P19-1252

In this paper , we investigate the importance of social network information compared to content information in the prediction of a Twitter user 's occupational class . We show that the content information of a user 's tweets , the profile descriptions of a user 's follower / following community , and the user 's social network provide useful information for classifying a user 's occupational group . In our study , we extend an existing dataset for this problem , and we achieve significantly better performance by using social network homophily that has not been fully exploited in previous work . In our analysis , we found that by using the graph convolutional network to exploit social homophily , we can achieve competitive performance on this dataset with just a small fraction of the training data . * Equal Contribution ; work performed while both authors were visiting Singapore University of Technology and Design ( SUTD ) . T1 : Day at the races . T2 : The new pitch era starts here . Groundsman . Wolves fan . Horse racing enthusiast . A Commercial and Domestic Grounds Maintenance company T1 : Pitch at wembley looks great . T2 : Spurs will have best stadium in uk . Twitter ( http://twitter.com ) is a microblogging service launched in 2006 , where , a user can publish messages with up to 280 characters , called " tweets " . Unlike many other social networking platforms , such as Facebook and LinkedIn , Twitter does not provide structured fields for users to fill in personal information . However , a user can write a 160-character-long small public summary about itself called a " Bio " . Besides linguistic information from tweets and Bios , online social media is a rich source of network information . People 's personal networks are homogeneous , i.e. , friends share more attributes such as race , ethnicity , religion , and occupation-known as the homophily principle ( McPherson et al . , 2001 ) . Such network information has been utilized in friend recommendation ( Guy et al . , 2010 ) , community detection Follow Figure 1 : User and Network information on Twitter Microblog . ( Yang and Leskovec , 2013 ) , etc . Figure 1 shows two users connected on Twitter . By looking at their Bio and tweets , it can be inferred that these users share the same occupational interest . Profiling users can enhance service quality and improve product recommendation , and hence is a widely studied problem . User occupational class prediction is an important component of user profiling and a sub-task of user demographic feature prediction . Existing approaches to predicting Twitter users ' demographic attributes explore , select , and combine various features generated from text and network to achieve the best predictive performances in respective classification tasks ( Han et al . , 2013 ; Miller et al . , 2012 ; Preot ¸iuc-Pietro et al . , 2015 ; Huang et al . , 2015 ; Aletras and Chamberlain , 2018 ) . The three categories of features are : account level features , tweet text features , and network based features . Past research have shown the distinctive usage of language across gender , age , location , etc . in tweets ( Sloan et al . , 2015 ; Cheng et al . , 2010 ; Burger et al . , 2011 ; Rao et al . , 2010 ) , which makes content based prediction effective . As for user occupational class prediction , Preot ¸iuc-Pietro et al . ( 2015 ) built a dataset where users are assigned to hierarchical job categories . They used word cluster distribution features of content information to predict a user 's occupational group . Aletras and Chamberlain ( 2018 ) constructed a user 's followings connections to learn the user embedding as a feature input to the classification models . Considering the regional disparities of economic development stages , the major job categories may vary significantly across regions . Sloan et al . ( 2015 ) summarized occupation distribution of Twitter users in the UK by looking into their profiles . In this paper , we analyze the usefulness of a user 's network information over the user 's tweets for predicting its occupational group . We extend the existing dataset for occupation classification ( Preot ¸iuc-Pietro et al . ( 2015 ) ) by introducing the network information about a user , i.e. follower / following IDs together with their Bio descriptions , and we construct a user-centric network to extract useful community and text based features . The acquired features from the network are then exploited using a graph neural network . The obtained results show the importance of a network information over tweet information from a user for such a task . Previous works have used tweets or a fraction of the network information to extract features for occupation classification . To analyze the importance of network information , we extended an existing Twitter dataset for a user 's social media connections ( follow information ) . We showed that by using only follow information as an input to graph convolutional networks , one can achieve a significantly higher accuracy on the prediction task as compared to the existing approaches utilizing tweet-only information or partial network structure . Directions of future research include adaptation of our methods to a large scale , sparsely connected social network . One might also want to investigate the inductive settings of GCN ( Hamilton et al . , 2017 ) to predict demographic information of a user from outside the black network .

Existing systems only use limited information from the tweets network to perform occupation classification.

They extend existing Twitter occupation classification graph-based models to exploit content information by adding textual data to existing datasets.

They show that textual feature enables graph neural networks to predict Twitter user occupation well even with a limited amount of training data.

P06-1073

Short vowels and other diacritics are not part of written Arabic scripts . Exceptions are made for important political and religious texts and in scripts for beginning students of Arabic . Script without diacritics have considerable ambiguity because many words with different diacritic patterns appear identical in a diacritic-less setting . We propose in this paper a maximum entropy approach for restoring diacritics in a document . The approach can easily integrate and make effective use of diverse types of information ; the model we propose integrates a wide array of lexical , segmentbased and part-of-speech tag features . The combination of these feature types leads to a state-of-the-art diacritization model . Using a publicly available corpus ( LDC 's Arabic Treebank Part 3 ) , we achieve a diacritic error rate of 5.1 % , a segment error rate 8.5 % , and a word error rate of 17.3 % . In case-ending-less setting , we obtain a diacritic error rate of 2.2 % , a segment error rate 4.0 % , and a word error rate of 7.2 % . Modern Arabic written texts are composed of scripts without short vowels and other diacritic marks . This often leads to considerable ambiguity since several words that have different diacritic patterns may appear identical in a diacritic-less setting . Educated modern Arabic speakers are able to accurately restore diacritics in a document . This is based on the context and their knowledge of the grammar and the lexicon of Arabic . However , a text without diacritics becomes a source of confusion for beginning readers and people with learning disabilities . A text without diacritics is also problematic for applications such as text-to-speech or speech-to-text , where the lack of diacritics adds another layer of ambiguity when processing the data . As an example , full vocalization of text is required for text-to-speech applications , where the mapping from graphemes to phonemes is simple compared to languages such as English and French ; where there is , in most cases , one-to-one relationship . Also , using data with diacritics shows an improvement in the accuracy of speech-recognition applications ( Afify et al . , 2004 ) . Currently , text-tospeech , speech-to-text , and other applications use data where diacritics are placed manually , which is a tedious and time consuming excercise . A diacritization system that restores the diacritics of scripts , i.e. supply the full diacritical markings , would be of interest to these applications . It also would greatly benefit nonnative speakers , sufferers of dyslexia and could assist in restoring diacritics of children 's and poetry books , a task that is currently done manually . We propose in this paper a statistical approach that restores diacritics in a text document . The proposed approach is based on the maximum entropy framework where several diverse sources of information are employed . The model implicitly learns the correlation between these types of information and the output diacritics . In the next section , we present the set of diacritics to be restored and the ambiguity we face when processing a non-diacritized text . Section 3 gives a brief summary of previous related works . Section 4 presents our diacritization model ; we explain the training and decoding process as well as the different feature categories employed to restore the diacritics . Section 5 describes a clearly defined and replicable split of the LDC 's Arabic Treebank Part 3 corpus , used to built and evaluate the system , so that the reproduction of the results and future comparison can accurately be established . Section 6 presents the experimental results . Section 7 reports a comparison of our approach to the finite state machine modeling technique that showed promissing results in ( Nelken and Shieber , 2005 ) . Finally , section 8 concludes the paper and discusses future directions . We presented in this paper a statistical model for Arabic diacritic restoration . The approach we propose is based on the Maximum entropy framework , which gives the system the ability to integrate different sources of knowledge . Our model has the advantage of successfully combining diverse sources of information ranging from lexical , segment-based and POS features . Both POS and segment-based features are generated by separate statistical systems -not extracted manually -in order to simulate real world applications . The segment-based features are extracted from a statistical morphological analysis system using WFST approach and the POS features are generated by a parsing model Table 5 : Error Rate in % for n-gram diacritic restoration and segmentation using FST and Kneser-Ney LM . Columns marked with " True shadda " represent results on documents containing the original consonant doubling " shadda " while columns marked with " Predicted shadda " represent results where the system restored all diacritics including shadda . that also uses Maximum entropy framework . Evaluation results show that combining these sources of information lead to state-of-the-art performance . As future work , we plan to incorporate Buckwalter morphological analyzer information to extract new features that reduce the search space . One idea will be to reduce the search to the number of hypotheses , if any , proposed by the morphological analyzer . We also plan to investigate additional conjunction features to improve the accuracy of the model .

Short vowels and other diacritics are not expressed in written Arabic, making it difficult to read for beginner readers or system developments.

They propose an approach that uses maximum entropy to restore diacritics in Arabic documents by learning relations between a wide varieties of features.

They show that by taking various kinds of features their system outperforms the existing state-of-the-art decritization model.

2021.eacl-main.251

Current state-of-the-art systems for joint entity relation extraction ( Luan et al . , 2019 ; Wadden et al . , 2019 ) usually adopt the multi-task learning framework . However , annotations for these additional tasks such as coreference resolution and event extraction are always equally hard ( or even harder ) to obtain . In this work , we propose a pre-training method ENPAR to improve the joint extraction performance . EN-PAR requires only the additional entity annotations that are much easier to collect . Unlike most existing works that only consider incorporating entity information into the sentence encoder , we further utilize the entity pair information . Specifically , we devise four novel objectives , i.e. , masked entity typing , masked entity prediction , adversarial context discrimination , and permutation prediction , to pretrain an entity encoder and an entity pair encoder . Comprehensive experiments show that the proposed pre-training method achieves significant improvement over BERT on ACE05 , SciERC , and NYT , and outperforms current state-of-the-art on ACE05 . Joint extraction of entities and relations is a fundamental task in information extraction , it aims to extract entities and relations with a unified model . Current approaches ( Luan et al . , 2019 ; Wadden et al . , 2019 ) usually adopt the multi-task learning framework that optimizes many objectives simultaneously , including entity recognition , relation extraction , coreference resolution , and event extraction . However , as large-scale manually labeled data required by these methods is unavailable in many domains , their applicability is severely restricted . Therefore , we expect to catch or even surpass the multi-task based joint models with less annotation cost . Compared with the annotations of coreference resolution and event extraction , entity annotations can be easily obtained through automatic NER annotation tools ( e.g. , spaCy1 ) . In this paper , we focus on improving the model 's performance with just extra entity annotations . Although pre-trained models , like BERT , have shown impressive performance in many downstream tasks , they have mainly two limitations when applied in the joint entity relation extraction task . One is that currently pre-trained objectives are insufficient for this task . Specifically , these commonly used universal pre-trained model ( e.g. , BERT ) do not consider the entity-related knowledge that is crucial for better extracting entities and relations . The other is that these models only provide pre-trained representations for tokens and sentences , but not entities and entity pairs . To obtain the representations for entities and entity pairs , additional parameters that are not pre-trained are introduced in the fine-tuning stage , which may futher impair the joint extraction performance . To address the first limitation , recent several works try to incorporate entity-related information into pre-training objectives . Zhang et al . ( 2019 ) fuses heterogeneous information from both texts and knowledge graphs and proposes a denoising entity auto-encoder objective based on BERT . Sun et al . ( 2019c ) presents two knowledge masking strategies in the pre-training stage ( entity-level masking and phrase-level masking ) . Both of them utilize extra entity annotations ( i.e. , entities in knowledge graphs and automatic entity annotations , respectively ) . In this paper , we follow this line of works and build a large-scale entity annotated corpus using the spaCy NER tool . For the second limitation , we propose ENPAR , a pre-training method customized for entity relation extraction . ENPAR consists of an underlying sentence encoder , an entity encoder , and an entity pair encoder . Compared with BERT ( Figure 1 ( a ) ) , the proposed entity encoder and entity pair encoder directly provide representations of entities and entity pairs . To train the three encoders , we devise four novel pre-training tasks : masked entity typing , masked entity prediction , adversarial context discrimination and permutation prediction ( Figure 1(b ) ) . In the first two tasks , we randomly mask some entity words and then predict the masked tokens and the entity type . These two tasks are natural extensions of the masked language model . To learn a better entity pair encoder , we draw inspirations from the denoising auto-encoder ( Zhang et al . , 2019 ) and propose the last two tasks . Specifically , when the entity pair or its context in a sentence are perturbed , we hope that the entity pair encoder is capable of tracking such changes . We employ the parameter-sharing method for these four tasks and train these objectives jointly . To sum up , our main contributions are as follows2 : • We introduce an entity encoder and an entity pair encoder to incorporate not only the entity information but also the entity pair information , which were ignored in current universal pre-trained models . • We propose four novel pre-training tasks that help to learn the proposed encoders . These tasks only require additional entity annotations ( with commonly used entity types ) , which can be automatically generated by public annotation tools , such as spaCy NER . • We conduct comprehensive experiments and demonstrate that the proposed method achieves significant improvement on ACE05 and NYT dataset and is comparable with the state-of-the-art on the SciERC dataset . We propose ENPAR , a pre-training method customized for entity relation extraction only with additional entity annotations . Instead of only pretraining sentence encoder in universal pre-trained models , we also pre-train an entity encoder and an entity pair encoder . Then the proposed four objec-tives can incorporate entity and entity-pair knowledge into the pre-trained encoders to enhance the encoders ' representations . Experiments on three datasets demonstrate that ENPAR achieves comparable or even superior performances compared with multi-task based joint models .

Current joint entity relation extraction models follow a multitask learning setup however datasets with multiple types of annotation are not available for many domains.

They propose to pre-train a language model entity relation extraction with four newly introduced objective functions which utilize automatically obtained annotations by NER models.

The models pre-trained by the proposed method significantly outperform BERT and current state-of-the-art models on three entity relation extraction datasets.

P16-1089

We present the Siamese Continuous Bag of Words ( Siamese CBOW ) model , a neural network for efficient estimation of highquality sentence embeddings . Averaging the embeddings of words in a sentence has proven to be a surprisingly successful and efficient way of obtaining sentence embeddings . However , word embeddings trained with the methods currently available are not optimized for the task of sentence representation , and , thus , likely to be suboptimal . Siamese CBOW handles this problem by training word embeddings directly for the purpose of being averaged . The underlying neural network learns word embeddings by predicting , from a sentence representation , its surrounding sentences . We show the robustness of the Siamese CBOW model by evaluating it on 20 datasets stemming from a wide variety of sources . Word embeddings have proven to be beneficial in a variety of tasks in NLP such as machine translation ( Zou et al . , 2013 ) , parsing ( Chen and Manning , 2014 ) , semantic search ( Reinanda et al . , 2015 ; Voskarides et al . , 2015 ) , and tracking the meaning of words and concepts over time ( Kim et al . , 2014 ; Kenter et al . , 2015 ) . It is not evident , however , how word embeddings should be combined to represent larger pieces of text , like sentences , paragraphs or documents . Surprisingly , simply averaging word embeddings of all words in a text has proven to be a strong baseline or feature across a multitude of tasks ( Faruqui et al . , 2014 ; Yu et al . , 2014 ; Gershman and Tenenbaum , 2015 ; Kenter and de Rijke , 2015 ) . Word embeddings , however , are not optimized specifically for representing sentences . In this paper we present a model for obtaining word embeddings that are tailored specifically for the task of averaging them . We do this by directly including a comparison of sentence embeddings-the averaged embeddings of the words they contain-in the cost function of our network . Word embeddings are typically trained in a fast and scalable way from unlabeled training data . As the training data is unlabeled , word embeddings are usually not task-specific . Rather , word embeddings trained on a large training corpus , like the ones from ( Collobert and Weston , 2008 ; Mikolov et al . , 2013b ) are employed across different tasks ( Socher et al . , 2012 ; Kenter and de Rijke , 2015 ; Hu et al . , 2014 ) . These two qualities-(i ) being trainable from large quantities of unlabeled data in a reasonable amount of time , and ( ii ) robust performance across different tasks-are highly desirable and allow word embeddings to be used in many large-scale applications . In this work we aim to optimize word embeddings for sentence representations in the same manner . We want to produce general purpose sentence embeddings that should score robustly across multiple test sets , and we want to leverage large amounts of unlabeled training material . In the word2vec algorithm , Mikolov et al . ( 2013a ) construe a supervised training criterion for obtaining word embeddings from unsupervised data , by predicting , for every word , its surrounding words . We apply this strategy at the sentence level , where we aim to predict a sentence from its adjacent sentences ( Kiros et al . , 2015 ; Hill et al . , 2016 ) . This allows us to use unlabeled training data , which is easy to obtain ; the only restriction is that documents need to be split into sentences and that the order between sentences is preserved . The main research question we address is whether directly optimizing word embeddings for the task of being averaged to produce sentence embeddings leads to word embeddings that are better suited for this task than word2vec does . Therefore , we test the embeddings in an unsupervised learning scenario . We use 20 evaluation sets that stem from a wide variety of sources ( newswire , video descriptions , dictionary descriptions , microblog posts ) . Furthermore , we analyze the time complexity of our method and compare it to our baselines methods . Summarizing , our main contributions are : We have presented Siamese CBOW , a neural network architecture that efficiently learns word embeddings optimized for producing sentence representations . The model is trained using only unla-beled text data . It predicts , from an input sentence representation , the preceding and following sentence . We evaluated the model on 20 test sets and show that in a majority of cases , 14 out of 20 , Siamese CBOW outperforms a word2vec baseline and a baseline based on the recently proposed skip-thought architecture . As further analysis on various choices of parameters show that the method is stable across settings , we conclude that Siamese CBOW provides a robust way of generating high-quality sentence representations . Word and sentence embeddings are ubiquitous and many different ways of using them in supervised tasks have been proposed . It is beyond the scope of this paper to provide a comprehensive analysis of all supervised methods using word or sentence embeddings and the effect Siamese CBOW would have on them . However , it would be interesting to see how Siamese CBOW embeddings would affect results in supervised tasks . Lastly , although we evaluated Siamese CBOW on sentence pairs , there is no theoretical limitation restricting it to sentences . It would be interesting to see how embeddings for larger pieces of texts , such as documents , would perform in document clustering or filtering tasks .

While an average of word embeddings has proven to be successful as sentence-level representations, it is suboptimal because they are not optimized to represent sentences.

They propose to train word embeddings directly for the purpose of being averaged by predicting sounding sentences from a sentence representation using unlabeled data.

Evaluations show that their word embeddings outperform existing methods in 14 out of 20 datasets and they are stable in choice of parameters.

D17-1222

We propose a new framework for abstractive text summarization based on a sequence-to-sequence oriented encoderdecoder model equipped with a deep recurrent generative decoder ( DRGN ) . Latent structure information implied in the target summaries is learned based on a recurrent latent random model for improving the summarization quality . Neural variational inference is employed to address the intractable posterior inference for the recurrent latent variables . Abstractive summaries are generated based on both the generative latent variables and the discriminative deterministic states . Extensive experiments on some benchmark datasets in different languages show that DRGN achieves improvements over the state-ofthe-art methods . Automatic summarization is the process of automatically generating a summary that retains the most important content of the original text document ( Edmundson , 1969 ; Luhn , 1958 ; Nenkova and McKeown , 2012 ) . Different from the common extraction-based and compression-based methods , abstraction-based methods aim at constructing new sentences as summaries , thus they require a deeper understanding of the text and the capability of generating new sentences , which provide an obvious advantage in improving the focus of a summary , reducing the redundancy , and keeping a good compression rate ( Bing et al . , 2015 ; Rush et al . , 2015 ; Nallapati et al . , 2016 ) . * The work described in this paper is supported by a grant from the Grant Council of the Hong Kong Special Administrative Region , China ( Project Code : 14203414 ) . Some previous research works show that human-written summaries are more abstractive ( Jing and McKeown , 2000 ) . Moreover , our investigation reveals that people may naturally follow some inherent structures when they write the abstractive summaries . To illustrate this observation , we show some examples in Figure 1 , which are some top story summaries or headlines from the channel " Technology " of CNN . After analyzing the summaries carefully , we can find some common structures from them , such as " What " , " What-Happened " , " Who Action What " , etc . For example , the summary " Apple sues Qualcomm for nearly $ 1 billion " can be structuralized as " Who ( Apple ) Action ( sues ) What ( Qualcomm ) " . Similarly , the summaries " [ Twitter ] [ fixes ] [ botched @POTUS account transfer ] " , " [ Uber ] [ to pay ] [ $ 20 million ] for misleading drivers " , and " [ Bipartisan bill ] aims to [ reform ] [ H-1B visa system ] " also follow the structure of " Who Action What " . The summary " The emergence of the ' cyber cold war " ' matches with the structure of " What " , and the summary " St. Louis ' public library computers hacked " follows the structure of " What-Happened " . Intuitively , if we can incorporate the latent structure information of summaries into the abstractive summarization model , it will improve the quality of the generated summaries . However , very few existing works specifically consider the latent structure information of summaries in their summarization models . Although a very popular neural network based sequence-to-sequence ( seq2seq ) framework has been proposed to tackle the abstractive summarization problem ( Lopyrev , 2015 ; Rush et al . , 2015 ; Nallapati et al . , 2016 ) , the calculation of the internal decoding states is entirely deterministic . The deterministic transformations in these discriminative models lead to limitations on the representation ability of the latent structure information . Miao and Blunsom ( 2016 ) extended the seq2seq framework and proposed a generative model to capture the latent summary information , but they did not consider the recurrent dependencies in their generative model leading to limited representation ability . To tackle the above mentioned problems , we design a new framework based on sequenceto-sequence oriented encoder-decoder model equipped with a latent structure modeling component . We employ Variational Auto-Encoders ( VAEs ) ( Kingma and Welling , 2013 ; Rezende et al . , 2014 ) as the base model for our generative framework which can handle the inference problem associated with complex generative modeling . However , the standard framework of VAEs is not designed for sequence modeling related tasks . Inspired by ( Chung et al . , 2015 ) , we add historical dependencies on the latent variables of VAEs and propose a deep recurrent generative decoder ( DRGD ) for latent structure modeling . Then the standard discriminative deterministic decoder and the recurrent generative decoder are integrated into a unified decoding framework . The target summaries will be decoded based on both the discriminative deterministic variables and the generative latent structural information . All the neural parameters are learned by back-propagation in an end-to-end training paradigm . The main contributions of our framework are summarized as follows : ( 1 ) We propose a sequence-to-sequence oriented encoder-decoder model equipped with a deep recurrent generative decoder ( DRGD ) to model and learn the latent structure information implied in the target summaries of the training data . Neural variational inference is employed to address the intractable posterior inference for the recurrent latent variables . ( 2 ) Both the generative latent structural information and the discriminative deterministic variables are jointly considered in the generation process of the abstractive summaries . ( 3 ) Experimental results on some benchmark datasets in different languages show that our framework achieves better performance than the state-of-the-art models . We propose a deep recurrent generative decoder ( DRGD ) to improve the abstractive summarization performance . The model is a sequenceto-sequence oriented encoder-decoder framework equipped with a latent structure modeling component . Abstractive summaries are generated based on both the latent variables and the deterministic states . Extensive experiments on benchmark

Although humans follow inherent structures in summary writing, currently there are no abstractive summarization models which take latent structure information and recurrent dependencies into account.

They propose a Variational Auto-Encoder-based sequence-to-sequence oriented encoder-decoder model with a deep recurrent generative decoder which learns latent structure information implied in the target summaries.

The proposed model outperforms the state-of-the-art models on some datasets in different languages.

2021.naacl-main.72

Multi-layer multi-head self-attention mechanism is widely applied in modern neural language models . Attention redundancy has been observed among attention heads but has not been deeply studied in the literature . Using BERT-base model as an example , this paper provides a comprehensive study on attention redundancy which is helpful for model interpretation and model compression . We analyze the attention redundancy with Five-Ws and How . ( What ) We define and focus the study on redundancy matrices generated from pre-trained and fine-tuned BERT-base model for GLUE datasets . ( How ) We use both token-based and sentence-based distance functions to measure the redundancy . ( Where ) Clear and similar redundancy patterns ( cluster structure ) are observed among attention heads . ( When ) Redundancy patterns are similar in both pre-training and fine-tuning phases . ( Who ) We discover that redundancy patterns are task-agnostic . Similar redundancy patterns even exist for randomly generated token sequences . ( " Why " ) We also evaluate influences of the pre-training dropout ratios on attention redundancy . Based on the phaseindependent and task-agnostic attention redundancy patterns , we propose a simple zero-shot pruning method as a case study . Experiments on fine-tuning GLUE tasks verify its effectiveness . The comprehensive analyses on attention redundancy make model understanding and zero-shot model pruning promising . Multi-layer multi-head self-attention architectures ( Transformer ( Vaswani et al . , 2017 ) ) are widely applied in modern language models , such as BERT ( Devlin et al . , 2019 ) , RoBERTa ( Liu et al . , 2019 ) , OpenAI GPT ( Radford et al . , 2018 ) , GPT-2 ( Radford et al . , 2019 ) and ERNIE2.0 ( Sun et al . , 2019 ) , to name a few . Redundancy phenomenon is discovered among attention heads . It demonstrates that many attention heads generate very similar attention matrices ( Clark et al . , 2019 ; Kovaleva et al . , 2019 ) . We take the pre-trained BERT-base model as an example . It learns 12-layer-12-head self-attention matrices describing dependencies between each pair of tokens in a sentence . Then for each token , there are 144 attention vectors . We use Jensen-Shannon distance to measure the relationship between each pair of vectors . Then for one sentence ( consisting of a sequence of tokens ) , the token-averaged distance is utilized to imply the redundancy between each pair of attention matrices . Smaller distance values reflect more redundancy . Figure 1 shows the redundancy ( distance ) among 144 × 144 pairs of attention matrices averaged over 1000 randomly sampled sentences . We can see clear redundancy patterns ( clusters with smaller distance areas ) in consecutive attention layers . Analyzing the attention redundancy helps to interpret the multi-layer multi-head self-attention architecture . Various studies have attempted to re-veal the relationship among attention heads . Examples are attention visualization ( Vig and Belinkov , 2019 ) , common attention patterns ( Kovaleva et al . , 2019 ) , attention head pruning ( Voita et al . , 2019 ) , and probing test ( Clark et al . , 2019 ) . Existing works either focus on the 12 × 12 attention matrices and their effects on ( pre-training or / and finetuning ) performances or focus on linguistic features extracted by latent token vectors and attention matrices . Though the redundancy phenomenon was discovered , no existing work studies the attention redundancy pattern itself ( i.e. , the 144 × 144 distance matrix in Figure 1 ) deeply . This motivates us to conduct a comprehensive and complementary study on the attention redundancy phenomenon . In this paper , we take the BERT-base model as a representative model to analyze the attention redundancy with Five Ws and How . As far as we know , many of the following discoveries are new to the research community . What is attention redundancy ? Given a distance function , we define the pairwise distance matrix ( ∈ R 144×144 ) of the 12 × 12 attention matrices of BERT-base model as attention redundancy matrix . In this paper , we obtain redundancy matrices from both pre-trained and finetuned BERT-base model for GLUE tasks as the research objects . How to measure attention redundancy ? Except for the two token-based measures , Jensen-Shannon distance ( Clark et al . , 2019 ) and cosine similarity ( Kovaleva et al . , 2019 ) used in literature , we employ two more token-based distance function and three sentence-based ones to measure attention redundancy and analyze their similar redundancy patterns ( please refer to Section 4.1 for more details ) . The purpose is to alleviate the measuring bias of just using one distance function . Sentence-based distances directly measure the relationship between two attention matrices without averaging over tokens . We visualize the redundancy patterns using various distance functions . Where does attention redundancy exist ? We find common hierarchical cluster structures in the set of token-based redundancy matrices and the set of sentence-based redundancy matrices , respectively . Attention heads of earlier , middle , and deeper attention layers are clearly clustered in the redundancy matrices . We also demonstrate that highly correlated similar redundancy patterns exist in redundancy matrices generated based on different type of distances . When does attention redundancy occur ? The redundancy is phase-independent . Common redundancy patterns are discovered in both the pre-trained phase and fine-tuned phases . For any downstream task with any distance function , we notice highly correlated attention redundancy patterns between two phases . Who ( which task ) has attention redundancy ? We surprisingly realize that the redundancy is task-agnostic . The redundancy patterns are highly correlated across different tasks . We even randomly generate token sequences as input in the pre-trained BERT-base model . Very similar attention redundancy patterns occur as well . Based on this astonishing discovery , as a case study application , we propose a simple zero-shot head-pruning strategy based on clustering results using redundancy matrices . Compared to other complex pruning strategies , e.g. , ( Tang et al . , 2019 ; Jiao et al . , 2019 ; Fan et al . , 2019 ; Wang et al . , 2019 ; McCarley , 2019 ) , the most important is that without knowing any data of fine-tuning tasks , this pruning can be effectively and efficiently conducted just based on some randomly generated token sequences with the pre-trained BERT-base model . The only effort is to compute one or several redundancy matrices . Results reflect that for most GLUE tasks , the proposed pruning strategy based on redundancy matrices can prune up to 75 % to 85 % of attention heads while keeping comparable fine-tuning performances . " Why " does the phase-independent and taskagnostic attention redundancy happen ? It 's hard to tell the reason of the redundancy patterns ( that 's why we use the quoted " Why " ) . However , we conduct experiments to evaluate the effects on attention redundancy of dropout ratios in the pre-training phase which are suspected as one of reasons ( Clark et al . , 2019 ) . When we use sentence-based distance , a monotonic trend is found . Attention heads tend to be more redundant when increasing dropout ratios . When we use token-based distances , a complex " N"-shape effect exists . We also notice that the redundancy is more sensitive to dropouts in hidden linear transformations than to dropouts in the self-attention mechanism . We believe that above-mentioned new findings in this paper make the redundancy analyses a promis-ing research direction in model interpretation and model compression , and so on . Using BERT-base model as an example , we comprehensively investigated the attention redundancy in multi-layer multi-head self-attention based language models . The redundancy was measured by distance functions at token level and sentence level . At both levels , we found that many heads are not distinct from each other , and clear clustering effects were observed . We discovered that the attention redundancy is phase-independent and taskagnostic . Specifically , compared to a pre-trained model , the redundancy patterns do not change much after fine-tuning on multiple downstream tasks . We also shown complex influences on redundancy of dropout ratios in hidden transformations and self-attention . Based on these discoveries , we design a zero-shot strategy to prune attention heads . Compared to existing methods , the zero-shot pruning is simple and robust ( task-agnostic ) . In the near future , we are interested in experimenting this method over more self-attention based pre-trained language models and more downstream tasks . This is the appendix for NAACL-HLT 2021 paper : Yuchen Bian , Jiaji Huang , Xingyu Cai , Jiahong Yuan , and Kenneth Church . On Attention Redundancy : A Comprehensive Study .

While there are works that report a redundancy among attention heads in modern language models, no works investigate its pattern deeply.

They perform token and sentence level analysis on redundancy matrices from pre-trained and fine-tuned BERT-base models and further propose a pruning method based on findings.

They find that many heads are redundant regardless of phase and task, and show the proposed pruning method can perform robustly.

D17-1220

Comprehending lyrics , as found in songs and poems , can pose a challenge to human and machine readers alike . This motivates the need for systems that can understand the ambiguity and jargon found in such creative texts , and provide commentary to aid readers in reaching the correct interpretation . We introduce the task of automated lyric annotation ( ALA ) . Like text simplification , a goal of ALA is to rephrase the original text in a more easily understandable manner . However , in ALA the system must often include additional information to clarify niche terminology and abstract concepts . To stimulate research on this task , we release a large collection of crowdsourced annotations for song lyrics . We analyze the performance of translation and retrieval models on this task , measuring performance with both automated and human evaluation . We find that each model captures a unique type of information important to the task . Song lyrics and poetry often make use of ambiguity , symbolism , irony , and other stylistic elements to evoke emotive responses . These characteristics sometimes make it challenging to interpret obscure lyrics , especially for readers or listeners who are unfamiliar with the genre . To address this problem , several online lyric databases have been created where users can explain , contextualize , or discuss lyrics . Examples include MetroLyrics1 and Genius.com2 . We refer to such How does it feel ? To be without a home Like a complete unknown , We presented and released the Genius dataset to study the task of Automated Lyric Annotation . As a first investigation , we studied automatic generation of context independent annotations as machine translation and information retrieval . Our baseline system tests indicate that our corpus is suitable to train machine translation systems . Standard SMT models are capable of rephrasing and simplifying song lyrics but tend to keep close to the structure of the song lyric . Seq2Seq models demonstrated potential to generate more fluent and informative text , dissimilar to the lyric . A large fraction of the annotations is heavily based on context and background knowledge ( CS ) , one of their most appealing aspects . As future work we suggest injection of structured and unstructured external knowledge ( Ahn et al . , 2016 ) and explicit modeling of references ( Yang et al . , 2016 ) .

Because of its creative nature, understanding lyrics can be challenging both for humans and machines.

They propose a task of automated lyric annotation with a dataset collected from an online platform which explains lyrics to readers.

They evaluate translation and retrieval models with automatic and human evaluation and show that different models capture different aspects well.

H05-1023

Most statistical translation systems are based on phrase translation pairs , or " blocks " , which are obtained mainly from word alignment . We use blocks to infer better word alignment and improved word alignment which , in turn , leads to better inference of blocks . We propose two new probabilistic models based on the innerouter segmentations and use EM algorithms for estimating the models ' parameters . The first model recovers IBM Model-1 as a special case . Both models outperform bidirectional IBM Model-4 in terms of word alignment accuracy by 10 % absolute on the F-measure . Using blocks obtained from the models in actual translation systems yields statistically significant improvements in Chinese-English SMT evaluation . Today 's statistical machine translation systems rely on high quality phrase translation pairs to acquire state-of-the-art performance , see ( Koehn et al . , 2003 ; Zens and Ney , 2004 ; Och and Ney , 2003 ) . Here , phrase pairs , or " blocks " are obtained automatically from parallel sentence pairs via the underlying word alignments . Word alignments traditionally are based on IBM Models 1 - 5 ( Brown et al . , 1993 ) or on HMMs ( Vogel et al . , 1996 ) . Automatic word alignment is challenging in that its accuracy is not yet close to inter-annotator agreement in some language pairs : for Chinese-English , inter-annotator agreement exceeds 90 on F-measure whereas IBM Model-4 or HMM accuracy is typically below 80s . HMMs assume that words " close-in-source " are aligned to words " close-in-target " . While this locality assumption is generally sound , HMMs do have limitations : the self-transition probability of a state ( word ) is the only control on the duration in the state , the length of the phrase aligned to the word . Also there is no natural way to control repeated non-contiguous visits to a state . Despite these problems , HMMs remain attractive for their speed and reasonable accuracy . We propose a new method for localizing word alignments . We use blocks to achieve locality in the following manner : a block in a sentence pair is a source phrase aligned to a target phrase . We assume that words inside the source phrase can not align to words outside the target phrase and that words outside the source phrase can not align to words inside the target phrase . Furthermore , a block divides the sentence pair into two smaller regions : the inner part of the block , which corresponds to the source and target phrase in the block , and the outer part of the block , which corresponds to the remaining source and target words in the parallel sentence pair . The two regions are non-overlapping ; and each of them is shorter than the original parallel sentence pair . The regions are thus easier to align than the original sentence pairs ( e.g. , using IBM Model-1 ) . While the model uses a single block to split the sentence pair into two independent regions , it is not clear which block we should select for this purpose . Therefore , we treat the splitting block as a hidden variable . This proposed approach is far simpler than treating the entire sentence as a sequence of nonoverlapping phrases ( or chunks ) and considering such sequential segmentation either explicitly or implicitly . For example , ( Marcu and Wong , 2002 ) for a joint phrase based model , ( Huang et al . , 2003 ) for a translation memory system ; and ( Watanabe et al . , 2003 ) for a complex model of insertion , deletion and head-word driven chunk reordering . Other approaches including ( Watanabe et al . , 2002 ) treat extracted phrase-pairs as new parallel data with limited success . Typically , they share a similar architecture of phrase level segmentation , reordering , translation as in ( Och and Ney , 2002 ; Koehn and Knight , 2002 ; Yamada and Knight , 2001 ) . The phrase level interaction has to be taken care of for the non-overlapping sequential segmentation in a complicated way . Our models model such interactions in a soft way . The hidden blocks are allowed to overlap with each other , while each block induced two non-overlapping regions , i.e. the model brackets the sentence pair into two independent parts which are generated synchronously . In this respect , it resembles bilingual bracketing ( Wu , 1997 ) , but our model has more lexical items in the blocks with many-to-many word alignment freedom in both inner and outer parts . We present our localization constraints using blocks for word alignment in Section 2 ; we detail our two new probabilistic models and their EM training algorithms in Section 3 ; our baseline system , a maximum-posterior inference for word alignment , is explained in Section 4 ; experimental results of alignments and translations are in Section 5 ; and Section 6 contains discussion and conclusions . Our main contributions are two novel Inner-Outer Bracket models based on segmentations induced by hidden blocks . Modeling the Inner-Outer hidden segmentations , we get significantly improved word alignments for both the small training set and the large training set over the widely-practiced bidirectional IBM Model-4 alignment . We also show significant improvements in translation quality using our proposed bracket models . Robustness to noisy blocks merits further investigation .

Automatic word alignment used in statistical machine translations, does not achieve satisfactory performance in some language pairs such as because of the limitations of HMMs.

They propose to use phrase translation pairs to get better word alignments using two new probabilistic models based on EM-algorithm that localizes the alignments.

The proposed models outperform IBM Model-4 by 10%, both on small and large training setups, and the translation models based on this result improve qualities.

E06-1051

We propose an approach for extracting relations between entities from biomedical literature based solely on shallow linguistic information . We use a combination of kernel functions to integrate two different information sources : ( i ) the whole sentence where the relation appears , and ( ii ) the local contexts around the interacting entities . We performed experiments on extracting gene and protein interactions from two different data sets . The results show that our approach outperforms most of the previous methods based on syntactic and semantic information . Information Extraction ( IE ) is the process of finding relevant entities and their relationships within textual documents . Applications of IE range from Semantic Web to Bioinformatics . For example , there is an increasing interest in automatically extracting relevant information from biomedical literature . Recent evaluation campaigns on bio-entity recognition , such as BioCreAtIvE and JNLPBA 2004 shared task , have shown that several systems are able to achieve good performance ( even if it is a bit worse than that reported on news articles ) . However , relation identification is more useful from an applicative perspective but it is still a considerable challenge for automatic tools . In this work , we propose a supervised machine learning approach to relation extraction which is applicable even when ( deep ) linguistic processing is not available or reliable . In particular , we explore a kernel-based approach based solely on shallow linguistic processing , such as tokeniza-tion , sentence splitting , Part-of-Speech ( PoS ) tagging and lemmatization . Kernel methods ( Shawe-Taylor and Cristianini , 2004 ) show their full potential when an explicit computation of the feature map becomes computationally infeasible , due to the high or even infinite dimension of the feature space . For this reason , kernels have been recently used to develop innovative approaches to relation extraction based on syntactic information , in which the examples preserve their original representations ( i.e. parse trees ) and are compared by the kernel function ( Zelenko et al . , 2003 ; Culotta and Sorensen , 2004 ; Zhao and Grishman , 2005 ) . Despite the positive results obtained exploiting syntactic information , we claim that there is still room for improvement relying exclusively on shallow linguistic information for two main reasons . First of all , previous comparative evaluations put more stress on the deep linguistic approaches and did not put as much effort on developing effective methods based on shallow linguistic information . A second reason concerns the fact that syntactic parsing is not always robust enough to deal with real-world sentences . This may prevent approaches based on syntactic features from producing any result . Another related issue concerns the fact that parsers are available only for few languages and may not produce reliable results when used on domain specific texts ( as is the case of the biomedical literature ) . For example , most of the participants at the Learning Language in Logic ( LLL ) challenge on Genic Interaction Extraction ( see Section 4.2 ) were unable to successfully exploit linguistic information provided by parsers . It is still an open issue whether the use of domainspecific treebanks ( such as the Genia treebank 1 ) can be successfully exploited to overcome this problem . Therefore it is essential to better investigate the potential of approaches based exclusively on simple linguistic features . In our approach we use a combination of kernel functions to represent two distinct information sources : the global context where entities appear and their local contexts . The whole sentence where the entities appear ( global context ) is used to discover the presence of a relation between two entities , similarly to what was done by Bunescu and Mooney ( 2005b ) . Windows of limited size around the entities ( local contexts ) provide useful clues to identify the roles of the entities within a relation . The approach has some resemblance with what was proposed by Roth and Yih ( 2002 ) . The main difference is that we perform the extraction task in a single step via a combined kernel , while they used two separate classifiers to identify entities and relations and their output is later combined with a probabilistic global inference . We evaluated our relation extraction algorithm on two biomedical data sets ( i.e. the AImed corpus and the LLL challenge data set ; see Section 4 ) . The motivations for using these benchmarks derive from the increasing applicative interest in tools able to extract relations between relevant entities in biomedical texts and , consequently , from the growing availability of annotated data sets . The experiments show clearly that our approach consistently improves previous results . Surprisingly , it outperforms most of the systems based on syntactic or semantic information , even when this information is manually annotated ( i.e. the LLL challenge ) . The good results obtained using only shallow linguistic features provide a higher baseline against which it is possible to measure improvements obtained using methods based on deep linguistic processing . In the near future , we plan to extend our work in several ways . First , we would like to evaluate the contribution of syntactic information to relation extraction from biomedical literature . With this aim , we will integrate the output of a parser ( possibly trained on a domain-specific resource such the Genia Treebank ) . Second , we plan to test the portability of our model on ACE and MUC data sets . Third , we would like to use a named entity recognizer instead of assuming that entities are already extracted or given by a dictionary . Our long term goal is to populate databases and ontologies by extracting information from large text collections such as Medline .

Deep linguistic features obtained by parsers are not always robust and available for limited languages and domains, however; applications of shallow features are under investigated.

They propose an approach for entity relation extraction using shallow linguistic information such as tokenization, sentence splitting, Part-of-Speech tagging and lemmatization coupled with kernel functions.

Evaluations of two biomedical datasets show that the proposed method outperforms existing systems which depend on syntactic or manually annotated semantic information.

N09-1032

Domain adaptation is an important problem in named entity recognition ( NER ) . NER classifiers usually lose accuracy in the domain transfer due to the different data distribution between the source and the target domains . The major reason for performance degrading is that each entity type often has lots of domainspecific term representations in the different domains . The existing approaches usually need an amount of labeled target domain data for tuning the original model . However , it is a labor-intensive and time-consuming task to build annotated training data set for every target domain . We present a domain adaptation method with latent semantic association ( LaSA ) . This method effectively overcomes the data distribution difference without leveraging any labeled target domain data . LaSA model is constructed to capture latent semantic association among words from the unlabeled corpus . It groups words into a set of concepts according to the related context snippets . In the domain transfer , the original term spaces of both domains are projected to a concept space using LaSA model at first , then the original NER model is tuned based on the semantic association features . Experimental results on English and Chinese corpus show that LaSA-based domain adaptation significantly enhances the performance of NER . Named entities ( NE ) are phrases that contain names of persons , organizations , locations , etc . NER is an important task in information extraction and natural language processing ( NLP ) applications . Supervised learning methods can effectively solve NER problem by learning a model from manually labeled data ( Borthwick , 1999 ; Sang and Meulder , 2003 ; Gao et al . , 2005 ; Florian et al . , 2003 ) . However , empirical study shows that NE types have different distribution across domains ( Guo et al . , 2006 ) . Trained NER classifiers in the source domain usually lose accuracy in a new target domain when the data distribution is different between both domains . Domain adaptation is a challenge for NER and other NLP applications . In the domain transfer , the reason for accuracy loss is that each NE type often has various specific term representations and context clues in the different domains . For example , { " economist " , " singer " , " dancer " , " athlete " , " player " , " philosopher " , ... } are used as context clues for NER . However , the distribution of these representations are varied with domains . We expect to do better domain adaptation for NER by exploiting latent semantic association among words from different domains . Some approaches have been proposed to group words into " topics " to capture important relationships between words , such as Latent Semantic Indexing ( LSI ) ( Deerwester et al . , 1990 ) , probabilistic Latent Semantic Indexing ( pLSI ) ( Hofmann , 1999 ) , Latent Dirichlet Allocation ( LDA ) ( Blei et al . , 2003 ) . These models have been successfully employed in topic modeling , dimensionality reduction for text categorization ( Blei et al . , 2003 ) , ad hoc IR ( Wei and Croft . , 2006 ) , and so on . In this paper , we present a domain adaptation method with latent semantic association . We focus on capturing the hidden semantic association among words in the domain adaptation . We introduce the LaSA model to overcome the distribution difference between the source domain and the target domain . LaSA model is constructed from the unlabeled corpus at first . It learns latent semantic association among words from their related context snippets . In the domain transfer , words in the corpus are associated with a low-dimension concept space using LaSA model , then the original NER model is tuned using these generated semantic association features . The intuition behind our method is that words in one concept set will have similar semantic features or latent semantic association , and share syntactic and semantic context in the corpus . They can be considered as behaving in the same way for discriminative learning in the source and target domains . The proposed method associates words from different domains on a semantic level rather than by lexical occurrence . It can better bridge the domain distribution gap without any labeled target domain samples . Experimental results on English and Chinese corpus show that LaSA-based adaptation significantly enhances NER performance across domains . The rest of this paper is organized as follows . Section 2 briefly describes the related works . Section 3 presents a domain adaptation method based on latent semantic association . Section 4 illustrates how to learn LaSA model from the unlabeled corpus . Section 5 shows experimental results on large-scale English and Chinese corpus across domains , respectively . The conclusion is given in Section 6 . We present a domain adaptation method with LaSA model in this paper . LaSA model captures latent semantic association among words from the unlabeled corpus . It better groups words into a set of concepts according to the related context snippets . LaSAbased domain adaptation method projects words to a low-dimension concept feature space in the transfer . It effectively overcomes the data distribution gap across domains without using any labeled target domain data . Experimental results on English and Chinese corpus show that LaSA-based domain adaptation significantly enhances the performance of NER across domains . Especially , LaSA model effectively increases the accuracy of each NE type recognition in the domain transfer . Moreover , LaSA-based domain adaptation method works well across languages . To further reduce the accuracy loss , we will explore informative sampling to capture fine-grained data difference in the domain transfer .

Named entity recognition classifiers lose accuracy in domain transfers because each entity type has domain-specific term representations and existing approaches require expensive labeled data.

They propose to capture latent semantic associations among words in the unlabeled corpus and use them to tune original named entity models.

The proposed model improves the performance on the English and Chinese corpus across domains especially on each NE type recognition.

D09-1030

Manual evaluation of translation quality is generally thought to be excessively time consuming and expensive . We explore a fast and inexpensive way of doing it using Amazon 's Mechanical Turk to pay small sums to a large number of non-expert annotators . For $ 10 we redundantly recreate judgments from a WMT08 translation task . We find that when combined non-expert judgments have a high-level of agreement with the existing gold-standard judgments of machine translation quality , and correlate more strongly with expert judgments than Bleu does . We go on to show that Mechanical Turk can be used to calculate human-mediated translation edit rate ( HTER ) , to conduct reading comprehension experiments with machine translation , and to create high quality reference translations . Conventional wisdom holds that manual evaluation of machine translation is too time-consuming and expensive to conduct . Instead , researchers routinely use automatic metrics like Bleu ( Papineni et al . , 2002 ) as the sole evidence of improvement to translation quality . Automatic metrics have been criticized for a variety of reasons ( Babych and Hartley , 2004 ; Callison-Burch et al . , 2006 ; Chiang et al . , 2008 ) , and it is clear that they only loosely approximate human judgments . Therefore , having people evaluate translation output would be preferable , if it were more practical . In this paper we demonstrate that the manual evaluation of translation quality is not as expensive or as time consuming as generally thought . We use Amazon 's Mechanical Turk , an online labor market that is designed to pay people small sums of money to complete human intelligence teststasks that are difficult for computers but easy for people . We show that : • Non-expert annotators produce judgments that are very similar to experts and that have a stronger correlation than Bleu . • Mechanical Turk can be used for complex tasks like human-mediated translation edit rate ( HTER ) and creating multiple reference translations . • Evaluating translation quality through reading comprehension , which is rarely done , can be easily accomplished through creative use of Mechanical Turk . 2 Related work Snow et al . ( 2008 ) examined the accuracy of labels created using Mechanical Turk for a variety of natural language processing tasks . These tasks included word sense disambiguation , word similarity , textual entailment , and temporal ordering of events , but not machine translation . Snow et al . measured the quality of non-expert annotations by comparing them against labels that had been previously created by expert annotators . They report inter-annotator agreement between expert and non-expert annotators , and show that the average of many non-experts converges on performance of a single expert for many of their tasks . Although it is not common for manual evaluation results to be reported in conference papers , several large-scale manual evaluations of machine translation quality take place annually . These include public forums like the NIST MT Evaluation Workshop , IWSLT and WMT , as well as the project-specific Go / No Go evaluations for the DARPA GALE program . Various types of human judgments are used . NIST collects 5-point fluency and adequacy scores ( LDC , 2005 ) , IWSLT and WMT collect relative rankings ( Callison-Burch et al . , 2008 ; Paul , 2006 ) , and DARPA evaluates using HTER ( Snover et al . , 2006 ) . The details of these are provided later in the paper . Public evaluation campaigns provide a ready source of goldstandard data that non-expert annotations can be compared to . Mechanical Turk is an inexpensive way of gathering human judgments and annotations for a wide variety of tasks . In this paper we demonstrate that it is feasible to perform manual evaluations of machine translation quality using the web service . The low cost of the non-expert labor found on Mechanical Turk is cheap enough to collect redundant annotations , which can be utilized to ensure translation quality . By combining the judgments of many non-experts we are able to achieve the equivalent quality of experts . The suggests that manual evaluation of translation quality could be straightforwardly done to validate performance improvements reported in conference papers , or even for mundane tasks like tracking incremental system updates . This challenges the conventional wisdom which has long held that automatic metrics must be used since manual evaluation is too costly and timeconsuming . We have shown that Mechanical Turk can be used creatively to produce quite interesting things . We showed how a reading comprehension test could be created , administered , and graded , with only very minimal intervention . We believe that it is feasible to use Mechanical Turk for a wide variety of other machine translated tasks like creating word alignments for sentence pairs , verifying the accuracy of document-and sentence-alignments , performing non-simulated active learning experiments for statistical machine translation , even collecting training data for low resource languages like Urdu . The cost of using Mechanical Turk is low enough that we might consider attempting quixotic things like human-in-the-loop minimum error rate training ( Zaidan and Callison-Burch , 2009 ) , or doubling the amount of training data available for Urdu .

Because of the high cost required for manual evaluation, most works rely on automatic evaluation metrics although there are several drawbacks.

They investigate whether judgements by non-experts from Amazon's Mechanical Turk can be a fast and inexpensive means of evaluation for machine translation systems.

They found that non-expert judgements with high agreement correlate better with gold standard judgements than BLEU while keeping the cost low.

D18-1133

State-of-the-art networks that model relations between two pieces of text often use complex architectures and attention . In this paper , instead of focusing on architecture engineering , we take advantage of small amounts of labelled data that model semantic phenomena in text to encode matching features directly in the word representations . This greatly boosts the accuracy of our reference network , while keeping the model simple and fast to train . Our approach also beats a tree kernel model that uses similar input encodings , and neural models which use advanced attention and compare-aggregate mechanisms . Modeling a match between pieces of text is at the core of many NLP tasks . Recently , manual feature engineering methods have been shadowed by neural network approaches . These networks model the interaction of two pieces of text , or word-toword interactions across sentences , using sophisticated attention mechanisms ( Wang et al . , 2016a ; Santos et al . , 2016 ) and compare-aggregate frameworks ( He and Lin , 2016 ; Wang et al . , 2017 ) . Architectural complexity is tied to longer training times 1 . Meaningful features may take long time to emerge by only leveraging word representations and the training data of the task at hand . This is especially problematic with little data , as it often happens in question answering ( QA ) tasks , e.g. , answer sentence selection ( Wang et al . , 2007 ; Yang et al . , 2015 ) . Thus , effective word representations are crucial in neural network models to get state-of-the-art performance . In this work , we try to answer the following research questions : ( i ) in addition to lexical links , can we incorporate higher-level semantic links between the words in a question and a candidate answer passage , and ( ii ) can we show that such information has an impact on the quality of our model , and also allows us to keep the architecture simple ? We show that modeling semantic relations improves the performance of a neural network for answer sentence selection with ( i ) a little number of semantic annotations , and ( ii ) a little increase in training time w.r.t . more complex architecture . In this paper , we presented a neural network that models semantic links between questions and answers , in addition to lexical links . The annotations for establishing such links are produced by a set of fast neural components for question analysis , trained on publicly available datasets . The evaluation on two QA datasets shows that our approach can achieve state-of-the-art performance using a simple CNN , leading to a low complexity and training time . Our approach is an interesting first step towards a future architecture , in which we will jointly optimize the semantic annotators and the answer sentence selection model , in an end-to-end fashion .

State-of-the-art models that model relations between two texts use complex architectures and attention which requires a long time and large data at training.

They propose a method that directly models higher-level semantic links between two texts that are annotated by a fast model.

The proposed model outperforms a tree kernel model and complex neural models while keeping the model simple and the training fast.

2021.emnlp-main.411

Language representations are known to carry certain associations ( e.g. , gendered connotations ) which may lead to invalid and harmful predictions in downstream tasks . While existing methods are effective at mitigating such unwanted associations by linear projection , we argue that they are too aggressive : not only do they remove such associations , they also erase information that should be retained . To address this issue , we propose OS-CAR ( Orthogonal Subspace Correction and Rectification ) , a balanced approach of mitigation that focuses on disentangling associations between concepts that are deemed problematic , instead of removing concepts wholesale . We develop new measurements for evaluating information retention relevant to the debiasing goal . Our experiments on genderoccupation associations show that OSCAR is a well-balanced approach that ensures that semantic information is retained in the embeddings and unwanted associations are also effectively mitigated . Keeping in mind that removing bias and retaining information have to be done in synergy , we present how to obtain aggregated measurements for these two components . We will first describe the de- Word embeddings are used extensively across natural language processing ( NLP ) and succinctly capture not only the syntactic and semantic structure of language , but also word meaning in context . As such , word embeddings are essential building blocks for today 's state-of-the-art in NLP . But they are also known to capture a significant amount of stereotypical associations ( e.g. , Bolukbasi et al . , 2016 ; Zhao et al . , 2017 ; Dev and Phillips , 2019 ; Sun et al . , 2019 ) related to gender , race , nationality , or religion , which can manifest in unwanted and/or potentially harmful ways in downstream tasks ( Webster et al . , 2018 ; Zhao et al . , 2019 ; Dev et al . , 2020 ) . Such potentially problematic associations , when embedded in word representations , can lead to incorrect and unfair decisions about large groups of people . While the term " bias " has many meanings , in this paper we use it to refer to these unwanted stereotypical associations . Existing methods to mitigate these effects either require expensive retraining of vectors ( Lemoine et al . , 2018 ) which can be inefficient , or projecting out information contained along an entire subspace representing a protected concept ( such as gender or race ) in the embedding space ( e.g. , Bolukbasi et al . , 2016 ; Dev and Phillips , 2019 ; Ravfogel et al . , 2020 ) . Projective approaches are difficult to control as they are either insufficient : removing a subspace can still leave residual bias ( Gonen and Goldberg , 2019 ; Lauscher et al . , 2020 ) , or too aggressive : in the case of gender , also unnecessarily altering the association between the word pregnant and words like female and mother . In tasks such as coreference resolution , removing such associations could hinder reference resolution . To quantify how much valid information is retained , we look at the output space of models . Following Dev et al . ( 2020 ) , we use Natural Language Inference ( NLI ) as an effective quantitative probe . Here , we construct the hypothesis by making minimal edits to a premise , and observe model prediction conditioned on such changes . For example , Premise : A matriarch sold a watch . Hypothesis : A woman sold a watch . Here , the objective is to determine if the hypothesis is entailed by the premise , contradicted by it , or neither ( neutral to it ) . A GloVe-based NLI ( Parikh et al . , 2016 ) model , without any explicit form of bias mitigation , predicts label entail with a high probability of 97 % ; the notion of a matriarch being a woman is correctly identified by the model . However , after projective debiasing , the model classifies the pair as neutral with a probability 62 % while the probability of the label entail drops to much lower at 16 % . That is , aggressive mitigation of gender representations erases valid gender associations . Ideally , we should correct problematic associations without erasing valid ones . To this end , we propose OSCAR ( Othogonal Subspace Correction and Rectification ) which orthogonalizes and rectifies identified subspaces of concepts that are incorrectly associated in an embedding space . Embeddings outside the subspace are stretched in a graded manner or untouched . Our contributions are : 1 . We argue that mitigating unwanted stereotypical associations should go beyond information removal ( e.g. , projecting out features ) , and should also preserve pertinent associations . 2 . We present OSCAR1 , a completely different method from the existing projective approaches ; it uses orthogonalization of subspaces desired not to have interdependence , and so minimal change is made to embeddings to prevent loss of desired associations . 3 . We develop a combination of tests based on NLI to evaluate both bias mitigation and information retention . 4 . Our experiments show that OSCAR is a wellbalanced approach that mitigates biases as good as projective approaches while retaining more valid associations . Our contributions , focusing specifically on representations rather than classifiers , are important given the preponderance of distributed representations of text across NLP . Predictions from systems that use such representations , if unchecked , could lead to real-world decisions ( involving , e.g. hiring ) that " systematically and unfairly discriminate against certain individuals or groups of individuals in favor of others " ( Friedman and Nissenbaum , 1996 ) . In an effort to prevent such transition of representational harms into allocational harms ( cf . Crawford , 2017 ; Abbasi et al . , 2019 ; Blodgett et al . , 2020 ) , we look at mitigating stereotyping biases at the source , i.e. the embedding space . 2 Bias , Gender and NLP Bias and Related Work . Social biases in tasks using machine learning have the potential to cause harms ( Barocas and Selbst , 2016 ) and are widely studied . Such biases in language technologies have been detected ( Bolukbasi et al . , 2016 ; Gonen and Goldberg , 2019 ) , measured ( Caliskan et al . , 2017 ; Webster et al . , 2018 ; Lauscher et al . , 2020 ) and mitigated ( Zhao et al . , 2018 ; Ravfogel et al . , 2020 ) . Most work focuses on gender bias , and in particular , the stereotypical associations of occupations to males and females ( Rudinger et al . , 2018 ; De-Arteaga et al . , 2019 ; Gaut et al . , 2019 ; Dev et al . , 2020 ) . In this work too , bias refers to stereotypical associations in language representations . Treatment of Gender . The concept of gender is a complex one ( cf . Larson , 2017 ; Dev et al . , 2021 ) . In our experiments , we identify a subspace in the representation associated with gender , typically derived by terms specifically associated with notions of male and female . This aligns with prominent recent NLP research ( Caliskan et al . , 2017 ; Bolukbasi et al . , 2016 ; Dev and Phillips , 2019 ; Ravfogel et al . , 2020 ) that try to measure and mitigate gender bias ( either implicitly or explicitly ) , directly focusing on male versus female associations . The underlying goal of this work is to disassociate two concepts . We mainly use concepts male-vsfemale and occupations to demonstrate this . There is explicit and measurable bias encoded in word vector representations ( Bolukbasi et al . , 2016 ) by the male-vs-female relation to occupations , and standard evaluation metrics ( Caliskan et al . , 2017 ; Dev et al . , 2020 ) are designed to measure this . As such many of our experiments are geared towards these male-vs-female issues . The notion of gender is not restricted to male and female . But because " conceptualizations of gender as binary are reinforced linguistically in English through pronoun conventions " ( Blodgett , 2021 ) , gender notions beyond male and female are either missing or poorly represented . As a result , their associations with other concepts are also not robust . Quantifying this effect , and evaluating its attenuation , is beyond the scope of this paper ; however the techniques we study for disassociating two concepts are amenable to such tasks in general . We remark that a potential harm towards nonmale / female people arises through omission ( Cao and Daumé III , 2020 ) , and as a result , we explicitly encourage further research on this topic . sign of our evaluation dataset in § 3.1 , then present evaluation metrics . Then , we will discuss how to measure biases in § 3.2 and information retention in § 3.3 . To be comprehensive , we will use metrics that are both intrinsic to embedding geometry and extrinsic ones that focus on downstream NLI model outputs . The propagation of undesirable and stereotypical associations learned from data into decisions made by language models maintains a vicious cycle of biases . Combating biases before deploying representations is thus extremely vital . But this poses its own challenges . Word embeddings capture a lot of information implicitly in relatively few dimensions . These implicit associations are what makes them state-of-the-art at tackling different language modeling tasks . Breaking down these associations for bias rectification , thus , has to be done carefully so as to not destroy the structure of the embeddings . OSCAR 's rectification of associations helps integrate both these aspects , allowing it to be more suitable at making word embeddings more usable . Moreover , being computationally lightweight and sub-differentiable , it is simple to apply adaptively without extensive retraining . OSCAR dissociates concepts which may be otherwise lost or overlapped in distributed representations by re-orthogonalizing them . We envision that this method can be used for many types of unwanted associations ( age-ability , religion-virtue perception , etc ) . Since it only decouples specific associations , informative components of these features will remain unchanged . All the experiments in this paper are based on English embeddings , models and tests ; extending our analysis to other languages is important future work . Further , we be-lieve OSCAR can extend in a straightforward way beyond NLP to other distributed , vectorized representations ( e.g. , of images , graphs , or spatial data ) which can also exhibit stereotypical associations .

Existing methods that remove harmful stereotypical associations from word embeddings either require inefficient retraining or remove information which should be retained.

They propose a method which orthogonalizes and rectifies incorrectly associated subspaces of concepts in an embedding space and a metric for evaluating information retention.

NLI-based evaluation on gender-occupation associations shows that the proposed approach is well-balanced ensuring semantic information is retained in the embeddings while mitigating biases.

2020.acl-main.75

Humor plays an important role in human languages and it is essential to model humor when building intelligence systems . Among different forms of humor , puns perform wordplay for humorous effects by employing words with double entendre and high phonetic similarity . However , identifying and modeling puns are challenging as puns usually involved implicit semantic or phonological tricks . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to perceive human humor , detect if a sentence contains puns and locate them in the sentence . PCPR derives contextualized representation for each word in a sentence by capturing the association between the surrounding context and its corresponding phonetic symbols . Extensive experiments are conducted on two benchmark datasets . Results demonstrate that the proposed approach significantly outperforms the state-of-the-art methods in pun detection and location tasks . In-depth analyses verify the effectiveness and robustness of PCPR . During the last decades , social media has promoted the creation of a vast amount of humorous web contents ( Nijholt et al . , 2017 ) . Automatic recognition of humor has become an important task in the area of figurative language processing , which can benefit various downstream NLP applications such as dialogue systems , sentiment analysis , and machine translation ( Melby and Warner , 1995 ; Augello et al . , 2008 ; Ghosh et al . , 2015 ; Bertero and Fung , 2016 ; Blinov et al . , 2019 ) . However , humor is one of the most complicated behaviors in natural language semantics and sometimes it is even difficult for humans to interpret . In most cases , understanding humor requires adequate background knowledge and a rich context . Puns are a form of humorous approaches using the different meanings of identical words or words with similar pronunciations to explain texts or utterances . There are two main types of puns . Homographic puns rely on multiple interpretations of the same word . As shown in Table 1 , the phrase all right means good condition or opposite to left ; the word reaction means chemical change or action . The two meanings of the same expression are consistent with its context , which creates a humorous pun in both sentences when there is a clear contrast between two meanings . On the other hand , heterographic puns take advantage of phonologically same or similar words . For example , the word pairs sale and sail , weak and week in Table 1 have the same or similar pronunciations . The sentences are funny because both words fit the same context . Understanding puns is a big fish to fry for deep comprehension of complex semantics . These two forms of puns have been studied in literature from different angles . To recognize puns in a sentence , word sense disambiguation techniques ( WSD ) ( Navigli , 2009 ) have been employed to identify the equitable intention of words in utterances ( Pedersen , 2017 ) . External knowledge bases such as WordNet ( Miller , 1998b ) have been applied in determining word senses of pun words ( Oele and Evang , 2017 ) . However , these methods can not tackle heterographic puns with distinct word spellings and knowledge bases that only contain a limited vocabulary . To resolve the issues of sparseness and heterographics , the word embedding techniques ( Mikolov et al . , 2013 ; Pennington et al . , 2014 ) provide flexible representations to model puns ( Hurtado et al . , 2017 ; Indurthi and Oota , 2017 ; Cai et al . , 2018 ) . However , a word may have different meanings regarding its contexts . Especially , an infrequent meaning of the word might be utilized for creating a pun . Therefore , static word embeddings are insufficient to represent words . In addition , some puns are created by replacing a word with another word with the same or similar pronunciation as examples shown in Table 1 . Therefore , to recognize puns , it is essential to model the association between words in the sentence and the pronunciation of words . Despite existing approaches attempt to leverage phonological structures to understand puns ( Doogan et al . , 2017 ; Jaech et al . , 2016 ) , there is a lack of a general framework to model these two types of signals in a whole . In this paper , we propose Pronunciation-attentive Contextualized Pun Recognition ( PCPR ) to jointly model the contextualized word embeddings and phonological word representations for pun recognition . To capture the phonological structures of words , we break each word into a sequence of phonemes as its pronunciation so that homophones can have similar phoneme sets . For instance , the phonemes of the word pun are { P , AH , N } . In PCPR , we construct a pronunciation attentive module to identify important phonemes of each word , which can be applied in other tasks related to phonology . We jointly encode the contextual and phonological features into a self-attentive embedding to tackle both pun detection and location tasks . We summarize our contributions as following . • To the best of our knowledge , PCPR is the first work to jointly model contextualized word embeddings and pronunciation embeddings to recognize puns . Both contexts and phonological properties are beneficial to pun recognition . In this paper , we propose a novel approach , PCPR , for pun detection and location by leveraging a contextualized word encoder and modeling phonemes as word pronunciations . Moreover , we would love to apply the proposed model to other problems , such as general humor recognition , irony discovery , and sarcasm detection , as the future work .

Puns involve implicit semantic or phonological tricks however there is no general framework to model these two types of signals as a whole.

They propose to jointly model contextualized word embeddings and phonological word representations by breaking each word into a sequence of phonemes for pun detection.

The proposed approach outperforms the state-of-the-art methods in pun detection and location tasks.

D10-1083

Part-of-speech ( POS ) tag distributions are known to exhibit sparsity -a word is likely to take a single predominant tag in a corpus . Recent research has demonstrated that incorporating this sparsity constraint improves tagging accuracy . However , in existing systems , this expansion come with a steep increase in model complexity . This paper proposes a simple and effective tagging method that directly models tag sparsity and other distributional properties of valid POS tag assignments . In addition , this formulation results in a dramatic reduction in the number of model parameters thereby , enabling unusually rapid training . Our experiments consistently demonstrate that this model architecture yields substantial performance gains over more complex tagging counterparts . On several languages , we report performance exceeding that of more complex state-of-the art systems . 1 Since the early days of statistical NLP , researchers have observed that a part-of-speech tag distribution exhibits " one tag per discourse " sparsity -words are likely to select a single predominant tag in a corpus , even when several tags are possible . Simply assigning to each word its most frequent associated tag in a corpus achieves 94.6 % accuracy on the WSJ portion of the Penn Treebank . This distributional sparsity of syntactic tags is not unique to English 1 The source code for the work presented in this paper is available at http://groups.csail.mit.edu / rbg / code / typetagging/. -similar results have been observed across multiple languages . Clearly , explicitly modeling such a powerful constraint on tagging assignment has a potential to significantly improve the accuracy of an unsupervised part-of-speech tagger learned without a tagging dictionary . In practice , this sparsity constraint is difficult to incorporate in a traditional POS induction system ( Mérialdo , 1994 ; Johnson , 2007 ; Gao and Johnson , 2008 ; Grac ¸a et al . , 2009 ; Berg-Kirkpatrick et al . , 2010 ) . These sequence models-based approaches commonly treat token-level tag assignment as the primary latent variable . By design , they readily capture regularities at the token-level . However , these approaches are ill-equipped to directly represent type-based constraints such as sparsity . Previous work has attempted to incorporate such constraints into token-level models via heavy-handed modifications to inference procedure and objective function ( e.g. , posterior regularization and ILP decoding ) ( Grac ¸a et al . , 2009 ; Ravi and Knight , 2009 ) . In most cases , however , these expansions come with a steep increase in model complexity , with respect to training procedure and inference time . In this work , we take a more direct approach and treat a word type and its allowed POS tags as a primary element of the model . The model starts by generating a tag assignment for each word type in a vocabulary , assuming one tag per word . Then , tokenlevel HMM emission parameters are drawn conditioned on these assignments such that each word is only allowed probability mass on a single assigned tag . In this way we restrict the parameterization of a token-level HMM to reflect lexicon sparsity . This model admits a simple Gibbs sampling algorithm where the number of latent variables is proportional to the number of word types , rather than the size of a corpus as for a standard HMM sampler ( Johnson , 2007 ) . There are two key benefits of this model architecture . First , it directly encodes linguistic intuitions about POS tag assignments : the model structure reflects the one-tag-per-word property , and a typelevel tag prior captures the skew on tag assignments ( e.g. , there are fewer unique determiners than unique nouns ) . Second , the reduced number of hidden variables and parameters dramatically speeds up learning and inference . We evaluate our model on seven languages exhibiting substantial syntactic variation . On several languages , we report performance exceeding that of state-of-the art systems . Our analysis identifies three key factors driving our performance gain : 1 ) selecting a model structure which directly encodes tag sparsity , 2 ) a type-level prior on tag assignments , and 3 ) a straightforward naïve-Bayes approach to incorporate features . The observed performance gains , coupled with the simplicity of model implementation , makes it a compelling alternative to existing more complex counterparts . We have presented a method for unsupervised partof-speech tagging that considers a word type and its allowed POS tags as a primary element of the model . This departure from the traditional token-based tagging approach allows us to explicitly capture typelevel distributional properties of valid POS tag as-signments as part of the model . The resulting model is compact , efficiently learnable and linguistically expressive . Our empirical results demonstrate that the type-based tagger rivals state-of-the-art tag-level taggers which employ more sophisticated learning mechanisms to exploit similar constraints . In this paper , we make a simplifying assumption of one-tag-per-word . This assumption , however , is not inherent to type-based tagging models . A promising direction for future work is to explicitly model a distribution over tags for each word type . We hypothesize that modeling morphological information will greatly constrain the set of possible tags , thereby further refining the representation of the tag lexicon .

Assuming there is only one tag for a word is a powerful heuristic for Part-of-speech tagging but incorporating this into a model leads to complexity.

They propose an unsupervised method that directly incorporates a one-tag-per-word assumption into a HMM-based model.

Their proposed method reduces the number of model parameters which results in faster training speed and also outperforms more complex systems.

P10-1072

We present a game-theoretic model of bargaining over a metaphor in the context of political communication , find its equilibrium , and use it to rationalize observed linguistic behavior . We argue that game theory is well suited for modeling discourse as a dynamic resulting from a number of conflicting pressures , and suggest applications of interest to computational linguists . A 13 Dec 1992 article in The Times starts thus : The European train chugged out of the station last night ; for most of the day it looked as if it might be stalled there for some time . It managed to pull away at around 10:30 pm only after the Spanish prime minister , Felipe Gonzalez , forced the passengers in the first class carriages into a last minute whip round to sweeten the trip for the European Community 's poor four : Spain , Portugal , Greece and Ireland . The fat controller , Helmut Kohl , beamed with satisfaction as the deal was done . The elegantlysuited Francois Mitterrand was equally satisfied . But nobody was as pleased as John Major , stationmaster for the UK presidency , for whom the agreement marked a scarce high point in a battered premiership . The departure had actually been delayed by seven months by Danes on the line . Just when that problem was solved , there was the voluble outbreak , orchestrated by Spain , from the poor four passengers demanding that they should travel free and be given spending money , too . The coupling of the carriages may not be reliably secure but the pan-European express is in motion . That few seem to agree the destination suggests that future arguments are inevitable at every set of points . Next stop : Copenhagen . Apart from an entertaining read , the extended metaphor provides an elaborate conceptual correspondence between a familiar domain of train journeys and the unfolding process of European integration . Carriages are likened to nation states ; passengers to their peoples ; treaties to stations ; politicians to responsible rail company employees . In a compact form , the metaphor gives expression to both the small and the large scale of the process . It provides for the recent history : Denmark 's failure to ratify the 1992 Maastricht treaty until opt-outs were negotiated later that year is compared to dissenters sabotaging the journey by laying on the tracks ( Danes on the line ) ; negotiations over the Cohesion Fund that would provide less developed regions with financial aid to help them comply with convergence criteria are likened to second class carriages with poor passengers for whom the journey had to be subsidized . At a more general level , the European integration is a purposeful movement towards some destination according to a worked out plan , getting safely through negotiation and implementation from one treaty to another , as a train moving on its rails through subsequent stations , with each nation being separate yet tied with everyone else . Numerous inferences regarding speed , timetables , stations , passengers , different classes of tickets , temporary obstacles on the tracks , and so on can be made by the reader based on the knowledge of train journeys , giving him or her a feeling of an enhanced understanding1 of the highly complex process of European integration . So apt was the metaphor that political fights were waged over its details ( Musolff , 2000 ) . Worries about destination were given an eloquent expression by Margaret Thatcher ( Sunday Times , 20 Sept 1992 ): She warned EC leaders to stop their endless round of summits and take notice of their own people . " There is a fear that the European train will thunder forward , laden with its customary cargo of gravy , towards a destination neither wished for nor understood by electorates . But the train can be stopped , " she said . The metaphor proved flexible enough for further elaboration . John Major , a Conservative PM of Britain , spoke on June 1st , 1994 about his vision of the decision making at the EU level , saying that he had never believed that Europe must act as one on every issue , and advocating " a sensible new approach , varying when it needs to , multitrack , multi-speed , multi-layered . " He attempted to turn a largely negative Conservative take on the European train ( see Thatcher above ) into a tenable positive vision -each nation-carriage is now presumably a rather autonomous entity , waiting on a side track for the right locomotive , in a huge yet smoothly operating railroad system . Major 's political opponents offered their counter-frames . In both cases , the imagery of a large transportation system was taken up , yet turned around to suggest that " multi , for everyone " amounts to Britain being in " the slow lane , " and a different image was suggested that makes the negative evaluation of Britain 's opt-outs more poignant -a football metaphor , where relegation to the second division is a sign of a weak performance , and a school metaphor , where Britain is portrayed as an under-achiever : John Cunningham , Labour He has admitted that his Government would let Britain fall behind in Europe . He is apparently willing to offer voluntary relegation to the second division in Europe , and he is n't even prepared to put up a fight . I believe that in any two-speed Europe , Britain must be up with those in the fast lane . Clearly Mr Major does not . Paddy Ashdown , Liberal Democrat Are you really saying that the best that Britain can hope for under your leadership is ... the slow lane of a two-speed Europe ? Most people in this country will want to aim higher , and will reject your view of a ' drop-out ' Britain . The pro-European camp rallied around the " Britain in the slow lane " version as a critical stance towards the government 's European policy . Of the alternative metaphors , the school metaphor has some traction in the Euro discourse , where the European ( mainly German ) financial officers are compared to school authorities , and governments struggling to meet the strict convergence criteria to enter the Euro are compared to pupils that barely make the grade with Britain as a ' drop-out ' who gave up even trying ( Musolff , 2000 ) . The fact that European policy is being communicated and negotiated via a metaphor is not surprising ; after all , " there is always someone willing to help us think by providing us with a metaphor that accords with HIS views . " 2 From the point of view of the dynamics of political discourse , the puzzle is rather the apparent tendency of politicians to be compelled by the rival 's metaphorical framework . Thatcher tries to turn the train metaphor used by the pro-EU camp around . Yet , assuming metaphors are matters of choice , why should Thatcher feel constrained by her rival 's choice , why does n't she ignore it and merely suggest a new metaphor of her own design ? As the evidence above suggests , this is not Thatcher 's idiosyncrasy , as Major and his rivals acted similarly . Can this dynamic be explained ? In this article , we use the explanatory framework of game theory , seeking to rationalize the observed behavior by designing a game that would produce , at equilibrium , the observed dynamics . Specifically , we formalize the notion that the price of " locking " the public into a metaphorical frame of reference is that a politician is coerced into staying within the metaphor as well , even if he or she is at the receiving end of a rival 's rhetorical move . Since the use of game theory is not common in computational linguistics , we first explain its main attributes , justify our decision to make use of it , and draw connections to research questions that can benefit from its application ( section 2 ) . Next , we design the game of bargaining over a metaphor , and find its equilibrium ( section 3 ) , followed by a discussion ( section 4 ) . This article addressed a specific communicative setting ( rival politicians trying to " sell " to the public their versions of the unfolding realities and necessary policies ) and a specific linguistic tool ( an extended metaphor ) , showing that the particular use made of metaphor in such setting can be rationalized based on the characteristics of the setting . Various questions now arise . Given the central role played by the public gratification constraint in our model , would conversational situations without the need to persuade the public , such as meetings of small groups of peers or phone conversations between friends , tend less to the use of extended metaphor ? Conversely , does the use of extended metaphor in other settings testify to the existence of presumed onlookers who need to be " captured " in a particular version of reality -as in pedagogic or poetic context ? Considerations of the participants ' agendas and their impact on the ensuing dynamics of the exchange would we believe lead to further interest in game theoretic models when addressing complex social dynamics in situations like collaborative authoring , debates , or dating , and will augment the existing mostly statistical approaches with a broader picture of the relevant communication .

Metaphors used in political arguments provide elaborate conceptual correspondences with a tendency of politicians to be compelled by the rival's metaphorical framework to be explained.

They propose a game-theoric model of bargaining over a metaphor which is suitable to model its dynamics and use to rationalize observed linguistic behavior.

They show that the proposed framework can rationalize political communications with the use of extended metaphors based on the characteristics of the setting.

2020.acl-main.47

We examine a methodology using neural language models ( LMs ) for analyzing the word order of language . This LM-based method has the potential to overcome the difficulties existing methods face , such as the propagation of preprocessor errors in count-based methods . In this study , we explore whether the LMbased method is valid for analyzing the word order . As a case study , this study focuses on Japanese due to its complex and flexible word order . To validate the LM-based method , we test ( i ) parallels between LMs and human word order preference , and ( ii ) consistency of the results obtained using the LM-based method with previous linguistic studies . Through our experiments , we tentatively conclude that LMs display sufficient word order knowledge for usage as an analysis tool . Finally , using the LMbased method , we demonstrate the relationship between the canonical word order and topicalization , which had yet to be analyzed by largescale experiments . Model Verbs whose type-A examples prefer the ACC-DAT order Verbs whose type-B examples prefer the ACC-DAT order CLM " 預ける " ( deposit ) , " 置く " ( place ) , " 持つ " ( have ) , " 入 れる " ( put in ) , " 納める " ( pay ) , " 郵送 " ( mail ) , " 供 給 " ( supply ) , " 出す " ( put out ) , " 運ぶ " ( transport ) , " 流 す " ( shed ) , " 掛ける " ( hang ) , " 飾る " ( decorate ) , " 広 げる " ( spread ) , " 移す " ( transfer ) , " 残す " ( leave ) , " 配 送 " ( deliver ) , " 送る " ( send ) , " 投げる " ( throw ) , " 送 付 " ( send ) , " 返却 " ( return ) , " 届ける " ( send ) , " 戻す " ( return ) , " 着ける " ( wear ) , " 上げる " ( increase ) , " 落と す " ( drop ) , " 載せる " ( publish ) , " 変更 " ( change ) , " 納 入 " ( deliver ) , " 卸す " ( unload ) , " 掲載 " ( publish ) , " 通 す " ( get X through ) " 配 布 " ( distribute ) , " 渡 す " ( pass ) , " プ レ ゼ ン ト " ( present ) , " 合わせる " ( match ) , " 見せる " ( show ) , " 提 供 " ( offer ) , " 与える " ( give ) , " 当てる " ( hit ) , " 回す " ( turn ) , " 追加 " ( add ) , " 貸す " ( lend ) , " 展示 " ( exhibit ) , " 据える " ( lay ) , " 依頼 " ( request ) , " 挿入 " ( insert ) , " 纏 める " ( collect ) , " 請求 " ( claim ) SLM " 預ける " ( deposit ) , " 置く " ( place ) , " 頼む " ( ask ) , " 入 れる " ( put in ) , " 納める " ( pay ) , " 郵送 " ( mail ) , " 出 す " ( put out ) , " 運ぶ " ( transport ) , " 流す " ( shed ) , " 掛け る " ( hang ) , " 広げる " ( spread ) , " 移す " ( transfer ) , " 残 す " ( leave ) , " リクエスト " ( request ) , " 配送 " ( deliver ) , " 送る " ( send ) , " 投げる " ( throw ) , " 送付 " ( send ) , " 求 める " ( ask ) , " 提出 " ( submit ) , " 届ける " ( deliver ) , " 要 求 " ( request ) , " 戻す " ( return ) , " 寄付 " ( donate ) , " 寄贈 " ( donation ) , " 着ける " ( wear ) , " 乗せる " ( place ) , " 上げ る " ( increase ) , " 落とす " ( drop ) , " 貼る " ( stick ) , " 分け る " ( divide ) , " ばらまく " ( scatter ) , " はめる " ( fit ) , " 支 払う " ( pay ) , " 配達 " ( deliver ) , " 卸す " ( unload ) , " 纏め る " ( collect ) , " 通す " ( get X through ) " プレゼント " ( present ) , " 持つ " ( have ) , " 合わせる " ( match ) , " 見せる " ( show ) , " 向ける " ( point ) , " 提供 " ( offer ) , " 装備 " ( equip ) , " 追加 " ( add ) , " 展示 " ( exhibit ) , " 据える " ( lay ) , " 採用 " ( adopt ) Speakers sometimes have a range of options for word order in conveying a similar meaning . A typical case in English is dative alternation : ( 1 ) a. A teacher gave a student a book . b. A teacher gave a book to a student . Even for such a particular alternation , several studies ( Bresnan et al . , 2007 ; Hovav and Levin , 2008 ; Colleman , 2009 ) investigated the factors determining this word order and found that the choice is not random . For analyzing such linguistic phenomena , linguists repeat the cycle of constructing hypotheses and testing their validity , usually through psychological experiments or count-based methods . However , these approaches sometimes face difficulties , such as scalability issues in psychological experiments and the propagation of preprocessor errors in count-based methods . Compared to the typical approaches for evaluating linguistic hypotheses , approaches using LMs have potential advantages ( Section 3.2 ) . In this study , we examine the methodology of using LMs for analyzing word order ( Figure 1 ) . To validate the LM-based method , we first examine if there is a parallel between canonical word order and generation probability of LMs for each word order . Futrell and Levy ( 2019 ) reported that English LMs have human-like word order preferences , which can be one piece of evidence for validating the LM-based method . However , it is not clear whether the above assumption is valid even in languages with more flexible word order . In this study , we specifically focus on the Japanese language due to its complex and flexible word order . There are many claims on the canonical word order of Japanese , and it has attracted considerable attention from linguists and natural language processing ( NLP ) researchers for decades ( Hoji , 1985 ; Saeki , 1998 ; Miyamoto , 2002 ; Matsuoka , 2003 ; Koizumi and Tamaoka , 2004 ; Nakamoto et al . , 2006 ; Shigenaga , 2014 ; Sasano and Okumura , 2016 ; Orita , 2017 ; Asahara et al . , 2018 ) . We investigated the validity of using Japanese LMs for canonical word order analysis by conducting two sets of experiments : ( i ) comparing word order preference in LMs to that in Japanese speakers ( Section 4 ) , and ( ii ) checking the consistency Topic Time Location Subject ( Adverb ) Indirect object Direct object Verb Notation TOP TIM LOC NOM -DAT ACC -Typical particle " は " ( wa ) " に " ( ni ) " で " ( de ) " が " ( ga ) -"に " ( ni ) " を " ( o ) -Related section 6 5.2 5.2 5.2 5.3 5.1 5.1 5.1 Table 1 : Overview of the typical cases in Japanese , their typical particles , and the sections where the corresponding case is analyzed . The well-known canonical word order of Japanese is listed from left to right . between the preference of LMs with previous linguistic studies ( Section 5 ) . From our experiments , we tentatively conclude that LMs display sufficient word order knowledge for usage as an analysis tool , and further explore potential applications . Finally , we analyzed the relationship between topicalization and word order of Japanese by taking advantage of the LM-based method ( Section 6 ) . In summary , we : • Discuss and validate the use of LMs as a tool for word order analysis as well as investigate the sensitivity of LMs against different word orders in non-European language ( Section 3 ) ; • Find encouraging parallels between the results obtained with the LM-based method and those with the previously established method on various hypotheses of canonical word order of Japanese ( Sections 4 and 5 ) ; and • Showcase the advantages of an LM-based method through analyzing linguistic phenomena that is difficult to explore with the previous data-driven methods ( Section 6 ) . We have proposed to use LMs as a tool for analyzing word order in Japanese . Our experimental results support the validity of using Japanese LMs for canonical word order analysis , which has the potential to broaden the possibilities of linguistic research . From an engineering view , this study supports the use of LMs for scoring Japanese word order automatically . From the viewpoint of the linguistic field , we provide additional empirical evidence to various word order hypotheses as well as demonstrate the validity of the LM-based method . We plan to further explore the capability of LMs on other linguistic phenomena related to word order , such as " given new ordering " ( Nakagawa , 2016 ; Asahara et al . , 2018 ) . Since LMs are language-agnostic , analyzing word order in another language with the LM-based method would also be an interesting direction to investigate . Furthermore , we would like to extend a comparison between machine and human language processing beyond the perspective of word order .

Linguistical approaches to analyze word order phenomena have scalability and preprocessor error propagation problems, and the use of language models is limited in English.

They validate language models as a tool to study word order in Japanese by examining the relationship between canonical word order and generation probability.

They show that language models have sufficient word order knowledge in Japanese to be used as a tool for linguists.

2021.naacl-main.150

A conventional approach to improving the performance of end-to-end speech translation ( E2E-ST ) models is to leverage the source transcription via pre-training and joint training with automatic speech recognition ( ASR ) and neural machine translation ( NMT ) tasks . However , since the input modalities are different , it is difficult to leverage source language text successfully . In this work , we focus on sequencelevel knowledge distillation ( SeqKD ) from external text-based NMT models . To leverage the full potential of the source language information , we propose backward SeqKD , SeqKD from a target-to-source backward NMT model . To this end , we train a bilingual E2E-ST model to predict paraphrased transcriptions as an auxiliary task with a single decoder . The paraphrases are generated from the translations in bitext via back-translation . We further propose bidirectional SeqKD in which SeqKD from both forward and backward NMT models is combined . Experimental evaluations on both autoregressive and non-autoregressive models show that SeqKD in each direction consistently improves the translation performance , and the effectiveness is complementary regardless of the model capacity . End-to-end speech translation ( E2E-ST ) ( Bérard et al . , 2016 ) , which aims to convert source speech to text in another language directly , is an active research area . Because direct ST is a more difficult task than automatic speech recognition ( ASR ) and machine translation ( MT ) , various techniques have been proposed to ease the training process by using source transcription . Examples include pretraining ( Bérard et al . , 2018 ; Wang et al . , 2020c ; Bansal et al . , 2019 ; Wang et al . , 2020d ) , multi-task learning ( Weiss et al . , 2017 ; Bérard et al . , 2018 ; Bahar et al . , 2019 ) , knowledge distillation ( Liu et al . , 2019 ) , meta-learning ( Indurthi et al . , 2020 ) , twopass decoding ( Anastasopoulos and Chiang , 2018 ; Sperber et al . , 2019 ) , and interactive decoding ( Liu et al . , 2020 ; Le et al . , 2020 ) . However , as input modalities between ST and MT tasks are different , an auxiliary MT task is not always helpful , especially when additional bitext is not available ( Bahar et al . , 2019 ) . Moreover , because monotonic speech-to-transcription alignments encourage the ASR task to see surface-level local information , an auxiliary ASR task helps the E2E-ST model to extract acoustic representations , not semantic ones , from speech . Sequence-level knowledge distillation ( Se-qKD ) ( Kim and Rush , 2016 ) is another approach to transferring knowledge from one model to another . Recent studies have shown that SeqKD has the effect of reducing the complexity of training data and thus eases the training of student models , e.g. , non-autoregressive ( NAR ) models ( Gu et al . , 2018 ; Zhou et al . , 2019a ; Ren et al . , 2020 ) . Paraphrasing , which represents text in a different form but with the same meaning , can also be regarded as SeqKD when using neural paraphrasing via back-translation ( Mallinson et al . , 2017 ; Wieting et al . , 2017 ; Federmann et al . , 2019 ) . It has been studied to improve the reference diversity for MT system evaluations ( Thompson and Post , 2020 ; Bawden et al . , 2020a , b ) and the performance of low-resource neural MT ( NMT ) models ( Zhou et al . , 2019b ; Khayrallah et al . , 2020 ) . In this work , due to its simplicity and effectiveness , we focus on SeqKD from text-based NMT models to improve the performance of a bilingual E2E-ST model . In order to fully leverage source language information , we propose backward Se-qKD , which targets paraphrased source transcriptions generated from a target-to-source backward NMT model as an auxiliary task . Then , a single ST decoder is trained to predict both source and target language text as in a multilingual setting ( Inaguma et al . , 2019 ) . This way , the decoder is biased to capture semantic representations from speech , un-like joint training with an auxiliary ASR task . We also propose bidirectional SeqKD , which combines SeqKD from two NMT models in both language directions . Therefore , the E2E-ST models can fully exploit the knowledge embedded in both forward and backward NMT models . Experimental evaluations demonstrate that Se-qKD from each direction consistently improves the translation performance of both autoregressive and non-autoregressive E2E-ST models . We also confirm that bidirectional SeqKD outperforms unidirectional SeqKD and that the effectiveness is maintained in large models . To fully leverage knowledge in both source and target language directions for bilingual E2E-ST models , we have proposed bidirectional SeqKD , in which both forward SeqKD from a source-to-target NMT model and backward SeqKD from a target-tosource NMT model are combined . Backward Se-qKD is performed by targeting source paraphrases generated via back-translation from the original translations in bitext . Then , the E2E-ST model is enhanced by training to generate both source and target language text with a single decoder . We experimentally confirmed that SeqKD from each direction boosted the translation performance of both autoregressive and non-autoregressive E2E-ST models , and the effectiveness was additive . Multi-referenced training with the original and distilled text gave further gains . We also showed that bidirectional SeqKD was effective regardless of model sizes .

Existing works on end-to-end speech recognition models use the source transcriptions for performance improvements but it is challenging due to the modality gap.

They propose a bidirectional sequence knowledge distillation which learns from text-based NMT systems with a single decoder to enhance the model to capture semantic representations.

Evaluations on autoregressive and non-autoregressive models show that the proposed method improves in both directions and the results are consistent in different model sizes.

N07-1072

This paper explores the problem of computing text similarity between verb phrases describing skilled human behavior for the purpose of finding approximate matches . Four parsers are evaluated on a large corpus of skill statements extracted from an enterprise-wide expertise taxonomy . A similarity measure utilizing common semantic role features extracted from parse trees was found superior to an information-theoretic measure of similarity and comparable to the level of human agreement . Knowledge-intensive industries need to become more efficient at deploying the right expertise as quickly and smoothly as possible , thus it is desired to have systems that can quickly match and deploy skilled individuals to meet customer needs . The searches in most of the current matching systems are based on exact matches between skill statements . However , exact matching is very likely to miss individuals who are very good matches to the job but did n't select the exact skills that appeared in the open job description . It is always hard for individuals to find the perfect skills to describe their skill sets . For example , an individual might not know whether to choose a skill stating that refers to " maintaining " a given product or " supporting " it or whether to choose a skill about maintaining a " database " or about maintaining " DB2 " . Thus , it is desirable for the job search system to be able to find approximate matches , instead of only exact matches , between available individuals and open job positions . More specifically , a skill similarity computation is needed to allow searches to be expanded to related skills , and return more potential matches . In this paper , we present our work on developing a skill similarity computation based upon semantic commonalities between skill statements . Although there has been much work on text similarity metrics ( Lin , 1998a ; Corley and Mihalcea , 2005 ) , most approaches treat texts as a bag of words and try to find shared words with certain statistical properties based on corpus frequencies . As a result , the structural information in the text is ignored in these approaches . We will describe a new semantic approach that takes the structural information of the text into consideration and matches skill statements on corresponding semantic roles . We will demonstrate that it can outperform standard statistical text similarity techniques , and reach the level of human agreement . In Section 2 , we first describe the skill statements we extracted from an enterprise-wide expertise taxonomy . In Section 3 , we describe the performance of a standard statistical approach on this task . This motivates our semantic approach of matching skill statements on corresponding semantic roles . We also compare and evaluate the performance of four natural language parsers ( the Charniak parser , the Stanford parser , the ESG parser , and MINIPAR ) for the purpose of our task . An inter-rater agreement study and evaluation of our approach will be presented in Section 4 . We end with a discussion and conclusion . In this paper , we have presented our work on a semantic similarity computation for skill statements in natural language . We compared and evaluated four different natural language parsers for our task , and matched skills on their corresponding semantic roles extracted from the parse trees generated by one of these parsers . The evaluation results showed that the skill similarity computation based on semantic role matching can outperform a standard statistical approach and reach the level of human agreement . The extracted semantic role information can also be incorporated into the standard statistical approaches as additional features . One way is to give higher weights to those semantic role features deemed most important . This approach has achieved a high performance for a text categorization task when combining extracted keywords with the full text ( Hulth and Megyesi , 2006 ) . We have shown that good results can be achieved for a domain-specific text matching task by performing a simple word-based feature comparison on corresponding structural elements of texts . We have shown that the structural elements of importance can be identified by domain-specific pattern analysis of corresponding parse trees . We believe this approach can generalize to other domains where phrases , sentences , or other short texts need to be compared .

Existing systems for skilled expertise matching use exact matching between skill statements resulting in missing good matches and calling for a system with approximate matching.

They evaluate four different parsers to take structural information into consideration by matching skill statements on corresponding semantic roles from generated parse trees.

The proposed similarity measure outperforms a standard statistical information-theoretic measure and is comparable to a human agreement.

2020.acl-main.573

Continual relation learning aims to continually train a model on new data to learn incessantly emerging novel relations while avoiding catastrophically forgetting old relations . Some pioneering work has proved that storing a handful of historical relation examples in episodic memory and replaying them in subsequent training is an effective solution for such a challenging problem . However , these memorybased methods usually suffer from overfitting the few memorized examples of old relations , which may gradually cause inevitable confusion among existing relations . Inspired by the mechanism in human long-term memory formation , we introduce episodic memory activation and reconsolidation ( EMAR ) to continual relation learning . Every time neural models are activated to learn both new and memorized data , EMAR utilizes relation prototypes for memory reconsolidation exercise to keep a stable understanding of old relations . The experimental results show that EMAR could get rid of catastrophically forgetting old relations and outperform the state-of-the-art continual learning models . The code and datasets are released on https://github.com / thunlp/ ContinualRE . Relation extraction aims at detecting relations between entities from text , e.g. , extracting the relation " the president of " from the given sentence " Newton served as the president of the Royal Society " , which could serve as external resource for various downstream applications ( Dong et al . , 2015 ; Xiong et al . , 2017 ; Schlichtkrull et al . , 2018 ) . The conventional RE methods ( Riedel et al . , 2013 ; Zeng et al . , 2014 ; Lin et al . , 2016 ) mostly focus on recognizing relations for a fixed pre-defined relation set , and can not handle rapidly emerging novel relations in the real world . Some researchers therefore explore to detect and learn incessantly emerging relations in an open scenario . As shown in Figure 1 , their efforts can be formulated into a two-step pipeline : ( 1 ) Open Relation Learning extracts phrases and arguments to construct patterns of specific relations , and then discovers unseen relation types by clustering patterns , and finally expands sufficient examples of new relation types from large-scale textual corpora ; ( 2 ) Continual Relation Learning continually uses those expanded examples of new relations to train an effective classifier . The classifier is trained on a sequence of tasks for handling both existing and novel relations , where each task has its own relation set . Although continual relation learning is vital for learning emerging relations , there are rare explorations for this field . A straightforward solution is to store all historical data and re-train models every time new relations and examples come in . Nevertheless , it is computationally expensive since relations are in sustainable growth . Moreover , the huge example number of each relation makes frequently mixing new and old examples become infeasible in the real world . Therefore , storing all data is not practical in continual relation learning . In view of this , the recent preliminary work ( Wang et al . , 2019 ) indicates that the main challenge of continual relation learning is the catastrophic forgetting problem , i.e. , it is hard to learn new relations and meanwhile avoid forgetting old relations , considering memorizing all the data is almost impossible . Recent work ( Shin et al . , 2017 ; Kemker and Kanan , 2018 ; Chaudhry et al . , 2019 ) has shown that the memory-based approaches , maintaining episodic memory to save a few training examples in old tasks and re-training memorized examples during training new tasks , are one of the most effective solutions to the catastrophic forgetting problem , especially for continual learning in NLP scenarios ( Wang et al . , 2019 ; d'Autume et al . , 2019 ) . However , existing memory-based models still suffer from an overfitting problem : when adapting them for continual relation learning , they may frequently change feature distribution of old relations , gradually overfit a few examples in memory , and finally become confused among old relations after long-term training . In fact , these memory-based methods are similar to long-term memory model of mammalian memory in neuroscience ( McClelland et al . , 1995 ; Bontempi et al . , 1999 ) . Although researchers in neuroscience are not clear about secrets inside the human brain , they reach a consensus that the formation of long-term memory relies on continually replaying and consolidating information ( Tononi and Cirelli , 2006 ; Boyce et al . , 2016 ; Yang et al . , 2014 ) , corresponding to the episodic memory and memory replay in continual learning models . Yet later work ( Nader et al . , 2000 ; Lee et al . , 2004 ; Alberini , 2005 ) in neuroscience indicates that reactivation of consolidated memory triggers a reconsolidation stage to continually maintain memory , and memory is easy to be changed or erased in this stage . To apply some reconsolidation exercises can help memory go through this stage and keep long-term memory stable . Intuitively , the ex-isting memory-based models seem like continual memory activation without reconsolidation exercises , and thus become sensitive and volatile . Inspired by the reconsolidation mechanism in human long-term memory formation , we introduce episodic memory activation and reconsolidation ( EMAR ) to continual relation learning in this paper . More specifically , when training models on new relations and their examples , we first adopt memory replay to activate neural models on examples of both new relations and memory , and then utilize a special reconsolidation module to let models avoid excessively changing and erasing feature distribution of old relations . As the core of relation learning is to grasp relation prototypes rather than rote memorization of relation examples , our reconsolidation module requires models to be able to distinguish old relation prototypes after each time memory is replayed and activated . As compared with pioneering explorations to improve episodic memory replay ( Chaudhry et al . , 2019 ; Wang et al . , 2019 ) , with toughly keeping feature distribution of old relations invariant , EMAR is more flexible in feature spaces and powerful in remembering relation prototypes . We conduct sufficient experiments on several RE datasets , and the results show that EMAR effectively alleviates the catastrophic forgetting problem and significantly outperforms the stateof-the-art continual learning models . Further experiments and analyses indicate the reasons for the effectiveness of EMAR , proving that it can utilize a few examples in old tasks to reconsolidate old relation prototypes and keep better distinction among old relations after long-term training . To alleviate catastrophically forgetting old relations in continual relation learning , we introduce episodic memory activation and reconsolidation ( EMAR ) , inspired by the mechanism in human long-term memory formation . Compared with existing memory-based methods , EMAR requires models to understand the prototypes of old relations rather than to overfit a few specific memorized examples , which can keep better distinction among relations after long-term training . We conduct experiments on three benchmarks in relation extraction and carry out extensive experimental results as well as empirical analyses , showing the effectiveness of EMAR on utilizing memorized examples . For future work , how to combine open relation learning and continual relation learning together to complete the pipeline for emerging relations still remains a problem , and we will continue to work on it .

Storing histories of examples is shown to be effective for continual relation learning however existing methods suffer from overfitting to memorize a few old memories.

They propose a human memory mechanism inspired by memory activation and reconsolidation method which aims to keep a stable understanding of old relations.

The proposed method mitigates catastrophic forgetting of old relations and achieves state-of-the-art on several relation extraction datasets showing it can use memorized examples.

2021.emnlp-main.66

This paper proposes to study a fine-grained semantic novelty detection task , which can be illustrated with the following example . It is normal that a person walks a dog in the park , but if someone says " A man is walking a chicken in the park , " it is novel . Given a set of natural language descriptions of normal scenes , we want to identify descriptions of novel scenes . We are not aware of any existing work that solves the problem . Although existing novelty or anomaly detection algorithms are applicable , since they are usually topic-based , they perform poorly on our fine-grained semantic novelty detection task . This paper proposes an effective model ( called GAT-MA ) to solve the problem and also contributes a new dataset . Experimental evaluation shows that GAT-MA outperforms 11 baselines by large margins . Novelty or anomaly detection has been an important research topic since 1970s ( Barnett and Lewis , 1994 ) due to numerous applications ( Chalapathy et al . , 2018 ; Pang et al . , 2021 ) . Recently , it has also become important for natural language processing ( NLP ) . Many researchers have studied the problem in the text classification setting ( Fei and Liu , 2016 ; Shu et al . , 2017 ; Xu et al . , 2019 ; Lin and Xu , 2019 ; Zheng et al . , 2020 ) . However , these text novelty classifiers are mainly coarse-grained , working at the document or topic level . Given a text document , their goal is to detect whether the text belongs to a known class or unknown class . This paper introduces a new text novelty detection problem -fine-grained semantic novelty detection . Specifically , given a text description d , we detect whether d represents a semantically novel fact or not . This work considers text data that describe scenes of real-world phenomena in natural language ( NL ) . In our daily lives , we observe different real-world phenomena ( events , activities , situations , etc . ) and often describe these observations ( referred as " scenes " onwards ) in NL to others or write about them . It is quite natural to observe scenes that we have not seen before ( i.e. , novel scenes ) . For example , it is a common scene that " A person walks a dog in the park " , but if someone says " A man is walking a chicken in the park " , it is quite unexpected and novel . Detecting such semantic novelty requires complex conceptual and semantic reasoning over text and thus , is a challenging NLP problem . Note that conceptually , the judgement of the novelty of a scene is subjective and might differ from person to person . However , there are some scenes for which a majority of people have agreement about their novelty . A good example of such majority-view of novelty is the widely-spread meme pictures on social media , which contain novel interactions between objects . In this work , we restrict our research to this majority-based view of novelty and leave the personalized novelty view angle for the future work . In this work , we leverage the captions of images from popular datasets like COCO , Flickr , etc . , to build a semantic novelty detection dataset ( Sec . 3 ) , 1where we consider an image as a scene and the corresponding image captions as different NL descriptions of the scene . Detecting text describing semantically novel observations have many applications , e.g. , recommending novel news , novel images & videos ( based on their text descriptions ) , social media posts and conversations . The problem of semantic novelty detection is defined as follows . Problem Definition : Given a set of natural language descriptions D = { d 1 , d 2 , ..... d n } of common scenes , build a model M using D to score the semantic novelty of a test NL description d with respect to D , i.e. , classifying d into one of the two classes { NORMAL , NOVEL } . " NORMAL " means that d is a description of a common scene and " NOVEL " means d is a description of a se-mantically novel scene . As the detection model M is built only with " NORMAL " class data , the task is an one-class text classification problem . We are unaware of any existing work that can effectively solve this problem . Although existing novelty / anomaly detection and one-class classification algorithms are applicable , since they are coarse-grained or topic-based , they perform poorly on our task ( see Sec . 5 ) . Note that although we focus on solving the problem of semantic novelty detection of NL descriptions of scenes , the proposed task and solution framework are generally applicable to other applications . This paper proposes a new technique , called GAT-MA ( Graph Attention network with Max-Margin loss and knowledge-based contrastive data Augmentation ) to identify NL description sentences of novel scenes . Since our task is at the sentence level and fine-grained , we exploit Graph Attention Network ( GAT ) on the parsed dependency graph of each sentence , which fuses both semantic and syntactic information in the sentence for reasoning with the internal interactions of entities and actions . To enable the model to capture long-range interactions , we stack multiple layers of GATs to build a deep GAT model with multi-hop graph attention . We also create the pseudo novel training data based on the given normal training data through contrastive data augmentation . Thus , GAT-MA is trained with the given original normal scene descriptions and the augmented pseudo novel scene descriptions ( Sec . 4 ) . GAT-MA is evaluated using our newly created Novel Scene Description Detection ( NSD2 ) Dataset . The results show that GAT-MA outperforms a wide range of latest novelty or anomaly detection baselines by very large margins . Our main contributions are as follows : 1 . We propose a new task of semantic novelty detection in text . Whereas the existing work focuses on coarse-grained document-or topiclevel novelty , our task requires fine-grained sentence-level semantic & syntactic analysis . 2 . We propose a highly effective technique called GAT-MA to solve the proposed semantic novelty detection problem , which is based on GAT with dependency parsing and knowledgebased contrastive data augmentation . 3 . We create a new dataset called NSD2 for the proposed task . The dataset can be used as a benchmark dataset by the NLP community . Novelty detection is an important problem because anything novel is of interest . This paper proposed a semantic novelty detection problem and designed a graph attention network based approach ( called GAT-MA ) exploiting parsing and data augmentation to solve the problem . As there is no existing evaluation dataset for the proposed task , an evaluation dataset has been created . Experimental comparisons with a wide range of baselines showed that GAT-MA outperforms them by very large margins .

Existing works on novelty or abnormally detection are coarse-grained only focusing on the document or sentence level as a text classification task.

They propose a fine-grained semantic novelty detection problem where systems detect whether a textual description is a novel fact, coupled with a graph attention-based model.

The proposed model outperforms 11 baseline models on the created dataset from an image caption dataset for the proposed task by large margins.

P12-1096

Long distance word reordering is a major challenge in statistical machine translation research . Previous work has shown using source syntactic trees is an effective way to tackle this problem between two languages with substantial word order difference . In this work , we further extend this line of exploration and propose a novel but simple approach , which utilizes a ranking model based on word order precedence in the target language to reposition nodes in the syntactic parse tree of a source sentence . The ranking model is automatically derived from word aligned parallel data with a syntactic parser for source language based on both lexical and syntactical features . We evaluated our approach on largescale Japanese-English and English-Japanese machine translation tasks , and show that it can significantly outperform the baseline phrasebased SMT system . * This work has been done while the first author was visiting Modeling word reordering between source and target sentences has been a research focus since the emerging of statistical machine translation . In phrase-based models ( Och , 2002 ; Koehn et al . , 2003 ) , phrase is introduced to serve as the fundamental translation element and deal with local reordering , while a distance based distortion model is used to coarsely depict the exponentially decayed word movement probabilities in language translation . Further work in this direction employed lexi-calized distortion models , including both generative ( Koehn et al . , 2005 ) and discriminative ( Zens and Ney , 2006 ; Xiong et al . , 2006 ) variants , to achieve finer-grained estimations , while other work took into account the hierarchical language structures in translation ( Chiang , 2005 ; Galley and Manning , 2008 ) . Long-distance word reordering between language pairs with substantial word order difference , such as Japanese with Subject-Object-Verb ( SOV ) structure and English with Subject-Verb-Object ( SVO ) structure , is generally viewed beyond the scope of the phrase-based systems discussed above , because of either distortion limits or lack of discriminative features for modeling . The most notable solution to this problem is adopting syntax-based SMT models , especially methods making use of source side syntactic parse trees . There are two major categories in this line of research . One is tree-to-string model ( Quirk et al . , 2005 ; Liu et al . , 2006 ) which directly uses source parse trees to derive a large set of translation rules and associated model parameters . The other is called syntax pre-reordering -an approach that re-positions source words to approximate target language word order as much as possible based on the features from source syntactic parse trees . This is usually done in a preprocessing step , and then followed by a standard phrase-based SMT system that takes the re-ordered source sentence as input to finish the translation . In this paper , we continue this line of work and address the problem of word reordering based on source syntactic parse trees for SMT . Similar to most previous work , our approach tries to rearrange the source tree nodes sharing a common parent to mimic the word order in target language . To this end , we propose a simple but effective ranking-based approach to word reordering . The ranking model is automatically derived from the word aligned parallel data , viewing the source tree nodes to be reordered as list items to be ranked . The ranks of tree nodes are determined by their relative positions in the target language -the node in the most front gets the highest rank , while the ending word in the target sentence gets the lowest rank . The ranking model is trained to directly minimize the mis-ordering of tree nodes , which differs from the prior work based on maximum likelihood estimations of reordering patterns ( Li et al . , 2007 ; Genzel , 2010 ) , and does not require any special tweaking in model training . The ranking model can not only be used in a pre-reordering based SMT system , but also be integrated into a phrasebased decoder serving as additional distortion features . We evaluated our approach on large-scale Japanese-English and English-Japanese machine translation tasks , and experimental results show that our approach can bring significant improvements to the baseline phrase-based SMT system in both preordering and integrated decoding settings . In the rest of the paper , we will first formally present our ranking-based word reordering model , then followed by detailed steps of modeling training and integration into a phrase-based SMT system . Experimental results are shown in Section 5 . Section 6 consists of more discussions on related work , and Section 7 concludes the paper . In this paper we present a ranking based reordering method to reorder source language to match the word order of target language given the source side parse tree . Reordering is formulated as a task to rank different nodes in the source side syntax tree according to their relative position in the target language . The ranking model is automatically trained to minimize the mis-ordering of tree nodes in the training data . Large scale experiment shows improvement on both reordering metric and SMT performance , with up to 1.73 point BLEU gain in our evaluation test . In future work , we plan to extend the ranking model to handle reordering between multiple levels of source trees . We also expect to explore better way to integrate ranking reorder model into SMT system instead of a simple penalty scheme . Along the research direction of preprocessing the source language to facilitate translation , we consider to not only change the order of the source language , but also inject syntactic structure of the target language into source language by adding pseudo words into source sentences .

Long distance word reordering remains a challenge for statistical machine translation and existing approaches do it during the preprocessing.

They propose a ranking-based reordering approach where the ranking model is automatically derived from the word aligned parallel data using a syntax parser.

Large scale evaluation of Japanese-English and English-Japanese shows that the proposed approach significantly outperforms the baseline phrase-based statistical machine translation system.

D08-1038

How can the development of ideas in a scientific field be studied over time ? We apply unsupervised topic modeling to the ACL Anthology to analyze historical trends in the field of Computational Linguistics from 1978 to 2006 . We induce topic clusters using Latent Dirichlet Allocation , and examine the strength of each topic over time . Our methods find trends in the field including the rise of probabilistic methods starting in 1988 , a steady increase in applications , and a sharp decline of research in semantics and understanding between 1978 and 2001 , possibly rising again after 2001 . We also introduce a model of the diversity of ideas , topic entropy , using it to show that COLING is a more diverse conference than ACL , but that both conferences as well as EMNLP are becoming broader over time . Finally , we apply Jensen-Shannon divergence of topic distributions to show that all three conferences are converging in the topics they cover . How can we identify and study the exploration of ideas in a scientific field over time , noting periods of gradual development , major ruptures , and the waxing and waning of both topic areas and connections with applied topics and nearby fields ? One important method is to make use of citation graphs ( Garfield , 1955 ) . This enables the use of graphbased algorithms like PageRank for determining researcher or paper centrality , and examining whether their influence grows or diminishes over time . However , because we are particularly interested in the change of ideas in a field over time , we have chosen a different method , following Kuhn ( 1962 ) . In Kuhn 's model of scientific change , science proceeds by shifting from one paradigm to another . Because researchers ' ideas and vocabulary are constrained by their paradigm , successive incommensurate paradigms will naturally have different vocabulary and framing . Kuhn 's model is intended to apply only to very large shifts in scientific thought rather than at the micro level of trends in research foci . Nonetheless , we propose to apply Kuhn 's insight that vocabulary and vocabulary shift is a crucial indicator of ideas and shifts in ideas . Our operationalization of this insight is based on the unsupervised topic model Latent Dirichlet Allocation ( LDA ; Blei et al . ( 2003 ) ) . For many fields , doing this kind of historical study would be very difficult . Computational linguistics has an advantage , however : the ACL Anthology , a public repository of all papers in the Computational Linguistics journal and the conferences and workshops associated with the ACL , COLING , EMNLP , and so on . The ACL Anthology ( Bird , 2008 ) , and comprises over 14,000 documents from conferences and the journal , beginning as early as 1965 through 2008 , indexed by conference and year . This resource has already been the basis of citation analysis work , for example , in the ACL Anthology Network of Joseph and Radev ( 2007 ) . We apply LDA to the text of the papers in the ACL Anthology to induce topics , and use the trends in these topics over time and over conference venues to address questions about the development of the field . Despite the relative youth of our field , computational linguistics has witnessed a number of research trends and shifts in focus . While some trends are obvious ( such as the rise in machine learning methods ) , others may be more subtle . Has the field gotten more theoretical over the years or has there been an increase in applications ? What topics have declined over the years , and which ones have remained roughly constant ? How have fields like Dialogue or Machine Translation changed over the years ? Are there differences among the conferences , for example between COLING and ACL , in their interests and breadth of focus ? As our field matures , it is important to go beyond anecdotal description to give grounded answers to these questions . Such answers could also help give formal metrics to model the differences between the many conferences and venues in our field , which could influence how we think about reviewing , about choosing conference topics , and about long range planning in our field . Our method discovers a number of trends in the field , such as the general increase in applications , the steady decline in semantics , and its possible reversal . We also showed a convergence over time in topic coverage of ACL , COLING , and EMNLP as well an expansion of topic diversity . This growth and convergence of the three conferences , perhaps influenced by the need to increase recall ( Church , 2005 ) seems to be leading toward a tripartite realization of a single new " latent " conference .

How topics or ideas have developed over time in NLP community remains unknown while there are analysis over the ACL anthology citation graph.

They propose to use Latent Dirichlet Allocation for studying topic shift over time and a model to compute the diversity of ideas and topic entropy.

They found that COLING has more diversity than ACL, but all the conferences are becoming to cover more topics over time, and applications increase generally.

P17-1024

In this paper , we aim to understand whether current language and vision ( LaVi ) models truly grasp the interaction between the two modalities . To this end , we propose an extension of the MS-COCO dataset , FOIL-COCO , which associates images with both correct and ' foil ' captions , that is , descriptions of the image that are highly similar to the original ones , but contain one single mistake ( ' foil word ' ) . We show that current LaVi models fall into the traps of this data and perform badly on three tasks : a ) caption classification ( correct vs. foil ) ; b ) foil word detection ; c ) foil word correction . Humans , in contrast , have near-perfect performance on those tasks . We demonstrate that merely utilising language cues is not enough to model FOIL-COCO and that it challenges the state-of-the-art by requiring a fine-grained understanding of the relation between text and image . Most human language understanding is grounded in perception . There is thus growing interest in combining information from language and vision in the NLP and AI communities . So far , the primary testbeds of Language and Vision ( LaVi ) models have been ' Visual Question Answering ' ( VQA ) ( e.g. Antol et al . ( 2015 ) ; Malinowski and Fritz ( 2014 ) ; Malinowski et al . ( 2015 ) ; Gao et al . ( 2015 ) ; Ren et al . ( 2015 ) ) and ' Image Captioning ' ( IC ) ( e.g. Hodosh et al . ( 2013 ) ; Fang et al . ( 2015 ) ; Chen and Lawrence Zitnick ( 2015 ) ; Donahue et al . ( 2015 ) ; Karpathy and Fei-Fei ( 2015 ) ; Vinyals et al . ( 2015 ) ) . Whilst some models have seemed extremely successful on those tasks , it remains unclear how the reported results should be interpreted and what those Figure 1 : Is the caption correct or foil ( T1 ) ? If it is foil , where is the mistake ( T2 ) and which is the word to correct the foil one ( T3 ) ? models are actually learning . There is an emerging feeling in the community that the VQA task should be revisited , especially as many current dataset can be handled by ' blind ' models which use language input only , or by simple concatenation of language and vision features ( Agrawal et al . , 2016 ; Jabri et al . , 2016 ; Zhang et al . , 2016 ; Goyal et al . , 2016a ) . In IC too , Hodosh and Hockenmaier ( 2016 ) showed that , contrarily to what prior research had suggested , the task is far from been solved , since IC models are not able to distinguish between a correct and incorrect caption . Such results indicate that in current datasets , language provides priors that make LaVi models successful without truly understanding and integrating language and vision . But problems do not stop at biases . Johnson et al . ( 2016 ) also point out that current data ' conflate multiple sources of error , making it hard to pinpoint model weaknesses ' , thus highlighting the need for diagnostic datasets . Thirdly , existing IC evaluation metrics are sensitive to n-gram overlap and there is a need for measures that better simulate human judgments ( Hodosh et al . , 2013 ; Elliott and Keller , 2014 ; Anderson et al . , 2016 ) . Our paper tackles the identified issues by proposing an automatic method for creating a large dataset of real images with minimal language bias and some diagnostic abilities . Our dataset , FOIL ( Find One mismatch between Image and Language caption),1 consists of images associated with incorrect captions . The captions are produced by introducing one single error ( or ' foil ' ) per caption in existing , human-annotated data ( Figure 1 ) . This process results in a challenging error-detection / correction setting ( because the caption is ' nearly ' correct ) . It also provides us with a ground truth ( we know where the error is ) that can be used to objectively measure the performance of current models . We propose three tasks based on widely accepted evaluation measures : we test the ability of the system to a ) compute whether a caption is compatible with the image ( T1 ) ; b ) when it is incompatible , highlight the mismatch in the caption ( T2 ) ; c ) correct the mistake by replacing the foil word ( T3 ) . The dataset presented in this paper ( Section 3 ) is built on top of MS-COCO ( Lin et al . , 2014 ) , and contains 297,268 datapoints and 97,847 images . We will refer to it as FOIL-COCO . We evaluate two state-of-the-art VQA models : the popular one by Antol et al . ( 2015 ) , and the attention-based model by Lu et al . ( 2016 ) , and one popular IC model by ( Wang et al . , 2016 ) . We show that those models perform close to chance level , while humans can perform the tasks accurately ( Section 4 ) . Section 5 provides an analysis of our results , allowing us to diagnose three failures of LaVi models . First , their coarse representations of language and visual input do not encode suitably structured information to spot mismatches between an utterance and the corresponding scene ( tested by T1 ) . Second , their language representation is not finegrained enough to identify the part of an utterance that causes a mismatch with the image as it is ( T2 ) . Third , their visual representation is also too poor to spot and name the visual area that corresponds to a captioning error ( T3 ) . We have introduced FOIL-COCO , a large dataset of images associated with both correct and foil captions . The error production is automatically generated , but carefully thought out , making the task of spotting foils particularly challenging . By associating the dataset with a series of tasks , we allow for diagnosing various failures of current LaVi systems , from their coarse understanding of the correspondence between text and vision to their grasp of language and image structure . Our hypothesis is that systems which , like humans , deeply integrate the language and vision modalities , should spot foil captions quite easily . The SoA LaVi models we have tested fall through that test , implying that they fail to integrate the two modalities . To complete the analysis of these results , we plan to carry out a further task , namely ask the system to detect in the image the area that produces the mismatch with the foil word ( the red box around the bird in Figure 1 . ) This extra step would allow us to fully diagnose the failure of the tested systems and confirm what is implicit in our results from task 3 : that the algorithms are unable to map particular elements of the text to their visual counterparts . We note that the addition of this extra step will move this work closer to the textual / visual explanation research ( e.g. , ( Park et al . , 2016 ; Selvaraju et al . , 2016 ) ) . We will then have a pipeline able to not only test whether a mistake can be detected , but also whether the system can explain its decision : ' the wrong word is dog because the cyclists are in fact approaching a bird , there , in the image ' . LaVi models are a great success of recent research , and we are impressed by the amount of ideas , data and models produced in this stimulating area . With our work , we would like to push the community to think of ways that models can better merge language and vision modalites , instead of merely using one to supplement the other .

Despite the success of language and vision models on visual question answering tasks, what these models are learning remains unknown because of coarse-grained datasets.

They propose to automatically inject one mistake to captions in the MS-COCO dataset as a foil word and three diagnostic tasks to study models' behaviors.

Using the introduced dataset, they find that best performing models fail on the proposed tasks indicating their abilities to integrate two modalities.

D17-1323

Language is increasingly being used to define rich visual recognition problems with supporting image collections sourced from the web . Structured prediction models are used in these tasks to take advantage of correlations between co-occurring labels and visual input but risk inadvertently encoding social biases found in web corpora . In this work , we study data and models associated with multilabel object classification and visual semantic role labeling . We find that ( a ) datasets for these tasks contain significant gender bias and ( b ) models trained on these datasets further amplify existing bias . For example , the activity cooking is over 33 % more likely to involve females than males in a training set , and a trained model further amplifies the disparity to 68 % at test time . We propose to inject corpus-level constraints for calibrating existing structured prediction models and design an algorithm based on Lagrangian relaxation for collective inference . Our method results in almost no performance loss for the underlying recognition task but decreases the magnitude of bias amplification by 47.5 % and 40.5 % for multilabel classification and visual semantic role labeling , respectively . Visual recognition tasks involving language , such as captioning ( Vinyals et al . , 2015 ) , visual question answering ( Antol et al . , 2015 ) , and visual semantic role labeling ( Yatskar et al . , 2016 ) , have emerged as avenues for expanding the diversity of information that can be recovered from images . These tasks aim at extracting rich seman-tics from images and require large quantities of labeled data , predominantly retrieved from the web . Methods often combine structured prediction and deep learning to model correlations between labels and images to make judgments that otherwise would have weak visual support . For example , in the first image of Figure 1 , it is possible to predict a spatula by considering that it is a common tool used for the activity cooking . Yet such methods run the risk of discovering and exploiting societal biases present in the underlying web corpora . Without properly quantifying and reducing the reliance on such correlations , broad adoption of these models can have the inadvertent effect of magnifying stereotypes . In this paper , we develop a general framework for quantifying bias and study two concrete tasks , visual semantic role labeling ( vSRL ) and multilabel object classification ( MLC ) . In vSRL , we use the imSitu formalism ( Yatskar et al . , 2016 ( Yatskar et al . , , 2017 ) ) , where the goal is to predict activities , objects and the roles those objects play within an activity . For MLC , we use MS-COCO ( Lin et al . , 2014 ; Chen et al . , 2015 ) , a recognition task covering 80 object classes . We use gender bias as a running example and show that both supporting datasets for these tasks are biased with respect to a gender binary1 . Our analysis reveals that over 45 % and 37 % of verbs and objects , respectively , exhibit bias toward a gender greater than 2:1 . For example , as seen in Figure 1 , the cooking activity in imSitu is a heavily biased verb . Furthermore , we show that after training state-of-the-art structured predictors , models amplify the existing bias , by 5.0 % for vSRL , and 3.6 % in MLC . Each image is paired with a table describing a situation : the verb , cooking , its semantic roles , i.e agent , and noun values filling that role , i.e. woman . In the imSitu training set , 33 % of cooking images have man in the agent role while the rest have woman . After training a Conditional Random Field ( CRF ) , bias is amplified : man fills 16 % of agent roles in cooking images . To reduce this bias amplification our calibration method adjusts weights of CRF potentials associated with biased predictions . After applying our methods , man appears in the agent role of 20 % of cooking images , reducing the bias amplification by 25 % , while keeping the CRF vSRL performance unchanged . To mitigate the role of bias amplification when training models on biased corpora , we propose a novel constrained inference framework , called RBA , for Reducing Bias Amplification in predictions . Our method introduces corpus-level constraints so that gender indicators co-occur no more often together with elements of the prediction task than in the original training distribution . For example , as seen in Figure 1 , we would like noun man to occur in the agent role of the cooking as often as it occurs in the imSitu training set when evaluating on a development set . We combine our calibration constraint with the original structured predictor and use Lagrangian relaxation ( Korte and Vygen , 2008 ; Rush and Collins , 2012 ) to reweigh bias creating factors in the original model . We evaluate our calibration method on imSitu vSRL and COCO MLC and find that in both instances , our models substantially reduce bias amplification . For vSRL , we reduce the average magnitude of bias amplification by 40.5 % . For MLC , we are able to reduce the average magnitude of bias amplification by 47.5 % . Overall , our calibration methods do not affect the performance of the underlying visual system , while substantially reducing the reliance of the system on socially biased correlations 2 . 2 Code and data are available at https://github . com / uclanlp / reducingbias Structured prediction models can leverage correlations that allow them to make correct predictions even with very little underlying evidence . Yet such models risk potentially leveraging social bias in their training data . In this paper , we presented a general framework for visualizing and quantifying biases in such models and proposed RBA to calibrate their predictions under two different settings . Taking gender bias as an example , our analysis demonstrates that conditional random fields can amplify social bias from data while our approach RBA can help to reduce the bias . Our work is the first to demonstrate structured prediction models amplify bias and the first to propose methods for reducing this effect but significant avenues for future work remain . While RBA can be applied to any structured predictor , it is unclear whether different predictors amplify bias more or less . Furthermore , we presented only one method for measuring bias . More extensive analysis could explore the interaction among predictor , bias measurement , and bias deamplification method . Future work also includes applying bias reducing methods in other structured domains , such as pronoun reference resolution ( Mitkov , 2014 ) .

Language is used for visual recognition problems such as captioning to improve performance however it can also encode social biases found in web corpora.

They propose a framework to quantify bias for visual semantic role labelling and multilabel object classification and a constraint inference framework to calibrate existing models.

They find that existing datasets contain gender bias the use of text can amplify it, and the proposed framework can reduce bias without performance loss.

P16-1177

We present a pairwise context-sensitive Autoencoder for computing text pair similarity . Our model encodes input text into context-sensitive representations and uses them to compute similarity between text pairs . Our model outperforms the state-of-the-art models in two semantic retrieval tasks and a contextual word similarity task . For retrieval , our unsupervised approach that merely ranks inputs with respect to the cosine similarity between their hidden representations shows comparable performance with the state-of-the-art supervised models and in some cases outperforms them . Representation learning algorithms learn representations that reveal intrinsic low-dimensional structure in data ( Bengio et al . , 2013 ) . Such representations can be used to induce similarity between textual contents by computing similarity between their respective vectors ( Huang et al . , 2012 ; Silberer and Lapata , 2014 ) . Recent research has made substantial progress on semantic similarity using neural networks ( Rothe and Schütze , 2015 ; Dos Santos et al . , 2015 ; Severyn and Moschitti , 2015 ) . In this work , we focus our attention on deep autoencoders and extend these models to integrate sentential or document context information about their inputs . We represent context information as low dimensional vectors that will be injected to deep autoencoders . To the best of our knowledge , this is the first work that enables integrating context into autoencoders . In representation learning , context may appear in various forms . For example , the context of a current sentence in a document could be either its neighboring sentences ( Lin et al . , 2015 ; Wang and Cho , 2015 ) , topics associated with the sentence ( Mikolov and Zweig , 2012 ; Le and Mikolov , 2014 ) , the document that contains the sentence ( Huang et al . , 2012 ) , as well as their combinations ( Ji et al . , 2016 ) . It is important to integrate context into neural networks because these models are often trained with only local information about their individual inputs . For example , recurrent and recursive neural networks only use local information about previously seen words in a sentence to predict the next word or composition . 1On the other hand , context information ( such as topical information ) often capture global information that can guide neural networks to generate more accurate representations . We investigate the utility of context information in three semantic similarity tasks : contextual word sense similarity in which we aim to predict semantic similarity between given word pairs in their sentential context ( Huang et al . , 2012 ; Rothe and Schütze , 2015 ) , question ranking in which we aim to retrieve semantically equivalent questions with respect to a given test question ( Dos Santos et al . , 2015 ) , and answer ranking in which we aim to rank single-sentence answers with respect to a given question ( Severyn and Moschitti , 2015 ) . The contributions of this paper are as follows : ( 1 ) integrating context information into deep autoencoders and ( 2 ) showing that such integration improves the representation performance of deep autoencoders across several different semantic similarity tasks . Our model outperforms the state-of-the-art su-pervised baselines in three semantic similarity tasks . Furthermore , the unsupervised version of our autoencoder show comparable performance with the supervised baseline models and in some cases outperforms them . 2 Context-sensitive Autoencoders We introduce an effective approach to integrate sentential or document context into deep autoencoders and show that such integration is important in semantic similarity tasks . In the future , we aim to investigate other types of linguistic context ( such as POS tag and word dependency information , word sense , and discourse relations ) and develop a unified representation learning framework that integrates such linguistic context with representation learning models .

Existing approaches for textual representation learning only use local information without contexts which capture global information that can guide neural networks in generating accurate representations.

They propose a pairwise context-sensitive Autoencoder which integrates sentential or document context for computing text pair similarity.

The proposed model outperforms the state-of-the-art models in two retrieval and word similarity tasks and an unsupervised version performs comparable with several supervised baselines.

D09-1066

Distance-based ( windowless ) word assocation measures have only very recently appeared in the NLP literature and their performance compared to existing windowed or frequency-based measures is largely unknown . We conduct a largescale empirical comparison of a variety of distance-based and frequency-based measures for the reproduction of syntagmatic human assocation norms . Overall , our results show an improvement in the predictive power of windowless over windowed measures . This provides support to some of the previously published theoretical advantages and makes windowless approaches a promising avenue to explore further . This study also serves as a first comparison of windowed methods across numerous human association datasets . During this comparison we also introduce some novel variations of window-based measures which perform as well as or better in the human association norm task than established measures . Automatic discovery of semantically associated words has attracted a large amount of attention in the last decades and a host of computational association measures have been proposed to deal with this task ( see Section 2 ) . These measures traditionally rely on the co-ocurrence frequency of two words in a corpus to estimate a relatedness score . There has been a recent emergence of distancebased language modelling techiques in NLP ( Savicki and Hlavacova , 2002 ; Terra and Clarke , 2004 ) in which the number of tokens separating words is the essential quantity . While some of this work has considered distance-based alternatives to conventional association measures ( Hardcastle , 2005 ; Washtell , 2009 ) , there has been no principled empirical evaluation of these measures as predictors of human association . We remedy this by conducting a thorough comparison of a wide variety of frequency-based and distance-based measures as predictors of human association scores as elicited in several different free word association tasks . In this work we focus on first-order association measures as predictors of syntagmatic associations . This is in contrast to second and higher-order measures which are better predictors of paradigmatic associations , or word similarity . The distinction between syntagmatic and paradigmatic relationship types is neither exact nor mutually exclusive , and many paradigmatic relationships can be observed syntagmatically in the text . Roughly in keeping with ( Rapp , 2002 ) , we hereby regard paradigmatic assocations as those based largely on word similarity ( i.e. including those typically classed as synonyms , antonyms , hypernyms , hyponyms etc ) , whereas syntagmatic associations are all those words which strongly invoke one another yet which can not readily be said to be similar . Typically these will have an identifiable semantic or grammatical relationship ( meronym / holonym : stem -flower , verb / object : eat -food etc ) , or may have harder-to-classify topical or idiomatic relationships ( family -Christmas , rock -roll ) . We will show in Section 3.2 that syntagmatic relations by themselves constitute a substantial 25 - 40 % of the strongest human responses to cue words . Although the automatic detection of these assocations in text has received less attention than that of paradigmatic associations , they are nonetheless important in applications such as the resolution of bridging anaphora ( Vieira and Poesio , 2000 ) . 1 Furthermore , first-order associations are often the basis of higher-order vector wordspace models used for predicting paradigmatic relationships : i.e. through the observation of words which share similar sets of syntagmatic associations . Therefore improvements made at the level we are concerned with may reasonably be expected to carry through to applications which hinge on the identification of paradigmatic relationships . After a discussion of previous work in Section 2 , we formulate the exact association measures and parameter settings which we compare in Section 3 , where we also introduce the corpora and human association sets used . Then , by using evaluations similar to those described in ( Baroni et al . , 2008 ) and by Rapp ( 2002 ) , we show that the best distance-based measures correlate better overall with human association scores than do the best window based configurations ( see Section 4 ) , and that they also serve as better predictors of the strongest human associations ( see Section 5 ) . This paper presented the first empirical comparison of window-based and the relatively recently introduced windowless association measures , using their ability to reproduce human association scores as a testbed . We show that the best windowless measures are always at least as good as the best window-based measures , both when it comes to overall correlation with human association scores and predicting the strongest human response . In addition , for several human association sets , they perform significantly better . Although not all parameter settings and corpus sizes could be explored , we conclude that it is worthwhile investigating windowless association measures further . As a side-benefit , we have also introduced new variants of existing frequency-based association measures and shown them to perform as well as or better than their existing counterparts . Although these measures were semi-principled in their construction , a deeper understanding of why they work so well is needed . This may in turn lead to the construction of superior windowless measures . In our own future work , we are especially interested in using higher-order windowless association measures for retrieving paradigmatic relations as well as exploring their use in various NLP applications .

The performance of new windowless word association measures which take the number of tokens separating words into account remains unknown.

They conduct large-scale empirical comparisons of window-based and windowless association measures for the reproduction of syntagmatic human association norms.

The best windowless measures perform on part with best window-based measures on correlation with human association scores.

D09-1042

This paper presents an effective method for generating natural language sentences from their underlying meaning representations . The method is built on top of a hybrid tree representation that jointly encodes both the meaning representation as well as the natural language in a tree structure . By using a tree conditional random field on top of the hybrid tree representation , we are able to explicitly model phrase-level dependencies amongst neighboring natural language phrases and meaning representation components in a simple and natural way . We show that the additional dependencies captured by the tree conditional random field allows it to perform better than directly inverting a previously developed hybrid tree semantic parser . Furthermore , we demonstrate that the model performs better than a previous state-of-the-art natural language generation model . Experiments are performed on two benchmark corpora with standard automatic evaluation metrics . One of the ultimate goals in the field of natural language processing ( NLP ) is to enable computers to converse with humans through human languages . To achieve this goal , two important issues need to be studied . First , it is important for computers to capture the meaning of a natural language sentence in a meaning representation . Second , computers should be able to produce a humanunderstandable natural language sentence from its meaning representation . These two tasks are referred to as semantic parsing and natural language generation ( NLG ) , respectively . In this paper , we use corpus-based statistical methods for constructing a natural language generation system . Given a set of pairs , where each pair consists of a natural language ( NL ) sentence and its formal meaning representation ( MR ) , a learning method induces an algorithm that can be used for performing language generation from other previously unseen meaning representations . A crucial question in any natural language processing system is the representation used . Meaning representations can be in the form of a tree structure . In Lu et al . ( 2008 ) , we introduced a hybrid tree framework together with a probabilistic generative model to tackle semantic parsing , where tree structured meaning representations are used . The hybrid tree gives a natural joint tree representation of a natural language sentence and its meaning representation . A joint generative model for natural language and its meaning representation , such as that used in Lu et al . ( 2008 ) has several advantages over various previous approaches designed for semantic parsing . First , unlike most previous approaches , the generative approach models a simultaneous generation process for both NL and MR . One elegant property of such a joint generative model is that it allows the modeling of both semantic parsing and natural language generation within the same process . Second , the generative process proceeds as a recursive top-down Markov process in a way that takes advantage of the tree structure of the MR . The hybrid tree generative model proposed in Lu et al . ( 2008 ) was shown to give stateof-the-art accuracy in semantic parsing on benchmark corpora . While semantic parsing with hybrid trees has been studied in Lu et al . ( 2008 ) , its inverse task -NLG with hybrid trees -has not yet been explored . We believe that the properties that make the hybrid trees effective for semantic parsing also make them effective for NLG . In this paper , we develop systems for the generation task by building on top of the generative model introduced in Lu et al . ( 2008 ) ( referred to as the LNLZ08 system ) . We first present a baseline model by directly " inverting " the LNLZ08 system , where an NL sentence is generated word by word . We call this model the direct inversion model . This model is unable to model some long range global dependencies over the entire NL sentence to be generated . To tackle several weaknesses exhibited by the baseline model , we next introduce an alternative , novel model that performs generation at the phrase level . Motivated by conditional random fields ( CRF ) ( Lafferty et al . , 2001 ) , a different parameterization of the conditional probability of the hybrid tree that enables the model to encode some longer range dependencies amongst phrases and MRs is used . This novel model is referred to as the tree CRF-based model . Evaluation results for both models are presented , through which we demonstrate that the tree CRF-based model performs better than the direct inversion model . We also compare the tree CRFbased model against the previous state-of-the-art model of Wong and Mooney ( 2007 ) . Furthermore , we evaluate our model on a dataset annotated with several natural languages other than English ( Japanese , Spanish , and Turkish ) . Evaluation results show that our proposed tree CRF-based model outperforms the previous model . In this paper , we presented two novel models for the task of generating natural language sentences from given meaning representations , under a hybrid tree framework . We first built a simple direct inversion model as a baseline . Next , to address the limitations associated with the direct inversion model , a tree CRF-based model was introduced . We evaluated both models on standard benchmark corpora . Evaluation results show that the tree CRF-based model performs better than the direct inversion model , and that the tree CRF-based model also outperforms WASP -1 + + , which was a previous state-of-the-art system reported in the literature .

While hybrid trees are shown to be effective for semantic parsing, their application for text generation is under explored.

They propose a phrase-level tree conditional random field that uses a hybrid tree of a meaning representation for the text generation model.

Experiments in four languages with automatic evaluation metrics show that the proposed conditional random field-based model outperforms the previous state-of-the-art system.

P98-1081

In this paper we examine how the differences in modelling between different data driven systems performing the same NLP task can be exploited to yield a higher accuracy than the best individual system . We do this by means of an experiment involving the task of morpho-syntactic wordclass tagging . Four well-known tagger generators ( Hidden Markov Model , Memory-Based , Transformation Rules and Maximum Entropy ) are trained on the same corpus data . After comparison , their outputs are combined using several voting strategies and second stage classifiers . All combination taggers outperform their best component , with the best combination showing a 19.1 % lower error rate than the best individual tagger . In all Natural Language Processing ( NLP ) systems , we find one or more language models which are used to predict , classify and/or interpret language related observations . Traditionally , these models were categorized as either rule-based / symbolic or corpusbased / probabilistic . Recent work ( e.g. Brill 1992 ) has demonstrated clearly that this categorization is in fact a mix-up of two distinct Categorization systems : on the one hand there is the representation used for the language model ( rules , Markov model , neural net , case base , etc . ) and on the other hand the manner in which the model is constructed ( hand crafted vs. data driven ) . Data driven methods appear to be the more popular . This can be explained by the fact that , in general , hand crafting an explicit model is rather difficult , especially since what is being modelled , natural language , is not ( yet ) wellunderstood . When a data driven method is used , a model is automatically learned from the implicit structure of an annotated training corpus . This is much easier and can quickly lead to a model which produces results with a ' reasonably ' good quality . Obviously , ' reasonably good quality ' is not the ultimate goal . Unfortunately , the quality that can be reached for a given task is limited , and not merely by the potential of the learning method used . Other limiting factors are the power of the hard-and software used to implement the learning method and the availability of training material . Because of these limitations , we find that for most tasks we are ( at any point in time ) faced with a ceiling to the quality that can be reached with any ( then ) available machine learning system . However , the fact that any given system can not go beyond this ceiling does not mean that machine learning as a whole is similarly limited . A potential loophole is that each type of learning method brings its own ' inductive bias ' to the task and therefore different methods will tend to produce different errors . In this paper , we are concerned with the question whether these differences between models can indeed be exploited to yield a data driven model with superior performance . In the machine learning literature this approach is known as ensemble , stacked , or combined classifiers . It has been shown that , when the errors are uncorrelated to a sufficient degree , the resulting combined classifier will often perform better than all the individual systems ( Ali and Pazzani 1996 ; Chan and Stolfo 1995 ; Tumer and Gosh 1996 ) . The underlying assumption is twofold . First , the combined votes will make the system more robust to the quirks of each learner 's particular bias . Also , the use of information about each individual method 's behaviour in principle even admits the possibility to fix collective errors . We will execute our investigation by means of an experiment . The NLP task used in the experiment is morpho-syntactic wordclass tagging . The reasons for this choice are several . First of all , tagging is a widely researched and well-understood task ( cf . van Halteren ( ed . ) 1998 ) . Second , current performance levels on this task still leave room for improvement : ' state of the art ' performance for data driven automatic wordclass taggers ( tagging English text with single tags from a low detail tagset ) is 96 - 97 % correctly tagged words . Finally , a number of rather different methods are available that generate a fully functional tagging system from annotated text . Our experiment shows that , at least for the task at hand , combination of several different systems allows us to raise the performance ceiling for data driven systems . Obviously there is still room for a closer examination of the differences between the combination methods , e.g. the question whether Memory-Based combination would have performed better if we had provided more training data than just Tune , and of the remaining errors , e.g. the effects of inconsistency in the data ( cf . Ratnaparkhi 1996 on such effects in the Penn Treebank corpus ) . Regardless of such closer investigation , we feel that our results are encouraging enough to extend our investigation of combination , starting with additional component taggers and selection strategies , and going on to shifts to other tagsets and/or languages . But the investigation need not be limited to wordclass tagging , for we expect that there are many other NLP tasks where combination could lead to worthwhile improvements .

Different data driven approaches tend to produce different errors and their qualities are limited due to the learning method and available training material.

They propose to combine four different modelling methods for the task of morpho-syntactic wordclass tagging by using several voting strategies and second stage classifiers.

All combinations outperform the best component, with the best one showing a 19.1% lower error rate and raising the performance ceiling.

2020.emnlp-main.500

Adversarial attacks for discrete data ( such as texts ) have been proved significantly more challenging than continuous data ( such as images ) since it is difficult to generate adversarial samples with gradient-based methods . Current successful attack methods for texts usually adopt heuristic replacement strategies on the character or word level , which remains challenging to find the optimal solution in the massive space of possible combinations of replacements while preserving semantic consistency and language fluency . In this paper , we propose BERT-Attack , a high-quality and effective method to generate adversarial samples using pre-trained masked language models exemplified by BERT . We turn BERT against its fine-tuned models and other deep neural models in downstream tasks so that we can successfully mislead the target models to predict incorrectly . Our method outperforms state-of-theart attack strategies in both success rate and perturb percentage , while the generated adversarial samples are fluent and semantically preserved . Also , the cost of calculation is low , thus possible for large-scale generations . The code is available at https://github.com/ LinyangLee / BERT-Attack . Despite the success of deep learning , recent works have found that these neural networks are vulnerable to adversarial samples , which are crafted with small perturbations to the original inputs ( Goodfellow et al . , 2014 ; Kurakin et al . , 2016 ; Chakraborty et al . , 2018 ) . That is , these adversarial samples are imperceptible to human judges while they can mislead the neural networks to incorrect predictions . Therefore , it is essential to explore these adversarial attack methods since the ultimate goal is to make sure the neural networks are highly reliable and robust . While in computer vision fields , both attack strategies and their defense countermeasures are well-explored ( Chakraborty et al . , 2018 ) , the adversarial attack for text is still challenging due to the discrete nature of languages . Generating of adversarial samples for texts needs to possess such qualities : ( 1 ) imperceptible to human judges yet misleading to neural models ; ( 2 ) fluent in grammar and semantically consistent with original inputs . Previous methods craft adversarial samples mainly based on specific rules ( Li et al . , 2018 ; Gao et al . , 2018 ; Yang et al . , 2018 ; Alzantot et al . , 2018 ; Ren et al . , 2019 ; Jin et al . , 2019 ; Zang et al . , 2020 ) . Therefore , these methods are difficult to guarantee the fluency and semantically preservation in the generated adversarial samples at the same time . Plus , these manual craft methods are rather complicated . They use multiple linguistic constraints like NER tagging or POS tagging . Introducing contextualized language models to serve as an automatic perturbation generator could make these rules designing much easier . The recent rise of pre-trained language models , such as BERT ( Devlin et al . , 2018 ) , push the performances of NLP tasks to a new level . On the one hand , the powerful ability of a fine-tuned BERT on downstream tasks makes it more challenging to be adversarial attacked ( Jin et al . , 2019 ) . On the other hand , BERT is a pre-trained masked language model on extremely large-scale unsupervised data and has learned general-purpose language knowledge . Therefore , BERT has the potential to generate more fluent and semantic-consistent substitutions for an input text . Naturally , both the properties of BERT motivate us to explore the possibility of attacking a fine-tuned BERT with another BERT as the attacker . In this paper , we propose an effective and high-quality adversarial sample generation method : BERT-Attack , using BERT as a language model to generate adversarial samples . The core algorithm of BERT-Attack is straightforward and consists of two stages : finding the vulnerable words in one given input sequence for the target model ; then applying BERT in a semantic-preserving way to generate substitutes for the vulnerable words . With the ability of BERT , the perturbations are generated considering the context around . Therefore , the perturbations are fluent and reasonable . We use the masked language model as a perturbation generator and find perturbations that maximize the risk of making wrong predictions ( Goodfellow et al . , 2014 ) . Differently from previous attacking strategies that require traditional single-direction language models as a constraint , we only need to inference the language model once as a perturbation generator rather than repeatedly using language models to score the generated adversarial samples in a trial and error process . Experimental results show that the proposed BERT-Attack method successfully fooled its finetuned downstream model with the highest attack success rate compared with previous methods . Meanwhile , the perturb percentage and the query number are considerably lower , while the semantic preservation is high . To summarize our main contributions : • We propose a simple and effective method , named BERT-Attack , to effectively generate fluent and semantically-preserved adversarial samples that can successfully mislead stateof-the-art models in NLP , such as fine-tuned BERT for various downstream tasks . • BERT-Attack has a higher attacking success rate and a lower perturb percentage with fewer access numbers to the target model compared with previous attacking algorithms , while does not require extra scoring models therefore extremely effective . In this work , we propose a high-quality and effective method BERT-Attack to generate adversarial samples using BERT masked language model . Experiment results show that the proposed method achieves a high success rate while maintaining a minimum perturbation . Nevertheless , candidates generated from the masked language model can sometimes be antonyms or irrelevant to the original words , causing a semantic loss . Thus , enhancing language models to generate more semantically related perturbations can be one possible solution to perfect BERT-Attack in the future .

Generating adversarial samples with gradient-based methods for text data is because of its discrete nature and existing complicated heuristic-based methods suffer from finding optimal solutions.

They propose to use BERT to generate adversarial samples by first finding the valuable words and generating substitutes for these words in a semantic-preserving way.

The proposed method outperforms state-of-the-art methods in success rate and perturb percentage while preserving fluency and sematic of generated samples with low cost.

E17-1060

We investigate the generation of onesentence Wikipedia biographies from facts derived from Wikidata slot-value pairs . We train a recurrent neural network sequence-to-sequence model with attention to select facts and generate textual summaries . Our model incorporates a novel secondary objective that helps ensure it generates sentences that contain the input facts . The model achieves a BLEU score of 41 , improving significantly upon the vanilla sequence-to-sequence model and scoring roughly twice that of a simple template baseline . Human preference evaluation suggests the model is nearly as good as the Wikipedia reference . Manual analysis explores content selection , suggesting the model can trade the ability to infer knowledge against the risk of hallucinating incorrect information . Despite massive effort , Wikipedia and other collaborative knowledge bases ( KBs ) have coverage and quality problems . Popular topics are covered in great detail , but there is a long tail of specialist topics with little or no text . Other text can be incorrect , whether by accident or vandalism . We report on the task of generating textual summaries for people , mapping slot-value facts to onesentence encyclopaedic biographies . In addition to initialising stub articles with only structured data , the resulting model could be used to improve consistency and accuracy of existing articles . Figure 1 shows a Wikidata entry for Mathias Tuomi , with fact keys and values flattened into a sequence , and the first sentence from his Wikipedia article . Some values are in the text , others are missing TITLE mathias tuomi SEX OR GENDER male DATE OF BIRTH 1985 - 09 - 03 OCCUPATION squash player CITIZENSHIP finland ( e.g. male ) or expressed differently ( e.g. dates ) . We treat this knowlege-to-text task like translation , using a recurrent neural network ( RNN ) sequence-to-sequence model ( Sutskever et al . , 2014 ) that learns to select and realise the most salient facts as text . This includes an attention mechanism to focus generation on specific facts , a shared vocabulary over input and output , and a multi-task autoencoding objective for the complementary extraction task . We create a reference dataset comprising more than 400,000 knowledgetext pairs , handling the 15 most frequent slots . We also describe a simple template baseline for comparison on BLEU and crowd-sourced human preference judgements over a heldout TEST set . Our model obtains a BLEU score of 41.0 , compared to 33.1 without the autoencoder and 21.1 for the template baseline . In a crowdsourced preference evaluation , the model outperforms the baseline and is preferred 40 % of the time to the Wikipedia reference . Manual analysis of content selection suggests that the model can infer knowledge but also makes mistakes , and that the autoencoding objective encourages the model to select more facts without increasing sentence length . The task formulation and models are a foundation for text completion and consistency in KBs . We present a neural model for mapping between structured and unstructured data , focusing on creating Wikipedia biographic summary sentences from Wikidata slot-value pairs . We introduce a sequence-to-sequence autoencoding RNN which improves upon base models by jointly learning to generate text and reconstruct facts . Our analysis of the task suggests evaluation in this domain is challenging . In place of a single score , we analyse statistical measures , human preference judgements and manual annotation to help characterise the task and understand system performance . In the human preference evaluation , our best model outperforms template baselines and is preferred 40 % of the time to the gold standard Wikipedia reference . Code and data is available at https:// github.com / andychisholm / mimo .

Wikipedia and other collaborative knowledge bases have coverage and quality issues especially on a long tail of specialist topics.

They propose a recurrent neural network sequence-to-sequence model with an attention mechanism trained on a multi-task autoencoding objective to generate one-sentence Wikipedia biographies from Wikidata.

The proposed model achieves 41 BLEU score outperforming the baseline model and human annotators prefer the 40% of outputs as good as Wikipedia gold references.

D08-1050

Most state-of-the-art wide-coverage parsers are trained on newspaper text and suffer a loss of accuracy in other domains , making parser adaptation a pressing issue . In this paper we demonstrate that a CCG parser can be adapted to two new domains , biomedical text and questions for a QA system , by using manually-annotated training data at the POS and lexical category levels only . This approach achieves parser accuracy comparable to that on newspaper data without the need for annotated parse trees in the new domain . We find that retraining at the lexical category level yields a larger performance increase for questions than for biomedical text and analyze the two datasets to investigate why different domains might behave differently for parser adaptation . Most state-of-the-art wide-coverage parsers are based on the Penn Treebank ( Marcus et al . , 1993 ) , making such parsers highly tuned to newspaper text . A pressing question facing the parsing community is how to adapt these parsers to other domains , such as biomedical research papers and web pages . A related question is how to improve the performance of these parsers on constructions that are rare in the Penn Treebank , such as questions . Questions are particularly important since a question parser is a component in most Question Answering ( QA ) systems ( Harabagiu et al . , 2001 ) . In this paper we investigate parser adaptation in the context of lexicalized grammars , by using a parser based on Combinatory Categorial Grammar ( CCG ) ( Steedman , 2000 ) . A key property of CCG is that it is lexicalized , meaning that each word in a sentence is associated with an elementary syntactic structure . In the case of CCG this is a lexical category expressing subcategorization information . We exploit this property of CCG by performing manual annotation in the new domain , but only up to this level of representation , where the annotation can be carried out relatively quickly . Since CCG lexical categories are so expressive , many of the syntactic characteristics of a domain are captured at this level . The two domains we consider are the biomedical domain and questions for a QA system . We use the term " domain " somewhat loosely here , since questions are best described as a particular set of syntactic constructions , rather than a set of documents about a particular topic . However , we consider question data to be interesting in the context of domain adaptation for the following reasons : 1 ) there are few examples in the Penn Treebank ( PTB ) and so PTB parsers typically perform poorly on them ; 2 ) questions form a fairly homogeneous set with respect to the syntactic constructions employed , and it is an interesting question how easy it is to adapt a parser to such data ; and 3 ) QA is becoming an important example of NLP technology , and question parsing is an important task for QA systems . The CCG parser we use ( Clark and Curran , 2007b ) makes use of three levels of representation : one , a POS tag level based on the fairly coarse-grained POS tags in the Penn Treebank ; two , a lexical category level based on the more fine-grained CCG lexical categories , which are assigned to words by a CCG su-pertagger ; and three , a hierarchical level consisting of CCG derivations . A key idea in this paper , following a pilot study in Clark et al . ( 2004 ) , is to perform manual annotation only at the first two levels . Since the lexical category level consists of sequences of tags , rather than hierarchical derivations , the annotation can be performed relatively quickly . For the biomedical and question domains we manually annotated approximately 1,000 and 2,000 sentences , respectively , with CCG lexical categories . We also created a gold standard set of grammatical relations ( GR ) in the Stanford format ( de Marneffe et al . , 2006 ) , using 500 of the questions . For the biomedical domain we used the BioInfer corpus ( Pyysalo et al . , 2007a ) , an existing gold-standard GR resource also in the Stanford format . We evaluated the parser on both lexical category assignment and recovery of GRs . The results show that the domain adaptation approach used here is successful in two very different domains , achieving parsing accuracy comparable to state-of-the-art accuracy for newspaper text . The results also show , however , that the two domains have different profiles with regard to the levels of representation used by the parser . We find that simply retraining the POS tagger used by the parser leads to a large improvement in performance for the biomedical domain , and that retraining the CCG supertagger on the annotated biomedical data improves the performance further . For the question data , retraining just the POS tagger also improves parser performance , but retraining the supertagger has a much greater effect . We perform some analysis of the two datasets in order to explain the different behaviours with regard to porting the CCG parser . We have targeted lower levels of representation in order to adapt a lexicalized-grammar parser to two new domains , biomedical text and questions . Although each of the lower levels has been targeted independently in previous work , this is the first study that examines both levels together to determine how they affect parsing accuracy . We achieved an accuracy on grammatical relations in the same range as that of the original parser for newspaper text , without requiring costly annotation of full parse trees . Both biomedical and question data are domains in which there is an immediate need for accurate parsing . The question dataset is in some ways an extreme example for domain adaptation , since the sentences are syntactically uniform ; on the other hand , it is of interest as a set of constructions where the parser initially performed poorly , and is a realistic parsing challenge in the context of QA systems . Interestingly , although an increase in accuracy at each stage of the pipeline did yield an increase at the following stage , these increases were not uniform across the two domains . The new POS tagger model was responsible for most of the improvement in parsing for the biomedical domain , while the new supertagger model was necessary to see a large improvement in the question domain . We attribute this to the fact that question syntax is significantly different from newspaper syntax . We expect these considerations to apply to any lexicalized-grammar parser . Of course , it would be useful to have a way of predicting which level of annotation would be most effective for adapting to a new domain before the annotation begins . The utility of measures such as unknown word rate ( which can be performed with unlabelled data ) and unknown POS n-gram rate ( which can be performed with only POS tags ) is not yet sufficiently clear to rely on them as predictive measures , but it seems a fruitful avenue for future work to investigate the importance of such measures for parser domain adaptation .

Most existing parsers are tuned for newspaper texts making them limited in applicable domains.

They propose a method to adapt a CCG parser to new domains using manually-annotated data only at POS and lexical category levels.

The proposed method achieves comparable results to in-domain parsers without expensive full annotations on biomedical texts and questions that are rare in existing benchmark datasets.