{ "paper_id": "2019", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:30:12.966452Z" }, "title": "Multi-linguality Helps: Event-Argument Extraction for Disaster Domain in Cross-lingual and Multi-lingual Setting", "authors": [ { "first": "Zishan", "middle": [], "last": "Ahmad", "suffix": "", "affiliation": { "laboratory": "", "institution": "Indian Institute of Technology", "location": { "addrLine": "Patna {1821cs18, 1821cs13" } }, "email": "" }, { "first": "Deeksha", "middle": [], "last": "Varshney", "suffix": "", "affiliation": { "laboratory": "", "institution": "Indian Institute of Technology", "location": { "addrLine": "Patna {1821cs18, 1821cs13" } }, "email": "" }, { "first": "Asif", "middle": [], "last": "Ekbal", "suffix": "", "affiliation": { "laboratory": "", "institution": "Indian Institute of Technology", "location": { "addrLine": "Patna {1821cs18, 1821cs13" } }, "email": "" }, { "first": "Pushpak", "middle": [], "last": "Bhattacharyya", "suffix": "", "affiliation": { "laboratory": "", "institution": "Indian Institute of Technology", "location": { "addrLine": "Patna {1821cs18, 1821cs13" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Automatic extraction of disaster-related events and their arguments from natural language text is vital for building a decision support system for crisis management. Event extraction from various news sources is a well-explored area for this objective. However, extracting events alone, without any context provides only partial help for this purpose. Extracting related arguments like Time, Place, Casualties, etc., provides a complete picture of the disaster event. In this paper, we create a disaster domain dataset in Hindi by annotating disasterrelated event and arguments. We also obtain equivalent datasets for Bengali and English from a collaboration. We build a multilingual deep learning model for argument extraction in all the three languages. We also compare our multilingual system with a similar baseline monolingual system trained for each language separately. It is observed that a single multilingual system is able to compensate for lack of training data, by using joint training of dataset from different languages in shared space, thus giving a better overall result.", "pdf_parse": { "paper_id": "2019", "_pdf_hash": "", "abstract": [ { "text": "Automatic extraction of disaster-related events and their arguments from natural language text is vital for building a decision support system for crisis management. Event extraction from various news sources is a well-explored area for this objective. However, extracting events alone, without any context provides only partial help for this purpose. Extracting related arguments like Time, Place, Casualties, etc., provides a complete picture of the disaster event. In this paper, we create a disaster domain dataset in Hindi by annotating disasterrelated event and arguments. We also obtain equivalent datasets for Bengali and English from a collaboration. We build a multilingual deep learning model for argument extraction in all the three languages. We also compare our multilingual system with a similar baseline monolingual system trained for each language separately. It is observed that a single multilingual system is able to compensate for lack of training data, by using joint training of dataset from different languages in shared space, thus giving a better overall result.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "The ability to extract real time news of disaster events automatically, can potentially help in better decision-making for planning and coordination of disaster relief efforts. Event extraction from text entails the extraction of particular types of events along with their arguments. Information obtained from extracted event mentions provides a more structured and clear picture when augmented with related arguments like Time, Place, Participant, Casualty etc. In a language rich world where each event is documented in multiple languages, argument extraction in multi-lingual setting stands as a crucial task.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Extraction of events from news is a well explored area in Natural Language Processing. Com-petitions such as ACE2005 (Doddington et al., 2004) and TAC-KBP2015 (Mitamura et al., 2015) have investigated the area and provided a large body of literature on event extraction from news articles. Event extraction was done on ACE2005 dataset by Ji and Grishman (2008) by combining global evidence from related documents with local decisions. Hou et al. (2012) introduced a method of event argument extraction based on CRFs model for ACE 2005 Chinese event corpus. Event and its arguments were extracted by Petroni et al. (2018) , for the purpose of extracting breaking news. Although extraction of events is quite well examined, there is a scarcity of work in extraction of detailed arguments for disaster domain like casualties, reason, after-effects etc. In this paper we create and publish a dataset annotated for events in disaster domain, for three different languages, i). Hindi, ii). Bengali and iii). English. This dataset is annotated for the task of argument extraction by expert annotators. We build a 'mono-lingual' deep learning system, based on CNN (Convolutional Neural Network) and Bi-LSTM (Bi-Directional Long Short Term Memory) for the task of argument extraction. In order to leverage the information from all the languages while training, and improve the performance of the system, we build a 'multi-lingual' argument extraction system. This is done by adding separate language layers for each language to our 'mono-lingual' system. To bring the datasets of all the languages to the same vector space, we make use of 'multi-lingual' word embeddings. We show that by training our model in this way we are able to utilize the dataset of all the three languages and improve the performance of our system for most arguments in the three languages. We also investigate how the syntactic difference of the languages is handled by our system. Through analysis, we show that 'multi-lingual' training is espe-cially helpful in improving the performance when some argument is under-represented in the 'monolingual' training data.", "cite_spans": [ { "start": 117, "end": 142, "text": "(Doddington et al., 2004)", "ref_id": "BIBREF6" }, { "start": 159, "end": 182, "text": "(Mitamura et al., 2015)", "ref_id": "BIBREF19" }, { "start": 338, "end": 360, "text": "Ji and Grishman (2008)", "ref_id": "BIBREF12" }, { "start": 435, "end": 452, "text": "Hou et al. (2012)", "ref_id": "BIBREF10" }, { "start": 599, "end": 620, "text": "Petroni et al. (2018)", "ref_id": "BIBREF26" }, { "start": 823, "end": 849, "text": "reason, after-effects etc.", "ref_id": null }, { "start": 1152, "end": 1186, "text": "CNN (Convolutional Neural Network)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Argument extraction entails classifying each word in the sentence into some argument or not argument. Therefore, it has been formulated as a sequence labelling task. Given a sentence of form w 1 , w 2 , ..., w n , the task is to predict the sequence of event-arguments, of the form l 1 , l 2 , ..., l n . Six different types of arguments were annotated in the dataset: i). Place, ii). Time, iii). Reason, iv). Casualties, v). Participant and vi). After-effects. To label multi-word event-arguments, IOB-style encoding is used where B, I and O denote the beginning, intermediate and outside token of an event.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Definition", "sec_num": "1.1" }, { "text": "\u2022 Input Hindi Sentence: \u0917\u0943 \u0939 \u092e\u0902 \u093e\u0932\u092f \u092e\u0941 \u0902 \u092c\u0908 \u0915\u0947 \u092c\u092e \u0935\u094d \u092b\u094b\u091f \u0915\u0947 \u092e \u0947 \u0928\u091c\u0930 \u0907\u0938 \u092c\u093e\u0924 \u0915 \u0935\u0936\u0947 \u0937 \u0924\u094c\u0930 \u092a\u0930 \u091c\u093e\u0902 \u091a \u0915\u0930 \u0930\u0939\u093e \u0939\u0948 \u0915 \u0905 \u0930\u0927\u093e\u092e \u092e\u0902 \u0926\u0930 \u0914\u0930 \u0967\u096f\u096f\u0969 \u0915\u0947 \u092e\u0941 \u0902 \u092c\u0908 \u092c\u092e \u0935-\u092b\u094b\u091f \u0915\u0947 \u092b\u0938\u0932 \u0915 \u0924 \u092f\u093e \u0915\u0947 \u092a \u092e \u0924\u094b \u092f\u0939 \u0939\u092e\u0932\u0947 \u0928\u0939 \u090f", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Definition", "sec_num": "1.1" }, { "text": "\u2022 Translation: In view of the Mumbai bomb blasts, the Home Ministry is specially investigating the fact that these attacks did not take place as response to the Akshardham Temple and the 1993 Bombay bomb blasts", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Definition", "sec_num": "1.1" }, { "text": "\u2022 Output: O O I_Place O O O O O O O O O O O O O O O O I_Place I_Place O I_Time O I_Place O O O O O O O O O O O O O O", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Problem Definition", "sec_num": "1.1" }, { "text": "A major task in information extraction is detection of event triggers, event classification and event argument extraction. Recent works on event trigger detection and classification discuss efficient feature representation techniques which can help in event extraction. Nguyen and Grishman (2015) proposed a convolutional neural network for event extraction which automatically learns features from text. Chen et al. (2015) Previously, in event argument extraction researchers have experimented with pattern based methods (Patwardhan and Riloff, 2007; Chambers and Jurafsky, 2011) and machine learning based methods (Patwardhan and Riloff, 2009; Lu and Roth, 2012) most of which utilise the various kinds of features obtained from the context of a sentence. Higher level representations such as crosssentence or cross-event information were also explored by Hong et al. (2011) and Huang and Riloff (2011) . Maximum Entropy based classifiers were applied for event and argument labeling by Ahn (2006) ; Chen and Ji (2009) ; Zhao et al. (2008) . The disadvantage with ME classifier is that it gets stuck in local optima and fails to fully capture the context features. To overcome this Hou et al. 2012proposes a event argument extraction system based on Conditional Random Fields (CRF) model that can select any features and normalizing these features in overall situation helps in obtaining optimal results. While, these models can get affected by the error propagated from upstream tasks, a joint model can help us utilise the close interaction between one or more similar tasks. Li et al. (2013) presented a joint model for Chinese Corpus which identifies arguments and determines their roles for event extraction using various kinds of discourse-level information. On ACE2005 dataset Sha et al. (2018) proposed a dependency bridge recurrent neural network (dbRNN) built upon LSTM units for event extraction. They use dependency bridges over Bi-LSTM to join syntactically similar words. A tensor layer is applied to get the various argument-argument interactions. Event triggers and arguments are then jointly extracted utilising a max-margin criterion. Nguyen et al. (2016) presented a GRU model to jointly predict events and its arguments.", "cite_spans": [ { "start": 405, "end": 423, "text": "Chen et al. (2015)", "ref_id": "BIBREF3" }, { "start": 522, "end": 551, "text": "(Patwardhan and Riloff, 2007;", "ref_id": "BIBREF24" }, { "start": 552, "end": 580, "text": "Chambers and Jurafsky, 2011)", "ref_id": "BIBREF2" }, { "start": 616, "end": 645, "text": "(Patwardhan and Riloff, 2009;", "ref_id": "BIBREF25" }, { "start": 646, "end": 664, "text": "Lu and Roth, 2012)", "ref_id": "BIBREF18" }, { "start": 858, "end": 876, "text": "Hong et al. (2011)", "ref_id": "BIBREF9" }, { "start": 881, "end": 904, "text": "Huang and Riloff (2011)", "ref_id": "BIBREF11" }, { "start": 989, "end": 999, "text": "Ahn (2006)", "ref_id": "BIBREF0" }, { "start": 1002, "end": 1020, "text": "Chen and Ji (2009)", "ref_id": "BIBREF4" }, { "start": 1023, "end": 1041, "text": "Zhao et al. (2008)", "ref_id": "BIBREF29" }, { "start": 1580, "end": 1596, "text": "Li et al. (2013)", "ref_id": "BIBREF15" }, { "start": 1786, "end": 1803, "text": "Sha et al. (2018)", "ref_id": "BIBREF28" } ], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "We introduce two systems for the task of event argument extraction. First is our monolingual system built using CNN (Convolutional Neural Network) and Bi-LSTM (Bi-Directional Long Short Term Memory). To exploit the information from related languages, we develop a second system that can use information from all the languages for training. This multi-lingual system is built by using shared vector space of embeddings while training, and by using separate language layers for each language to accommodate for diversity in syntax of the languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Works", "sec_num": "2" }, { "text": "In this paper, we propose that joint training of IE system on different language datasets, using 'multi-lingual' word embeddings and language layers helps in better extraction of arguments. This is particularly true when the dataset is limited in size. To corroborate our claim, we device two different systems, i). monolingual baseline system, and ii). multi-lingual system. The 'monolingual baseline' system only takes input data (sentence wise) from one language and extracts the arguments. For word representation, it uses monolingual word embeddings. The 'multi-lingual' argument extraction system uses separate language layers and multi-lingual word embeddings for joint training on all the three languages.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Methodology", "sec_num": "3" }, { "text": "The monolingual word-embeddings that are used in our experiments are also known as fastText 1 . It was proposed by Bojanowski et al. (2017) , and is based on the skipgram model. However instead of using one-hot vector encoding for each word while training, a vector representation of a word that considers character n-grams occurring in the word is formed. To get this representation, the n-grams from all the words for 'n' greater than 2 and smaller than 7 are extracted. After this, a dictionary of all the extracted n-grams is created. A given word w, can now be denoted by \u0393 w \u2282 {1, ...., G} i.e the set of n-grams appearing in the word; where G is the size of the n-gram dictionary. With each n-gram in G, a vector representation z g is associated. A word represention is obtained by summing up all the n-grams, as described in Equation 1:", "cite_spans": [ { "start": 115, "end": 139, "text": "Bojanowski et al. (2017)", "ref_id": "BIBREF1" } ], "ref_spans": [], "eq_spans": [], "section": "Monolingual Word Embedding", "sec_num": "3.0.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "V w = \u2211 g\u2208Gw z g", "eq_num": "(1)" } ], "section": "Monolingual Word Embedding", "sec_num": "3.0.1" }, { "text": "The continuous skip-gram model used these word vectors V w , to obtain word-embedding representa-tions of words. The main advantage of this technique is that, even in the absence of some word in the training corpus, some representations of the word is still obtained as the n-gram representation of words is considered. This skip-gram model is trained using Wikipedia data dump of each language. The dimension of the word vector to is set to 300.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Monolingual Word Embedding", "sec_num": "3.0.1" }, { "text": "Multi-lingual embeddings are obtained by learning a mapping matrix W , between source embeddings X = {x 1 , x 2 , x 3 ..., x n } and target embeddings Y = {y 1 , y 2 , y 3 , ..., y n } without crosslingual supervision.Adversarial training was used in this method proposed by Conneau et al. (2017) . A discriminator is trained to discriminate between a randomly sampled element from W X = {W x 1 , ..., W x n } and Y . At the same time W is trained to prevent the discriminator from making correct prediction. Thus making it a two-player game, where the discriminator tries to maximize its capability of identifying the origins of an embedding, and W aims to prevent the discriminator from doing so by making W X and Y as indistinguishable as possible. The W matrix is trained with near orthogonality constraint, to ensure that while transforming the source vector to the target vector space, the angles and distances between words in the embeddings are not distorted during transformation. To achieve this near orthogonality constraint, weight updation for W is done using Equation 2.", "cite_spans": [ { "start": 275, "end": 296, "text": "Conneau et al. (2017)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Multi-lingual Word Embedding", "sec_num": "3.0.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "W \u2190 (1 + \u03b2)W \u2212 \u03b2(W W T )W", "eq_num": "(2)" } ], "section": "Multi-lingual Word Embedding", "sec_num": "3.0.2" }, { "text": "Here, \u03b2 was set to 0.01 for the transformation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Multi-lingual Word Embedding", "sec_num": "3.0.2" }, { "text": "For our experiments we trained mapping matrices W hindi and W bengali that map the Hindi and Bengali word embeddings to the vector space of English embeddings.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Multi-lingual Word Embedding", "sec_num": "3.0.2" }, { "text": "The 'monolingual baseline' model (c.f Figure 1 ) is based on Bi-Directional Long Short Term Memory (Bi-LSTM) (Hochreiter and Schmidhuber, 1997; Schuster and Paliwal, 1997) and Convolutional Neural Networks (CNN) (Kim, 2014) . The input to the model is a sentence, represented by a sequence of monolingual word embeddings. Since Bi-LSTM and CNN take sequences of equal lengths, the shorter sequences are padded by zero vectors. This sequence is passed through Bi-LSTM and CNN having filter size 2 and 3. The Bi-LSTM gives contextual representation of each word, while the CNN extracts the 'bi-gram' and 'tri-gram' features for the sequence. These features are concatenated and passed through a fully connected layer. This layer gives shared representation for the task of argument extraction. Since the arguments in the dataset are not mutually exclusive (E.g: Place or Participant argument can also be a part of Reason or After-effect argument), we have different layers to predict different arguments independently. We have 6 different fullyconnected layers in parallel, each of them specialized for detection of one of the 6 arguments. 'Softmax' is used after each of the final layers to classify the representation into I, O or B of an argument.", "cite_spans": [ { "start": 109, "end": 143, "text": "(Hochreiter and Schmidhuber, 1997;", "ref_id": "BIBREF8" }, { "start": 144, "end": 171, "text": "Schuster and Paliwal, 1997)", "ref_id": "BIBREF27" }, { "start": 212, "end": 223, "text": "(Kim, 2014)", "ref_id": "BIBREF13" } ], "ref_spans": [ { "start": 38, "end": 46, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Monolingual Baseline Model", "sec_num": "3.1" }, { "text": "For multi-lingual system, we build a model based on the baseline model, by adding separate language layers (L 1 , L 2 and L 3 ) for each language (c.f Figure 2) . A layer L i and its subsequent layers are only trained when input data is also of language L i . We represent the input sentence as a sequence of multi-lingual word embeddings, and padding with zero vectors is used to make the sequence equal in length. Similar to the 'monolingual baseline' model, Bi-LSTM, CNN and a fully connected layer is used. This fully connected layer ", "cite_spans": [], "ref_spans": [ { "start": 151, "end": 160, "text": "Figure 2)", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Multi-lingual Model", "sec_num": "3.2" }, { "text": "In this section, we describe the dataset used and the experiments conducted.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Dataset and Experiments", "sec_num": "4" }, { "text": "To create the dataset, we crawled news articles in disaster domain from popular news websites in Hindi. These news articles were annotated by three annotators, with good language abilities and having satisfactory knowledge in the relevant area. The guidelines for annotation used were similar to the guidelines given by TAC KBP 2017 Event Sequence Annotation Guidelines 2 . We recorded that the annotators had Kappa agreement score of 0.85 Time 3,953 11,042 822 Place 12,410 10,576 3,018 Reason 1,573 1,744 544 Casualties 12,171 15,870 4,823 Participant 2,264 4,311 639 After-effects 13,355 9,731 274 ", "cite_spans": [], "ref_spans": [ { "start": 440, "end": 608, "text": "Time 3,953 11,042 822 Place 12,410 10,576 3,018 Reason 1,573 1,744 544 Casualties 12,171 15,870 4,823 Participant 2,264 4,311 639 After-effects 13,355", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Dataset", "sec_num": "4.1" }, { "text": "We conduct two separate experiments to show that dataset from different languages (L 1 and L 2 ) can be leveraged to improve the performance of argument extraction system of a different language (L 3 ). First we conduct experiment to obtain baseline results on 'mono-lingual' setup. Next, we perform experiment using the combined dataset of all the three languages using 'multi-lingual' argument extraction model.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Experiments", "sec_num": "4.2" }, { "text": "This experiment is conducted separately on each dataset using the 'monolingual baseline model' (c.f. Figure 1 ) and monolingual fastText embeddings. The results of this experiment is used as a baseline, against which the results of the other experiment is compared. The following set-up is used for the experiment: i). learning rate: 1 \u00d7 10 \u22122 , ii). batch size: 32, iii). optimizer: Adam (Kingma and Ba, 2014), iv). loss function: Binary cross-entropy. The best model based on validation-set accuracy was saved after 100 epochs.", "cite_spans": [], "ref_spans": [ { "start": 101, "end": 109, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Monolingual Experiment", "sec_num": "4.2.1" }, { "text": "This experiment is conducted on the combined dataset of three languages, using the 'multi-lingual model' (c.f Figure 2 ). Multi-lingual word embeddings (described in Section 3.2) were used for word representation in all the three languages, in this experiment. The same experimental set-up used for the 'monolingual baseline' experiment, is also used for this experiment. The training of multilingual system was done batch wise, i.e. each language branch was trained for one batch alternatively. The number of steps per epochs was decided by the number of batches needed to complete one epoch of the largest training set, among the different language datasets.", "cite_spans": [], "ref_spans": [ { "start": 110, "end": 118, "text": "Figure 2", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Multi-lingual Experiment", "sec_num": "4.2.2" }, { "text": "In this section, we discuss the results obtained for the two experiments described in Section 4.2. We also provide analysis of the results. F1-Score is used as an evaluation metric, and all the results reported are 5-Fold cross-validated. The results for both, 'monolingual' and 'cross-lingual' experiments are reported in Table 2 . From the results, it can be observed that F1-score for Hindi and English datasets improve for most arguments (5 out of 6 arguments), while the results for Bengali dataset improves for three out of the six arguments.", "cite_spans": [], "ref_spans": [ { "start": 323, "end": 330, "text": "Table 2", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "We also test the statistical significance of each increment in F1-Score for argument extraction. The 'p-values' obtained after 't-test' are shown in Table 3 . It can be seen that most improvements in F1-score are statistically significant.", "cite_spans": [], "ref_spans": [ { "start": 149, "end": 156, "text": "Table 3", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "It is observed that multi-word Time arguments are better captured by 'multi-lingual' model than by the 'monolingual baseline' model. An example of this can be seen in the following sentence:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Hindi Text: \u090f\u0938\u090f\u0938\u092a\u0940 \u0938\u0902 \u0924\u094b\u0937 \u0915\u0941 \u092e\u093e\u0930 \u0938 \u0939 \u0928\u0947 \u092c\u0924\u093e\u092f\u093e \u0915 \u0930 \u0935\u0935\u093e\u0930 \u0930\u093e\u0924 \u0915\u094b \u091c\u0932\u093e\u0932\u092a\u0941 \u0930 \u092a\u0930 \u0924\u0948 \u0928\u093e\u0924 \u092a\u0941 \u0932\u0938\u0915 \u092e \u092f \u0928\u0947 \u092c\u093e\u0907\u0915 \u092a\u0930 \u0938\u0935\u093e\u0930 \u0926\u094b \u092f\u0941 \u0935\u0915 \u0915\u094b \u0930\u094b\u0915\u0928\u0947 \u0915 \u0915\u094b \u0936\u0936 \u0915", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Transliteration: esesapee santosh kumaar sinh ne bataaya ki ravivaar raat ko jalaalapur par tainaat pulisakarmiyon ne baik par savaar do yuvakon ko rokane kee koshish kee", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Translation: SSP Santosh Kumar Singh said that on Sunday night, policemen stationed at Jalalpur tried to stop two youths riding on bikes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "In the aforementioned sentence the actual phrase denoting time is '\u0930 \u0935\u0935\u093e\u0930 \u0930\u093e\u0924' (Sunday night). However the 'monolingual' model only detects '\u0930 \u0935\u0935\u093e\u0930' (Sunday) as the Time argument. However, after multi-lingual training the entire time phrase is correctly detected. This is because the lack of training data for multi-word time arguments in Hindi, is supplemented by training data from Bengali and English.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "Another interesting observation is that, for Casualty argument of English dataset, the 'monolingual' system often confuses people as casualties, even when they are not. An example of such observation is as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Actual: Over 200000 people in 36 villages located 6 miles (10 km) from the volcano were advised to evacuate immediately.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Monolingual Prediction: Over 200000 people in 36 villages located 6 miles (10 km) from the volcano were advised to evacuate immediately.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Multi-lingual Prediction: Over 200000 people in 36 villages located 6 miles (10 km) from the volcano were advised to evacuate immediately.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "In the above example the phrase '200000 people' does not denote casualty, however the 'monolingual' model confuses it as casualty. This is due to the lack of training data in English to learn the difference between some count of people and actual casualty. However, after 'multi-lingual' train-ing the model is able to make this distinction correctly.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "The F1-score for Place arguments for all the datasets, is better for the 'monolingual baseline' model. This is because Place argument is present in good numbers for all the datasets, therefore there are enough instances for proper training of deep learning model, even in monolingual setting. Using 'multi-lingual model' for such cases is of little help. Furthermore, the syntactic difference between languages confuses the system, thus degrading the performance of the 'multi-lingual' system. A good example of this phenomenon is show below:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Actual: Three youths lost their lives when the car they were travelling in collided with a truck near Gaddoli village of Naraingarh in Ambala.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Monolingual Prediction: Three youths lost their lives when the car they were travelling in collided with a truck near Gaddoli village of Naraingarh in Ambala.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Multi-lingual Prediction: Three youths lost their lives when the car they were travelling in collided with a truck near Gaddoli village of Naraingarh in Ambala.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "It can be observed that the 'monolingual baseline' model predicts the entire phrase describing the Place argument correctly. However the prediction by 'multi-lingual model' misses the preposition 'in', which is present between 'Naraingarh' and 'Ambala'. The same sentence can be written in Bengali as follows:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "\u2022 Bengali Transliteration: Amb\u0101l\u0101ra n\u0101r\u0101y\u0227naga\u1e5b\u0113ra g\u0101ddali gr\u0101m\u0113ra k\u0101ch\u0113 \u0113ka\u1e6di \u1e6dr\u0101k\u0113ra s\u0101th\u0113 \u1e6dr\u0113n\u0113ra mukh\u014dmukhi sa\u1e45ghar\u1e63\u0113 tinajana yubaka pr\u0101\u1e47a h\u0101r\u0101y\u0227.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "The phrase 'in Ambala' is represented by a single word 'Amb\u0101l\u0101ra', in Bengali. This difference in syntax between languages, makes the 'multilingual' system miss the word 'in' thus degrading the performance of the system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "The best improvement in F1-score is observed for the arguments Reason and After-effects for the English language. This is because these two arguments have least support in the dataset, and thus multi-lingual training helps by mitigating the scarcity in training examples. The same phenomenon can also be observed for Reason argument which has a low support in Hindi dataset. Thus through our analysis we can conclude that, 'multi-lingual' training can help in improving the performance of the system for low support classes. However, it can also cause confusion and deteriorate the performance for high support classes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5" }, { "text": "In this paper we create a dataset for argument extraction for disaster domain, for three languages Hindi, Bengali and English. We then build a deep learning model for extraction of these argument in each language separately. Since the data is limited in size, we build another model that leverages data from all the languages. To make use of different language datasets, we first bring the word embeddings of all the three languages to the same vector space. We also use separate language layers to accommodate divergence in syntax of the languages. Through our experiments we show that training in shared vector space by using 'multi-lingual' system helps in improving the performance of low support arguments. We also show that the for high support arguments, the syntactic difference in language can sometimes overcome the benefit of 'multi-lingual' training and cost in performance of our proposed 'multi-lingual' system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6" }, { "text": "In future we would like to explore how to handle these syntactic differences so that the performance can be further improved. It would also be interesting to explore the range of languages that can be trained successfully in a multi-lingual setting.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "6" }, { "text": "https://github.com/facebookresearch/ fastText", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "https://cairo.lti.cs.cmu.edu/kbp/2017/ event/TAC_KBP_2017_Event_Coreference_and_ Sequence_Annotation_Guidelines_v1.1.pdf", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "The research reported in this paper is an outcome of the project titled \"A Platform for Cross-lingual and Multi-lingual Event Monitoring in Indian Languages\", supported by IMPRINT-1, MHRD, Govt. of India, and MeiTY, Govt. of India.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgement", "sec_num": "7" } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "The stages of event extraction", "authors": [ { "first": "David", "middle": [], "last": "Ahn", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the Workshop on Annotating and Reasoning about Time and Events", "volume": "", "issue": "", "pages": "1--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "David Ahn. 2006. The stages of event extraction. In Proceedings of the Workshop on Annotating and Reasoning about Time and Events, pages 1-8.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Enriching word vectors with subword information", "authors": [ { "first": "Piotr", "middle": [], "last": "Bojanowski", "suffix": "" }, { "first": "Edouard", "middle": [], "last": "Grave", "suffix": "" }, { "first": "Armand", "middle": [], "last": "Joulin", "suffix": "" }, { "first": "Tomas", "middle": [], "last": "Mikolov", "suffix": "" } ], "year": 2017, "venue": "Transactions of the Association for Computational Linguistics", "volume": "5", "issue": "", "pages": "135--146", "other_ids": {}, "num": null, "urls": [], "raw_text": "Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Associa- tion for Computational Linguistics, 5:135-146.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Template-based information extraction without the templates", "authors": [ { "first": "Nathanael", "middle": [], "last": "Chambers", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Jurafsky", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "976--986", "other_ids": {}, "num": null, "urls": [], "raw_text": "Nathanael Chambers and Dan Jurafsky. 2011. Template-based information extraction without the templates. In Proceedings of the 49th Annual Meeting of the Association for Computational Lin- guistics: Human Language Technologies-Volume 1, pages 976-986. Association for Computational Linguistics.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Event extraction via dynamic multipooling convolutional neural networks", "authors": [ { "first": "Yubo", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Liheng", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Kang", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Daojian", "middle": [], "last": "Zeng", "suffix": "" }, { "first": "Jun", "middle": [], "last": "Zhao", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing", "volume": "1", "issue": "", "pages": "167--176", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multi- pooling convolutional neural networks. In Proceed- ings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th Interna- tional Joint Conference on Natural Language Pro- cessing (Volume 1: Long Papers), pages 167-176.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Language specific issue and feature exploration in chinese event extraction", "authors": [ { "first": "Zheng", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Heng", "middle": [], "last": "Ji", "suffix": "" } ], "year": 2009, "venue": "The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers", "volume": "", "issue": "", "pages": "209--212", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zheng Chen and Heng Ji. 2009. Language specific issue and feature exploration in chinese event ex- traction. In Proceedings of Human Language Tech- nologies: The 2009 Annual Conference of the North American Chapter of the Association for Computa- tional Linguistics, Companion Volume: Short Pa- pers, pages 209-212.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Word translation without parallel data", "authors": [ { "first": "Alexis", "middle": [], "last": "Conneau", "suffix": "" }, { "first": "Guillaume", "middle": [], "last": "Lample", "suffix": "" }, { "first": "Marc'aurelio", "middle": [], "last": "Ranzato", "suffix": "" }, { "first": "Ludovic", "middle": [], "last": "Denoyer", "suffix": "" }, { "first": "Herv\u00e9", "middle": [], "last": "J\u00e9gou", "suffix": "" } ], "year": 2017, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1710.04087" ] }, "num": null, "urls": [], "raw_text": "Alexis Conneau, Guillaume Lample, Marc'Aurelio Ranzato, Ludovic Denoyer, and Herv\u00e9 J\u00e9gou. 2017. Word translation without parallel data. arXiv preprint arXiv:1710.04087.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "The automatic content extraction (ace) program-tasks, data, and evaluation", "authors": [ { "first": "Alexis", "middle": [], "last": "George R Doddington", "suffix": "" }, { "first": "", "middle": [], "last": "Mitchell", "suffix": "" }, { "first": "A", "middle": [], "last": "Mark", "suffix": "" }, { "first": "", "middle": [], "last": "Przybocki", "suffix": "" }, { "first": "A", "middle": [], "last": "Lance", "suffix": "" }, { "first": "Stephanie", "middle": [ "M" ], "last": "Ramshaw", "suffix": "" }, { "first": "Ralph", "middle": [ "M" ], "last": "Strassel", "suffix": "" }, { "first": "", "middle": [], "last": "Weischedel", "suffix": "" } ], "year": 2004, "venue": "LREC", "volume": "2", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "George R Doddington, Alexis Mitchell, Mark A Przy- bocki, Lance A Ramshaw, Stephanie M Strassel, and Ralph M Weischedel. 2004. The automatic content extraction (ace) program-tasks, data, and evaluation. In LREC, volume 2, page 1.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "A language-independent neural network for event detection", "authors": [ { "first": "Xiaocheng", "middle": [], "last": "Feng", "suffix": "" }, { "first": "Bing", "middle": [], "last": "Qin", "suffix": "" }, { "first": "Ting", "middle": [], "last": "Liu", "suffix": "" } ], "year": 2018, "venue": "Science China Information Sciences", "volume": "61", "issue": "9", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Xiaocheng Feng, Bing Qin, and Ting Liu. 2018. A language-independent neural network for event detection. Science China Information Sciences, 61(9):092106.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Long short-term memory", "authors": [ { "first": "Sepp", "middle": [], "last": "Hochreiter", "suffix": "" }, { "first": "J\u00fcrgen", "middle": [], "last": "Schmidhuber", "suffix": "" } ], "year": 1997, "venue": "Neural computation", "volume": "9", "issue": "8", "pages": "1735--1780", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735-1780.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Using cross-entity inference to improve event extraction", "authors": [ { "first": "Yu", "middle": [], "last": "Hong", "suffix": "" }, { "first": "Jianfeng", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Bin", "middle": [], "last": "Ma", "suffix": "" }, { "first": "Jianmin", "middle": [], "last": "Yao", "suffix": "" }, { "first": "Guodong", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "Qiaoming", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "1127--1136", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yu Hong, Jianfeng Zhang, Bin Ma, Jianmin Yao, Guodong Zhou, and Qiaoming Zhu. 2011. Us- ing cross-entity inference to improve event extrac- tion. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Hu- man Language Technologies-Volume 1, pages 1127- 1136. Association for Computational Linguistics.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Event argument extraction based on crf", "authors": [ { "first": "Libin", "middle": [], "last": "Hou", "suffix": "" }, { "first": "Peifeng", "middle": [], "last": "Li", "suffix": "" }, { "first": "Qiaoming", "middle": [], "last": "Zhu", "suffix": "" }, { "first": "Yuan", "middle": [], "last": "Cao", "suffix": "" } ], "year": 2012, "venue": "Workshop on Chinese Lexical Semantics", "volume": "", "issue": "", "pages": "32--39", "other_ids": {}, "num": null, "urls": [], "raw_text": "Libin Hou, Peifeng Li, Qiaoming Zhu, and Yuan Cao. 2012. Event argument extraction based on crf. In Workshop on Chinese Lexical Semantics, pages 32- 39. Springer.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Peeling back the layers: detecting event role fillers in secondary contexts", "authors": [ { "first": "Ruihong", "middle": [], "last": "Huang", "suffix": "" }, { "first": "Ellen", "middle": [], "last": "Riloff", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", "volume": "1", "issue": "", "pages": "1137--1147", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ruihong Huang and Ellen Riloff. 2011. Peeling back the layers: detecting event role fillers in secondary contexts. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies-Volume 1, pages 1137-1147. Association for Computational Linguis- tics.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Refining event extraction through cross-document inference", "authors": [ { "first": "Heng", "middle": [], "last": "Ji", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2008, "venue": "Proceedings of ACL-08: HLT", "volume": "", "issue": "", "pages": "254--262", "other_ids": {}, "num": null, "urls": [], "raw_text": "Heng Ji and Ralph Grishman. 2008. Refining event extraction through cross-document inference. Pro- ceedings of ACL-08: HLT, pages 254-262.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Convolutional neural networks for sentence classification", "authors": [ { "first": "Yoon", "middle": [], "last": "Kim", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1408.5882" ] }, "num": null, "urls": [], "raw_text": "Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Adam: A method for stochastic optimization", "authors": [ { "first": "P", "middle": [], "last": "Diederik", "suffix": "" }, { "first": "Jimmy", "middle": [], "last": "Kingma", "suffix": "" }, { "first": "", "middle": [], "last": "Ba", "suffix": "" } ], "year": 2014, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1412.6980" ] }, "num": null, "urls": [], "raw_text": "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Joint modeling of argument identification and role determination in chinese event extraction with discourse-level information", "authors": [ { "first": "Peifeng", "middle": [], "last": "Li", "suffix": "" }, { "first": "Qiaoming", "middle": [], "last": "Zhu", "suffix": "" }, { "first": "Guodong", "middle": [], "last": "Zhou", "suffix": "" } ], "year": 2013, "venue": "Twenty-Third International Joint Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peifeng Li, Qiaoming Zhu, and Guodong Zhou. 2013. Joint modeling of argument identification and role determination in chinese event extraction with discourse-level information. In Twenty-Third Inter- national Joint Conference on Artificial Intelligence.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "A multi-lingual multi-task architecture for low-resource sequence labeling", "authors": [ { "first": "Ying", "middle": [], "last": "Lin", "suffix": "" }, { "first": "Shengqi", "middle": [], "last": "Yang", "suffix": "" }, { "first": "Veselin", "middle": [], "last": "Stoyanov", "suffix": "" }, { "first": "Heng", "middle": [], "last": "Ji", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "799--809", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ying Lin, Shengqi Yang, Veselin Stoyanov, and Heng Ji. 2018. A multi-lingual multi-task architecture for low-resource sequence labeling. In Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 799-809.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Event detection via gated multilingual attention mechanism", "authors": [ { "first": "Jian", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Yubo", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Kang", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Jun", "middle": [], "last": "Zhao", "suffix": "" } ], "year": 2018, "venue": "Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jian Liu, Yubo Chen, Kang Liu, and Jun Zhao. 2018. Event detection via gated multilingual attention mechanism. In Thirty-Second AAAI Conference on Artificial Intelligence.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Automatic event extraction with structured preference modeling", "authors": [ { "first": "Wei", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Roth", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers", "volume": "1", "issue": "", "pages": "835--844", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wei Lu and Dan Roth. 2012. Automatic event extrac- tion with structured preference modeling. In Pro- ceedings of the 50th Annual Meeting of the Associ- ation for Computational Linguistics: Long Papers- Volume 1, pages 835-844. Association for Compu- tational Linguistics.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Overview of tac kbp 2015 event nugget track", "authors": [ { "first": "Teruko", "middle": [], "last": "Mitamura", "suffix": "" }, { "first": "Zhengzhong", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Eduard", "middle": [ "H" ], "last": "Hovy", "suffix": "" } ], "year": 2015, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Teruko Mitamura, Zhengzhong Liu, and Eduard H Hovy. 2015. Overview of tac kbp 2015 event nugget track. In TAC.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Joint event extraction via recurrent neural networks", "authors": [ { "first": "Kyunghyun", "middle": [], "last": "Thien Huu Nguyen", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Cho", "suffix": "" }, { "first": "", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "300--309", "other_ids": {}, "num": null, "urls": [], "raw_text": "Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grish- man. 2016. Joint event extraction via recurrent neu- ral networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, pages 300-309.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Event detection and domain adaptation with convolutional neural networks", "authors": [ { "first": "Huu", "middle": [], "last": "Thien", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Nguyen", "suffix": "" }, { "first": "", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing", "volume": "2", "issue": "", "pages": "365--371", "other_ids": {}, "num": null, "urls": [], "raw_text": "Thien Huu Nguyen and Ralph Grishman. 2015. Event detection and domain adaptation with convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Lin- guistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 365-371.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Modeling skip-grams for event detection with convolutional neural networks", "authors": [ { "first": "Huu", "middle": [], "last": "Thien", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Nguyen", "suffix": "" }, { "first": "", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "886--891", "other_ids": {}, "num": null, "urls": [], "raw_text": "Thien Huu Nguyen and Ralph Grishman. 2016. Mod- eling skip-grams for event detection with convolu- tional neural networks. In Proceedings of the 2016 Conference on Empirical Methods in Natural Lan- guage Processing, pages 886-891.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Graph convolutional networks with argument-aware pooling for event detection", "authors": [ { "first": "Huu", "middle": [], "last": "Thien", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Nguyen", "suffix": "" }, { "first": "", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2018, "venue": "Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Thien Huu Nguyen and Ralph Grishman. 2018. Graph convolutional networks with argument-aware pool- ing for event detection. In Thirty-Second AAAI Con- ference on Artificial Intelligence.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Effective information extraction with semantic affinity patterns and relevant regions", "authors": [ { "first": "Siddharth", "middle": [], "last": "Patwardhan", "suffix": "" }, { "first": "Ellen", "middle": [], "last": "Riloff", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)", "volume": "", "issue": "", "pages": "717--727", "other_ids": {}, "num": null, "urls": [], "raw_text": "Siddharth Patwardhan and Ellen Riloff. 2007. Effective information extraction with semantic affinity pat- terns and relevant regions. In Proceedings of the 2007 Joint Conference on Empirical Methods in Nat- ural Language Processing and Computational Nat- ural Language Learning (EMNLP-CoNLL), pages 717-727.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "A unified model of phrasal and sentential evidence for information extraction", "authors": [ { "first": "Siddharth", "middle": [], "last": "Patwardhan", "suffix": "" }, { "first": "Ellen", "middle": [], "last": "Riloff", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing", "volume": "1", "issue": "", "pages": "151--160", "other_ids": {}, "num": null, "urls": [], "raw_text": "Siddharth Patwardhan and Ellen Riloff. 2009. A uni- fied model of phrasal and sentential evidence for in- formation extraction. In Proceedings of the 2009 Conference on Empirical Methods in Natural Lan- guage Processing: Volume 1-Volume 1, pages 151- 160. Association for Computational Linguistics.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "An extensible event extraction system with cross-media event resolution", "authors": [ { "first": "Fabio", "middle": [], "last": "Petroni", "suffix": "" }, { "first": "Natraj", "middle": [], "last": "Raman", "suffix": "" }, { "first": "Tim", "middle": [], "last": "Nugent", "suffix": "" }, { "first": "Armineh", "middle": [], "last": "Nourbakhsh", "suffix": "" }, { "first": "\u017darko", "middle": [], "last": "Pani\u0107", "suffix": "" }, { "first": "Sameena", "middle": [], "last": "Shah", "suffix": "" }, { "first": "Jochen L", "middle": [], "last": "Leidner", "suffix": "" } ], "year": 2018, "venue": "Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining", "volume": "", "issue": "", "pages": "626--635", "other_ids": {}, "num": null, "urls": [], "raw_text": "Fabio Petroni, Natraj Raman, Tim Nugent, Armineh Nourbakhsh, \u017darko Pani\u0107, Sameena Shah, and Jochen L Leidner. 2018. An extensible event ex- traction system with cross-media event resolution. In Proceedings of the 24th ACM SIGKDD Interna- tional Conference on Knowledge Discovery & Data Mining, pages 626-635. ACM.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Bidirectional recurrent neural networks", "authors": [ { "first": "Mike", "middle": [], "last": "Schuster", "suffix": "" }, { "first": "K", "middle": [], "last": "Kuldip", "suffix": "" }, { "first": "", "middle": [], "last": "Paliwal", "suffix": "" } ], "year": 1997, "venue": "IEEE Transactions on Signal Processing", "volume": "45", "issue": "11", "pages": "2673--2681", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mike Schuster and Kuldip K Paliwal. 1997. Bidirec- tional recurrent neural networks. IEEE Transactions on Signal Processing, 45(11):2673-2681.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Jointly extracting event triggers and arguments by dependency-bridge rnn and tensor-based argument interaction", "authors": [ { "first": "Lei", "middle": [], "last": "Sha", "suffix": "" }, { "first": "Feng", "middle": [], "last": "Qian", "suffix": "" }, { "first": "Baobao", "middle": [], "last": "Chang", "suffix": "" }, { "first": "Zhifang", "middle": [], "last": "Sui", "suffix": "" } ], "year": 2018, "venue": "Thirty-Second AAAI Conference on Artificial Intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lei Sha, Feng Qian, Baobao Chang, and Zhifang Sui. 2018. Jointly extracting event triggers and argu- ments by dependency-bridge rnn and tensor-based argument interaction. In Thirty-Second AAAI Con- ference on Artificial Intelligence.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Research on chinese event extraction", "authors": [ { "first": "Yan-Yan", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "Bing", "middle": [], "last": "Qin", "suffix": "" }, { "first": "Che", "middle": [], "last": "Wan-Xiang", "suffix": "" }, { "first": "Ting", "middle": [], "last": "Liu", "suffix": "" } ], "year": 2008, "venue": "Journal of Chinese Information Processing", "volume": "22", "issue": "1", "pages": "3--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yan-yan Zhao, Bing Qin, Wan-xiang Che, and Ting Liu. 2008. Research on chinese event extraction. Journal of Chinese Information Processing, 22(1):3-8.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "text": "monolingual baseline model for argument extraction", "uris": null, "num": null }, "FIGREF1": { "type_str": "figure", "text": "Multi-lingual baseline model for argument extraction produces shared language and task representation as output. Three separate language layers for the languages Hindi, Bengali and English are used in parallel. These language layers decode the language specific representation from shared representation. After each language layer we have 6 fully connected layers for each of the 6 arguments. 'Softmax' classifier is used to classify the representation into I, O or B of an argument.", "uris": null, "num": null }, "TABREF1": { "type_str": "table", "html": null, "text": "", "content": "
: Distribution of number of arguments in Hindi,
Bengali and English datasets
on average. We also obtained equivalent dataset
in Bengali and English language from a collabora-
tion. The total dataset is comprised of 2,191 doc-
uments (Hindi: 922, Bengali: 999 and English:
270). It contains 44,615 sentences (Hindi: 17,116,
Bengali: 25,717 and English: 1,782). The six ar-
guments in the dataset and their distribution in the
three languages are detailed in the Table 1.
", "num": null }, "TABREF3": { "type_str": "table", "html": null, "text": "", "content": "
: Results (F1-Scores) for 'mono-lingual' and 'multi-lingual' experiments on Hindi, Bengali and English
datasets: 5-Fold cross-validated
ArgumentHindi Bengali English
Time0.46n/a0.03
Placen/an/an/a
Reason0.030.180.04
Casualties0.39n/a0.10
Participant0.010.110.54
After-effects 0.040.090.01
", "num": null }, "TABREF4": { "type_str": "table", "html": null, "text": "", "content": "
: The 'p-values' obtained for each improvement
in results from the baseline 'mono-lingual' to 'multi-
lingual' experiment (n/a is used for instances where no
improvement was observed)
", "num": null } } } }