{ "paper_id": "P13-1008", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T09:33:58.677669Z" }, "title": "Joint Event Extraction via Structured Prediction with Global Features", "authors": [ { "first": "Qi", "middle": [], "last": "Li", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of New York New York", "location": { "postCode": "10016", "region": "NY", "country": "USA" } }, "email": "" }, { "first": "Heng", "middle": [], "last": "Ji", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of New York New York", "location": { "postCode": "10016", "region": "NY", "country": "USA" } }, "email": "" }, { "first": "Liang", "middle": [], "last": "Huang", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of New York New York", "location": { "postCode": "10016", "region": "NY", "country": "USA" } }, "email": "liang.huang.sh@gmail.com" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Traditional approaches to the task of ACE event extraction usually rely on sequential pipelines with multiple stages, which suffer from error propagation since event triggers and arguments are predicted in isolation by independent local classifiers. By contrast, we propose a joint framework based on structured prediction which extracts triggers and arguments together so that the local predictions can be mutually improved. In addition, we propose to incorporate global features which explicitly capture the dependencies of multiple triggers and arguments. Experimental results show that our joint approach with local features outperforms the pipelined baseline, and adding global features further improves the performance significantly. Our approach advances state-ofthe-art sentence-level event extraction, and even outperforms previous argument labeling methods which use external knowledge from other sentences and documents.", "pdf_parse": { "paper_id": "P13-1008", "_pdf_hash": "", "abstract": [ { "text": "Traditional approaches to the task of ACE event extraction usually rely on sequential pipelines with multiple stages, which suffer from error propagation since event triggers and arguments are predicted in isolation by independent local classifiers. By contrast, we propose a joint framework based on structured prediction which extracts triggers and arguments together so that the local predictions can be mutually improved. In addition, we propose to incorporate global features which explicitly capture the dependencies of multiple triggers and arguments. Experimental results show that our joint approach with local features outperforms the pipelined baseline, and adding global features further improves the performance significantly. Our approach advances state-ofthe-art sentence-level event extraction, and even outperforms previous argument labeling methods which use external knowledge from other sentences and documents.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Event extraction is an important and challenging task in Information Extraction (IE), which aims to discover event triggers with specific types and their arguments. Most state-of-the-art approaches (Ji and Grishman, 2008; Liao and Grishman, 2010; Hong et al., 2011) use sequential pipelines as building blocks, which break down the whole task into separate subtasks, such as trigger identification/classification and argument identification/classification. As a common drawback of the staged architecture, errors in upstream component are often compounded and propagated to the downstream classifiers. The downstream components, however, cannot impact earlier deci-sions. For example, consider the following sentences with an ambiguous word \"fired\":", "cite_spans": [ { "start": 198, "end": 221, "text": "(Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 222, "end": 246, "text": "Liao and Grishman, 2010;", "ref_id": "BIBREF13" }, { "start": 247, "end": 265, "text": "Hong et al., 2011)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "(1) In Baghdad, a cameraman died when an American tank fired on the Palestine Hotel.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "(2) He has fired his air defense chief .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In sentence (1), \"fired\" is a trigger of type Attack.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Because of the ambiguity, a local classifier may miss it or mislabel it as a trigger of End-Position. However, knowing that \"tank\" is very likely to be an Instrument argument of Attack events, the correct event subtype assignment of \"fired\" is obviously Attack. Likewise, in sentence (2), \"air defense chief\" is a job title, hence the argument classifier is likely to label it as an Entity argument for End-Position trigger. In addition, the local classifiers are incapable of capturing inter-dependencies among multiple event triggers and arguments. Consider sentence (1) again. Figure 1 depicts the corresponding event triggers and arguments. The dependency between \"fired\" and \"died\" cannot be captured by the local classifiers, which may fail to attach \"cameraman\" to \"fired\" as a Target argument. By using global features, we can propagate the Victim argument of the Die event to the Target argument of the Attack event. As another example, knowing that an Attack event usually only has one Attacker argument, we could penalize assignments in which one trigger has more than one Attacker. Such global features cannot be easily exploited by a local classifier.", "cite_spans": [], "ref_spans": [ { "start": 580, "end": 588, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Therefore, we take a fresh look at this problem and formulate it, for the first time, as a structured learning problem. We propose a novel joint event extraction algorithm to predict the triggers and arguments simultaneously, and use the structured perceptron (Collins, 2002) to train the joint model. This way we can capture the dependencies between triggers and argument as well as explore", "cite_spans": [ { "start": 260, "end": 275, "text": "(Collins, 2002)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In Baghdad, a cameraman died when an American tank fired on the Palestine Hotel. Figure 1 : Event mentions of example (1). There are two event mentions that share three arguments, namely the Die event mention triggered by \"died\", and the Attack event mention triggered by \"fired\". arbitrary global features over multiple local predictions. However, different from easier tasks such as part-of-speech tagging or noun phrase chunking where efficient dynamic programming decoding is feasible, here exact joint inference is intractable. Therefore we employ beam search in decoding, and train the model using the early-update perceptron variant tailored for beam search (Collins and Roark, 2004; Huang et al., 2012) .", "cite_spans": [ { "start": 665, "end": 690, "text": "(Collins and Roark, 2004;", "ref_id": "BIBREF4" }, { "start": 691, "end": 710, "text": "Huang et al., 2012)", "ref_id": "BIBREF9" } ], "ref_spans": [ { "start": 81, "end": 89, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We make the following contributions:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "1. Different from traditional pipeline approach, we present a novel framework for sentencelevel event extraction, which predicts triggers and their arguments jointly (Section 3).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "2. We develop a rich set of features for event extraction which yield promising performance even with the traditional pipeline (Section 3.4.1). In this paper we refer to them as local features.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "3. We introduce various global features to exploit dependencies among multiple triggers and arguments (Section 3.4.2). Experiments show that our approach outperforms the pipelined approach with the same set of local features, and significantly advances the state-of-the-art with the addition of global features which brings a notable further improvement (Section 4).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper we focus on the event extraction task defined in Automatic Content Extraction (ACE) evaluation. 1 The task defines 8 event types and 33 subtypes such as Attack, End-Position etc. We introduce the terminology of the ACE event extraction that we used in this paper:", "cite_spans": [ { "start": 110, "end": 111, "text": "1", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Event Extraction Task", "sec_num": "2" }, { "text": "1 http://projects.ldc.upenn.edu/ace/ \u2022 Event mention: an occurrence of an event with a particular type and subtype. \u2022 Event trigger: the word most clearly expresses the event mention. \u2022 Event argument: an entity mention, temporal expression or value (e.g. Job-Title) that serves as a participant or attribute with a specific role in an event mention. \u2022 Event mention: an instance that includes one event trigger and some arguments that appear within the same sentence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Event Extraction Task", "sec_num": "2" }, { "text": "Given an English text document, an event extraction system should predict event triggers with specific subtypes and their arguments from each sentence. Figure 1 depicts the event triggers and their arguments of sentence (1) in Section 1. The outcome of the entire sentence can be considered a graph in which each argument role is represented as a typed edge from a trigger to its argument.", "cite_spans": [], "ref_spans": [ { "start": 152, "end": 160, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Event Extraction Task", "sec_num": "2" }, { "text": "In this work, we assume that argument candidates such as entities are part of the input to the event extraction, and can be from either gold standard or IE system output.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Event Extraction Task", "sec_num": "2" }, { "text": "Based on the hypothesis that facts are interdependent, we propose to use structured perceptron with inexact search to jointly extract triggers and arguments that co-occur in the same sentence. In this section, we will describe the training and decoding algorithms for this model.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Joint Framework for Event Extraction", "sec_num": "3" }, { "text": "Structured perceptron is an extension to the standard linear perceptron for structured prediction, which was proposed in (Collins, 2002) . Given a sentence instance x \u2208 X , which in our case is a sentence with argument candidates, the structured perceptron involves the following decoding prob-lem which finds the best configuration z \u2208 Y according to the current model w:", "cite_spans": [ { "start": 121, "end": 136, "text": "(Collins, 2002)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "z = argmax y \u2208Y(x) w \u2022 f (x, y )", "eq_num": "(1)" } ], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "where f (x, y ) represents the feature vector for instance x along with configuration y . The perceptron learns the model w in an online fashion. Let D = {(x (j) , y (j) )} n j=1 be the set of training instances (with j indexing the current training instance). In each iteration, the algorithm finds the best configuration z for x under the current model (Eq. 1). If z is incorrect, the weights are updated as follows:", "cite_spans": [ { "start": 158, "end": 161, "text": "(j)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "w = w + f (x, y) \u2212 f (x, z)", "eq_num": "(2)" } ], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "The key step of the training and test is the decoding procedure, which aims to search for the best configuration under the current parameters. In simpler tasks such as part-of-speech tagging and noun phrase chunking, efficient dynamic programming algorithms can be employed to perform exact inference. Unfortunately, it is intractable to perform the exact search in our framework because: (1) by jointly modeling the trigger labeling and argument labeling, the search space becomes much more complex. (2) we propose to make use of arbitrary global features, which makes it infeasible to perform exact inference efficiently.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "To address this problem, we apply beam-search along with early-update strategy to perform inexact decoding. Collins and Roark (2004) proposed the early-update idea, and Huang et al. (2012) later proved its convergence and formalized a general framework which includes it as a special case. Figure 2 describes the skeleton of perceptron training algorithm with beam search. In each step of the beam search, if the prefix of oracle assignment y falls out from the beam, then the top result in the beam is returned for early update. One could also use the standard-update for inference, however, with highly inexact search the standardupdate generally does not work very well because of \"invalid updates\", i.e., updates that do not fix a violation (Huang et al., 2012) . In Section 4.5 we will show that the standard perceptron introduces many invalid updates especially with smaller beam sizes, also observed by Huang et al. (2012) .", "cite_spans": [ { "start": 108, "end": 132, "text": "Collins and Roark (2004)", "ref_id": "BIBREF4" }, { "start": 169, "end": 188, "text": "Huang et al. (2012)", "ref_id": "BIBREF9" }, { "start": 745, "end": 765, "text": "(Huang et al., 2012)", "ref_id": "BIBREF9" }, { "start": 910, "end": 929, "text": "Huang et al. (2012)", "ref_id": "BIBREF9" } ], "ref_spans": [ { "start": 290, "end": 296, "text": "Figure", "ref_id": null } ], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "To reduce overfitting, we used averaged parameters after training to decode test instances in our experiments. The resulting model is called averaged perceptron (Collins, 2002) .", "cite_spans": [ { "start": 161, "end": 176, "text": "(Collins, 2002)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "Input: Training set D = {(x (j) , y (j) )} n i=1 , maximum iteration number T Output: Model parameters w 1 Initialization: Set w = 0; 2 for t \u2190 1...T do 3 foreach (x, y) \u2208 D do 4 z \u2190 beamSearch (x, y, w) 5 if z = y then 6 w \u2190 w + f (x, y [1:|z|] ) \u2212 f (x, z)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "Figure 2: Perceptron training with beamsearch (Huang et al., 2012) . Here y [1:i] denotes the prefix of y that has length i, e.g., y [1:3] = (y 1 , y 2 , y 3 ).", "cite_spans": [ { "start": 46, "end": 66, "text": "(Huang et al., 2012)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Structured perceptron with beam search", "sec_num": "3.1" }, { "text": "Here we introduce the label sets for trigger and argument in the model. We use L \u222a {\u22a5} to denote the trigger label alphabet, where L represents the 33 event subtypes, and \u22a5 indicates that the token is not a trigger. Similarly, R \u222a {\u22a5} denotes the argument label sets, where R is the set of possible argument roles, and \u22a5 means that the argument candidate is not an argument for the current trigger. It is worth to note that the set R of each particular event subtype is subject to the entity type constraints defined in the official ACE annotation guideline 2 . For example, the Attacker argument for an Attack event can only be one of PER, ORG and GPE (Geo-political Entity).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Label sets", "sec_num": "3.2" }, { "text": "Let x = (x 1 , x 2 , ..., x s )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": ", E denote the sentence instance, where x i represents the i-th token in the sentence and E = {e k } m k=1 is the set of argument candidates. We use", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "y = (t 1 , a 1,1 , . . . , a 1,m , . . . , t s , a s,1 , . . . , a s,m )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "to denote the corresponding gold standard structure, where t i represents the trigger assignment for the token x i , and a i,k represents the argument role label for the edge between x i and argument candidate e k . For simplicity, throughout this paper we use y g(i) and y h(i,k) to represent t i and a i,k , respectively. Figure 3 demonstrates the notation with s = 3 and m = 2. The variables for the toy sentence \"Jobs founded Apple\" are as follows:", "cite_spans": [], "ref_spans": [ { "start": 324, "end": 332, "text": "Figure 3", "ref_id": "FIGREF1" } ], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "y = (t 1 , a 1,1 , a 1,2 , t 2 , a 2,1 , a 2,2 , | {z } arguments for x 2 t 3 , a 3,1 , a 3,2 ) g(1) g(2) h(2, 1) h(3, 2)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "x = (Jobs, x 2 founded, Apple), E {Jobs PER , Apple ORG } y = (\u22a5, \u22a5, \u22a5, Start Org t 2 , Agent, Org args for founded , \u22a5, \u22a5, \u22a5)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "Figure 4 describes the beam-search procedure with early-update for event extraction. During each step with token i, there are two sub-steps:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "\u2022 Trigger labeling We enumerate all possible trigger labels for the current token. The linear model defined in Eq. 1is used to score each partial configuration. Then the K-best partial configurations are selected to the beam, assuming the beam size is K. \u2022 Argument labeling After the trigger labeling step, we traverse all configurations in the beam. Once a trigger label for x i is found in the beam, the decoder searches through the argument candidates E to label the edges between each argument candidate and the trigger. After labeling each argument candidate, we again score each partial assignment and select the K-best results to the beam.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "After the second step, the rank of different trigger assignments can be changed because of the argument edges. Likewise, the decision on later argument candidates may be affected by earlier argument assignments. The overall time complexity for decoding is O(K \u2022 s \u2022 m).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Decoding", "sec_num": "3.3" }, { "text": "In this framework, we define two types of features, namely local features and global features. We first introduce the definition of local and global features in this paper, and then describe the implementation details later. Recall that in the linear model defined in Eq. (1), f (x, y) denotes the features extracted from the input instance x along Input: Instance x = (x 1 , x 2 , ..., x s ), E and the oracle output y if for training. K: Beam size. L \u222a {\u22a5}: trigger label alphabet. R \u222a {\u22a5}: argument label alphabet.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "Output: 1-best prediction z for x 1 Set beam B \u2190 [ ] /*empty configuration*/ 2 for i \u2190 1...s do 3 buf \u2190 {z \u2022 l | z \u2208 B, l \u2208 L \u222a {\u22a5}} B \u2190K-best(buf ) 4 if y [1:g(i)] \u2208 B then 5 return B[0] /*for early-update*/ 6", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "for e k \u2208 E do /*search for arguments*/ with configuration y. In general, each feature instance f in f is a function f : X \u00d7 Y \u2192 R, which maps x and y to a feature value. Local features are only related to predictions on individual trigger or argument. In the case of unigram tagging for trigger labeling, each local feature takes the form of f (x, i, y g(i) ), where i denotes the index of the current token, and y g(i) is its trigger label. In practice, it is convenient to define the local feature function as an indicator function, for example:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "7 buf \u2190 \u2205 8 for z \u2208 B do 9 buf \u2190 buf \u222a {z \u2022 \u22a5} 10 if z g(i) = \u22a5 then /*x i is a trigger*/ 11 buf \u2190 buf \u222a {z \u2022 r | r \u2208 R} 12 B \u2190K-best(buf ) 13 if y [1:h(i,k)] \u2208 B then 14 return B[0] /*for early-update*/ 15 return B[0]", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "f 1 (x, i, y g(i) ) = 1 if y g(i) = Attack and xi = \"fire\" 0 otherwise", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "The global features, by contrast, involve longer range of the output structure. Formally, each global feature function takes the form of f (x, i, k, y), where i and k denote the indices of the current token and argument candidate in decoding, respectively. The following indicator function is a simple example of global features: . unigrams/bigrams of the current and context words within the window of size 2 2. unigrams/bigrams of part-of-speech tags of the current and context words within the window of size 2 3. lemma and synonyms of the current token 4. base form of the current token extracted from Nomlex (Macleod et al., 1998) 5. Brown clusters that are learned from ACE English corpus (Brown et al., 1992; Miller et al., 2004; Sun et al., 2011) . We used the clusters with prefixes of length 13, 16 and 20 for each token.", "cite_spans": [ { "start": 613, "end": 635, "text": "(Macleod et al., 1998)", "ref_id": "BIBREF14" }, { "start": 695, "end": 715, "text": "(Brown et al., 1992;", "ref_id": "BIBREF1" }, { "start": 716, "end": 736, "text": "Miller et al., 2004;", "ref_id": "BIBREF16" }, { "start": 737, "end": 754, "text": "Sun et al., 2011)", "ref_id": "BIBREF21" } ], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "f 101 (x, i, k, y) = \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 1 if y g(i) =", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Features", "sec_num": "3.4" }, { "text": "6. dependent and governor words of the current token 7. dependency types associated the current token 8. whether the current token is a modifier of job title 9. whether the current token is a non-referential pronoun Entity Information 10. unigrams/bigrams normalized by entity types 11. dependency features normalized by entity types 12. nearest entity type and string in the sentence/clause Argument Basic 1. context words of the entity mention 2. trigger word and subtype 3. entity type, subtype and entity role if it is a geo-political entity mention 4. entity mention head, and head of any other name mention from co-reference chain 5. lexical distance between the argument candidate and the trigger 6. the relative position between the argument candidate and the trigger: {before, after, overlap, or separated by punctuation} 7. whether it is the nearest argument candidate with the same type 8. whether it is the only mention of the same entity type in the sentence Syntactic 9. dependency path between the argument candidate and the trigger 10. path from the argument candidate and the trigger in constituent parse tree 11. length of the path between the argument candidate and the trigger in dependency graph 12. common root node and its depth of the argument candidate and parse tree 13. whether the argument candidate and the trigger appear in the same clause ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Syntactic", "sec_num": null }, { "text": "In general there are two kinds of local features:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "Trigger features The local feature function for trigger labeling can be factorized as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "f (x, i, y g(i) ) = p(x, i) \u2022 q(y g(i) ), where p(x, i)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "is a predicate about the input, which we call text feature, and q(y g(i) ) is a predicate on the trigger label. In practice, we define two versions of q(y g(i) ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "q 0 (y g(i) ) = y g(i) (event subtype) q 1 (y g(i) ) = event type of y g(i) q 1 (y g(i)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": ") is a backoff version of the standard unigram feature. Some text features for the same event type may share a certain distributional similarity regardless of the subtypes. For example, if the nearest entity mention is \"Company\", the current token is likely to be Personnel no matter whether it is End-Postion or Start-Position.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "Argument features Similarly, the local feature function for argument labeling can be rep-", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "resented as f (x, i, k, y g(i) , y h(i,k) ) = p(x, i, k) \u2022 q(y g(i) , y h(i,k) ),", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "where y h(i,k) denotes the argument assignment for the edge between trigger word i and argument candidate e k . We define two versions of q(y g(i) , y h(i,k) ):", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "q 0 (y g(i) , y h(i,k) ) = \uf8f1 \uf8f4 \uf8f2 \uf8f4 \uf8f3 y h(i,k) if y h(i,k) is Place, Time or None y g(i) \u2022 y h(i,k) otherwise q 1 (y g(i) , y h(i,k) ) = 1 if y h(i,k) =None 0 otherwise", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "It is notable that Place and Time arguments are applicable and behave similarly to all event subtypes. Therefore features for these arguments are not conjuncted with trigger labels. q 1 (y h(i,k) ) can be considered as a backoff version of q 0 (y h(i,k)) , which does not discriminate different argument roles but only focuses on argument identification. Table 1 summarizes the text features about the input for trigger and argument labeling. In our experiments, we used the Stanford parser (De Marneffe et al., 2006) to create dependency parses. ", "cite_spans": [ { "start": 495, "end": 517, "text": "Marneffe et al., 2006)", "ref_id": "BIBREF6" } ], "ref_spans": [ { "start": 355, "end": 362, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "Local features", "sec_num": "3.4.1" }, { "text": "Feature Description Trigger", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "1. bigram of trigger types occur in the same sentence or the same clause 2. binary feature indicating whether synonyms in the same sentence have the same trigger label 3. context and dependency paths between two triggers conjuncted with their types Argument 4. context and dependency features about two argument candidates which share the same role within the same event mention 5. features about one argument candidate which plays as arguments in two event mentions in the same sentence 6. features about two arguments of an event mention which are overlapping 7. the number of arguments with each role type of an event mention conjuncted with the event subtype 8. the pairs of time arguments within an event mention conjuncted with the event subtype Trigger global feature This type of feature captures the dependencies between two triggers within the same sentence. For instance: feature (1) captures the co-occurrence of trigger types. This kind of feature is motivated by the fact that two event mentions in the same sentence tend to be semantically coherent. As an example, from Table 3 we can see that Attack event often co-occur with Die event in the same sentence, but rarely co-occur with Start-Position event. Feature (2) encourages synonyms or identical tokens to have the same label. Feature (3) exploits the lexical and syntactic relation between two triggers. A simple example is whether an Attack trigger and a Die trigger are linked by the dependency relation conj and.", "cite_spans": [], "ref_spans": [ { "start": 1085, "end": 1092, "text": "Table 3", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "Argument global feature This type of feature is defined over multiple arguments for the same or different triggers. Consider the following sentence:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "(3) Trains running to southern Sudan were used to transport abducted women and children.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "The Transport event mention \"transport\" has two Artifact arguments, \"women\" and \"children\".", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "The dependency edge conj and between \"women\" and \"children\" indicates that they should play the same role in the event mention. The triangle structure in Figure 5(a) is an example of feature (4) for the above example. This feature encourages entities that are linked by dependency relation conj and to play the same role Artifact in any Transport event.", "cite_spans": [], "ref_spans": [ { "start": 154, "end": 165, "text": "Figure 5(a)", "ref_id": "FIGREF4" } ], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "Similarly, Figure 5 (b) depicts an example of feature (5) for sentence (1) in Section 1. In this example, an entity mention is Victim argument to Die event and Target argument to Attack event, and the two event triggers are connected by the typed dependency advcl. Here advcl means that the word \"fired\" is an adverbial clause modier of \"died\". The job title \"co-chief executive of Vivendi Universal Entertainment\" overlaps with the Organization mention \"Vivendi Universal Entertainment\". The feature in the triangle shape can be considered as a soft constraint such that if a Job-Title mention is a Position argument to an End-Position trigger, then the Organization mention which appears at the end of it should be labeled as Entity argument for the same trigger. Feature (7-8) are based on the statistics about different arguments for the same trigger. For instance, in many cases, a trigger can only have one Place argument. If a partial configuration mistakenly classifies more than one entity mention as Place arguments for the same trigger, then it will be penalized.", "cite_spans": [], "ref_spans": [ { "start": 11, "end": 19, "text": "Figure 5", "ref_id": "FIGREF4" } ], "eq_spans": [], "section": "Global features", "sec_num": "3.4.2" }, { "text": "We utilized the ACE 2005 corpus as our testbed. For comparison, we used the same test set with 40 newswire articles (672 sentences) as in (Ji and Grishman, 2008; Liao and Grishman, 2010) for the experiments, and randomly selected 30 other documents (863 sentences) from different genres as the development set. The rest 529 documents (14, 840 sentences) are used for training.", "cite_spans": [ { "start": 138, "end": 161, "text": "(Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 162, "end": 186, "text": "Liao and Grishman, 2010)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Data set and evaluation metric", "sec_num": "4.1" }, { "text": "Following previous work (Ji and Grishman, 2008; Liao and Grishman, 2010; Hong et al., 2011) , we use the following criteria to determine the correctness of an predicted event mention:", "cite_spans": [ { "start": 24, "end": 47, "text": "(Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 48, "end": 72, "text": "Liao and Grishman, 2010;", "ref_id": "BIBREF13" }, { "start": 73, "end": 91, "text": "Hong et al., 2011)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Data set and evaluation metric", "sec_num": "4.1" }, { "text": "\u2022 A trigger is correct if its event subtype and offsets match those of a reference trigger. \u2022 An argument is correctly identified if its event subtype and offsets match those of any of the reference argument mentions. \u2022 An argument is correctly identified and classified if its event subtype, offsets and argument role match those of any of the reference argument mentions.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data set and evaluation metric", "sec_num": "4.1" }, { "text": "Finally we use Precision (P), Recall (R) and Fmeasure (F 1 ) to evaluate the overall performance.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data set and evaluation metric", "sec_num": "4.1" }, { "text": "Chen and Ng (2012) have proven that performing identification and classification in one step is better than two steps. To compare our proposed method with the previous pipelined approaches, we implemented two Maximum Entropy (Max-Ent) classifiers for trigger labeling and argument labeling respectively. To make a fair comparison, the feature sets in the baseline are identical to the local text features we developed in our framework (see Figure 1 ).", "cite_spans": [], "ref_spans": [ { "start": 440, "end": 448, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Baseline system", "sec_num": "4.2" }, { "text": "We use the harmonic mean of the trigger's F 1 measure and argument's F 1 measure to measure the performance on the development set. Figure 6 shows the training curves of the averaged perceptron with respect to the performance on the development set when the beam size is 4. As we can see both curves converge around iteration 20 and the global features improve the overall performance, compared to its counterpart with only local features. Therefore we set the number of iterations as 20 in the remaining experiments.", "cite_spans": [], "ref_spans": [ { "start": 132, "end": 140, "text": "Figure 6", "ref_id": "FIGREF6" } ], "eq_spans": [], "section": "Training curves", "sec_num": "4.3" }, { "text": "The beam size is an important hyper parameter in both training and test. Larger beam size will increase the computational cost while smaller beam size may reduce the performance. Table 4 shows the performance on the development set with several different beam sizes. When beam size = 4, the algorithm achieved the highest performance on the development set with trigger F 1 = 67.9, argument F 1 = 51.5, and harmonic mean = 58.6. When the size is increased to 32, the accuracy was not improved. Based on this observation, we chose beam size as 4 for the remaining experiments. -update vs. standard-update Huang et al. (2012) define \"invalid update\" to be an update that does not fix a violation (and instead reinforces the error), and show that it strongly (anti-)correlates with search quality and learning quality. Figure 7 depicts the percentage of invalid updates in standard-update with and without global features, respectively. With global features, there are numerous invalid updates when the Figure 7 : Percentage of the so-called \"invalid updates\" (Huang et al., 2012) 67.9 51.5 67.5 52.7 Table 5 : Comparison between the performance (%) of standard-update and early-update with global features. Here b stands for beam size.", "cite_spans": [ { "start": 576, "end": 623, "text": "-update vs. standard-update Huang et al. (2012)", "ref_id": null }, { "start": 1057, "end": 1077, "text": "(Huang et al., 2012)", "ref_id": "BIBREF9" } ], "ref_spans": [ { "start": 179, "end": 186, "text": "Table 4", "ref_id": null }, { "start": 816, "end": 824, "text": "Figure 7", "ref_id": null }, { "start": 1000, "end": 1008, "text": "Figure 7", "ref_id": null }, { "start": 1098, "end": 1105, "text": "Table 5", "ref_id": null } ], "eq_spans": [], "section": "Impact of beam size", "sec_num": "4.4" }, { "text": "beam size is small. The ratio decreases monotonically as beam size increases. The model with only local features made much smaller numbers of invalid updates, which suggests that the use of global features makes the search problem much harder. This observation justify the application of early-update in this work. To further investigate the difference between early-update and standardupdate, we tested the performance of both strategies, which is summarized in Table 5 . As we can see the performance of standard-update is generally worse than early-update. When the beam size is increased (b = 4), the gap becomes smaller as the ratio of invalid updates is reduced. Table 6 shows the overall performance on the blind test set. In addition to our baseline, we compare against the sentence-level system reported in Hong et al. (2011) , which, to the best of our knowledge, is the best-reported system in the literature based on gold standard argument candidates. The proposed joint framework with local features achieves comparable performance for triggers and outperforms the staged baseline especially on arguments. By adding global features, the overall performance is further improved significantly. Compared to the staged baseline, it gains 1.6% improvement on trigger's F-measure and 8.8% improvement on argument's F-measure. Remarkably, compared to the cross-entity approach reported in (Hong et al., 2011) , which attained 68.3% F 1 for triggers and 48.3% for arguments, our approach with global features achieves even better performance on argument labeling although we only used sentencelevel information. We also tested the performance with argument candidates automatically extracted by a highperforming name tagger (Li et al., 2012b ) and an IE system (Grishman et al., 2005) . The results are summarized in Table 7 . The joint approach with global features significantly outperforms the baseline and the model with only local features. We also show that it outperforms the sentencelevel baseline reported in (Ji and Grishman, 2008; Liao and Grishman, 2010) , both of which attained 59.7% F 1 for triggers and 36.6% for arguments. Our approach aims to tackle the problem of sentence-level event extraction, thereby only used intra-sentential evidence. Nevertheless, the performance of our approach is still comparable with the best-reported methods based on cross-document and cross-event inference (Ji and Grishman, 2008; Liao and Grishman, 2010) .", "cite_spans": [ { "start": 816, "end": 834, "text": "Hong et al. (2011)", "ref_id": "BIBREF8" }, { "start": 1395, "end": 1414, "text": "(Hong et al., 2011)", "ref_id": "BIBREF8" }, { "start": 1729, "end": 1746, "text": "(Li et al., 2012b", "ref_id": "BIBREF12" }, { "start": 1766, "end": 1789, "text": "(Grishman et al., 2005)", "ref_id": "BIBREF7" }, { "start": 2023, "end": 2046, "text": "(Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 2047, "end": 2071, "text": "Liao and Grishman, 2010)", "ref_id": "BIBREF13" }, { "start": 2413, "end": 2436, "text": "(Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 2437, "end": 2461, "text": "Liao and Grishman, 2010)", "ref_id": "BIBREF13" } ], "ref_spans": [ { "start": 463, "end": 470, "text": "Table 5", "ref_id": null }, { "start": 669, "end": 676, "text": "Table 6", "ref_id": "TABREF8" }, { "start": 1822, "end": 1829, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Early", "sec_num": "4.5" }, { "text": "Most recent studies about ACE event extraction rely on staged pipeline which consists of separate local classifiers for trigger labeling and argument labeling (Grishman et al., 2005; Ahn, 2006; Ji and Grishman, 2008; Chen and Ji, 2009; Liao and Grishman, 2010; Hong et al., 2011; Li et al., 2012a; Chen and Ng, 2012) . To the best of our knowledge, our work is the first attempt to jointly model these two ACE event subtasks. Argument Role (%)", "cite_spans": [ { "start": 159, "end": 182, "text": "(Grishman et al., 2005;", "ref_id": "BIBREF7" }, { "start": 183, "end": 193, "text": "Ahn, 2006;", "ref_id": "BIBREF0" }, { "start": 194, "end": 216, "text": "Ji and Grishman, 2008;", "ref_id": "BIBREF10" }, { "start": 217, "end": 235, "text": "Chen and Ji, 2009;", "ref_id": "BIBREF2" }, { "start": 236, "end": 260, "text": "Liao and Grishman, 2010;", "ref_id": "BIBREF13" }, { "start": 261, "end": 279, "text": "Hong et al., 2011;", "ref_id": "BIBREF8" }, { "start": 280, "end": 297, "text": "Li et al., 2012a;", "ref_id": "BIBREF11" }, { "start": 298, "end": 316, "text": "Chen and Ng, 2012)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "P R F 1 P R F 1 P R F 1 P R F 1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "Sentence-level in Hong et al. (2011) Joint w/ local + global 65.6 (\u21931.9) 41.8 (\u219310.9) Table 7 : Overall performance (%) with predicted entities, timex, and values. \u2193 indicates the performance drop from experiments with gold-standard argument candidates (see Table 6 ).", "cite_spans": [ { "start": 18, "end": 36, "text": "Hong et al. (2011)", "ref_id": "BIBREF8" } ], "ref_spans": [ { "start": 86, "end": 93, "text": "Table 7", "ref_id": null }, { "start": 258, "end": 265, "text": "Table 6", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "For the Message Understanding Conference (MUC) and FAS Program for Monitoring Emerging Diseases (ProMED) event extraction tasks, Patwardhan and Riloff (2009) proposed a probabilistic framework to extract event role fillers conditioned on the sentential event occurrence. Besides having different task definitions, the key difference from our approach is that their role filler recognizer and sentential event recognizer are trained independently but combined in the test stage. Our experiments, however, have demonstrated that it is more advantageous to do both training and testing with joint inference.", "cite_spans": [ { "start": 129, "end": 157, "text": "Patwardhan and Riloff (2009)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "There has been some previous work on joint modeling for biomedical events (Riedel and Mc-Callum, 2011a; Riedel et al., 2009; McClosky et al., 2011; Riedel and McCallum, 2011b) . (Mc-Closky et al., 2011) is most closely related to our approach. They casted the problem of biomedical event extraction as a dependency parsing problem. The key assumption that event structure can be considered as trees is incompatible with ACE event extraction. In addition, they used a separate classifier to predict the event triggers before applying the parser, while we extract the triggers and argument jointly. Finally, the features in the parser are edge-factorized. To exploit global features, they applied a MaxEnt-based global re-ranker. In comparison, our approach is a unified framework based on beam search, which allows us to exploit arbitrary global features efficiently.", "cite_spans": [ { "start": 74, "end": 103, "text": "(Riedel and Mc-Callum, 2011a;", "ref_id": null }, { "start": 104, "end": 124, "text": "Riedel et al., 2009;", "ref_id": "BIBREF20" }, { "start": 125, "end": 147, "text": "McClosky et al., 2011;", "ref_id": "BIBREF15" }, { "start": 148, "end": 175, "text": "Riedel and McCallum, 2011b)", "ref_id": "BIBREF19" }, { "start": 178, "end": 202, "text": "(Mc-Closky et al., 2011)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "5" }, { "text": "We presented a joint framework for ACE event extraction based on structured perceptron with inexact search. As opposed to traditional pipelined approaches, we re-defined the task as a structured prediction problem. The experiments proved that the perceptron with local features outperforms the staged baseline and the global features further improve the performance significantly, surpassing the current state-of-the-art by a large margin.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "As shown in Table 7 , the overall performance drops substantially when using predicted argument candidates. To improve the accuracy of endto-end IE system, we plan to develop a complete joint framework to recognize entities together with event mentions for future work. Also we are interested in applying this framework to other IE tasks such as relation extraction.", "cite_spans": [], "ref_spans": [ { "start": 12, "end": 19, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "http://projects.ldc.upenn.edu/ace/docs/English-Events-Guidelines v5.4.3.pdf", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "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": "Class-based n-gram models of natural language", "authors": [ { "first": "", "middle": [], "last": "Peter F Brown", "suffix": "" }, { "first": "V", "middle": [], "last": "Peter", "suffix": "" }, { "first": "", "middle": [], "last": "Desouza", "suffix": "" }, { "first": "L", "middle": [], "last": "Robert", "suffix": "" }, { "first": "Vincent J Della", "middle": [], "last": "Mercer", "suffix": "" }, { "first": "Jenifer C", "middle": [], "last": "Pietra", "suffix": "" }, { "first": "", "middle": [], "last": "Lai", "suffix": "" } ], "year": 1992, "venue": "Computational linguistics", "volume": "18", "issue": "4", "pages": "467--479", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peter F Brown, Peter V Desouza, Robert L Mercer, Vincent J Della Pietra, and Jenifer C Lai. 1992. Class-based n-gram models of natural language. Computational linguistics, 18(4):467-479.", "links": null }, "BIBREF2": { "ref_id": "b2", "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 }, "BIBREF3": { "ref_id": "b3", "title": "Joint modeling for chinese event extraction with rich linguistic features", "authors": [ { "first": "Chen", "middle": [], "last": "Chen", "suffix": "" }, { "first": "Vincent", "middle": [], "last": "Ng", "suffix": "" } ], "year": 2012, "venue": "COLING", "volume": "", "issue": "", "pages": "529--544", "other_ids": {}, "num": null, "urls": [], "raw_text": "Chen Chen and Vincent Ng. 2012. Joint modeling for chinese event extraction with rich linguistic features. In COLING, pages 529-544.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Incremental parsing with the perceptron algorithm", "authors": [ { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Roark", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael Collins and Brian Roark. 2004. Incremental parsing with the perceptron algorithm. In Proceed- ings of the 42nd Annual Meeting on Association for Computational Linguistics, page 111.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Discriminative training methods for hidden markov models: Theory and experiments with perceptron algorithms", "authors": [ { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the ACL-02 conference on Empirical methods in natural language processing", "volume": "10", "issue": "", "pages": "1--8", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael Collins. 2002. Discriminative training meth- ods for hidden markov models: Theory and exper- iments with perceptron algorithms. In Proceedings of the ACL-02 conference on Empirical methods in natural language processing-Volume 10, pages 1-8.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Generating typed dependency parses from phrase structure parses", "authors": [ { "first": "Marie-Catherine De", "middle": [], "last": "Marneffe", "suffix": "" }, { "first": "Bill", "middle": [], "last": "Maccartney", "suffix": "" }, { "first": "Christopher D", "middle": [], "last": "Manning", "suffix": "" } ], "year": 2006, "venue": "Proceedings of LREC", "volume": "6", "issue": "", "pages": "449--454", "other_ids": {}, "num": null, "urls": [], "raw_text": "Marie-Catherine De Marneffe, Bill MacCartney, and Christopher D Manning. 2006. Generating typed dependency parses from phrase structure parses. In Proceedings of LREC, volume 6, pages 449-454.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Nyu's english ace 2005 system description", "authors": [ { "first": "Ralph", "middle": [], "last": "Grishman", "suffix": "" }, { "first": "David", "middle": [], "last": "Westbrook", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Meyers", "suffix": "" } ], "year": 2005, "venue": "Proceedings of ACE 2005 Evaluation Workshop", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ralph Grishman, David Westbrook, and Adam Meyers. 2005. Nyu's english ace 2005 system description. In Proceedings of ACE 2005 Evaluation Workshop. Washington.", "links": null }, "BIBREF8": { "ref_id": "b8", "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": "Jian-Min", "middle": [], "last": "Yao", "suffix": "" }, { "first": "Guodong", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "Qiaoming", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2011, "venue": "Proceedings of ACL", "volume": "", "issue": "", "pages": "1127--1136", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yu Hong, Jianfeng Zhang, Bin Ma, Jian-Min Yao, Guodong Zhou, and Qiaoming Zhu. 2011. Using cross-entity inference to improve event extraction. In Proceedings of ACL, pages 1127-1136.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Structured perceptron with inexact search", "authors": [ { "first": "Liang", "middle": [], "last": "Huang", "suffix": "" }, { "first": "Suphan", "middle": [], "last": "Fayong", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Guo", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "142--151", "other_ids": {}, "num": null, "urls": [], "raw_text": "Liang Huang, Suphan Fayong, and Yang Guo. 2012. Structured perceptron with inexact search. In Pro- ceedings of the 2012 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 142-151.", "links": null }, "BIBREF10": { "ref_id": "b10", "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", "volume": "", "issue": "", "pages": "254--262", "other_ids": {}, "num": null, "urls": [], "raw_text": "Heng Ji and Ralph Grishman. 2008. Refining event ex- traction through cross-document inference. In Pro- ceedings of ACL, pages 254-262.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Employing compositional semantics and discourse consistency in chinese event extraction", "authors": [ { "first": "Peifeng", "middle": [], "last": "Li", "suffix": "" }, { "first": "Guodong", "middle": [], "last": "Zhou", "suffix": "" }, { "first": "Qiaoming", "middle": [], "last": "Zhu", "suffix": "" }, { "first": "Libin", "middle": [], "last": "Hou", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning", "volume": "", "issue": "", "pages": "1006--1016", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peifeng Li, Guodong Zhou, Qiaoming Zhu, and Li- bin Hou. 2012a. Employing compositional seman- tics and discourse consistency in chinese event ex- traction. In Proceedings of the 2012 Joint Confer- ence on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1006-1016.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Joint bilingual name tagging for parallel corpora", "authors": [ { "first": "Qi", "middle": [], "last": "Li", "suffix": "" }, { "first": "Haibo", "middle": [], "last": "Li", "suffix": "" }, { "first": "Heng", "middle": [], "last": "Ji", "suffix": "" }, { "first": "Wen", "middle": [], "last": "Wang", "suffix": "" }, { "first": "Jing", "middle": [], "last": "Zheng", "suffix": "" }, { "first": "Fei", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the 21st ACM international conference on Information and knowledge management", "volume": "", "issue": "", "pages": "1727--1731", "other_ids": {}, "num": null, "urls": [], "raw_text": "Qi Li, Haibo Li, Heng Ji, Wen Wang, Jing Zheng, and Fei Huang. 2012b. Joint bilingual name tagging for parallel corpora. In Proceedings of the 21st ACM international conference on Information and knowl- edge management, pages 1727-1731.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Using document level cross-event inference to improve event extraction", "authors": [ { "first": "Shasha", "middle": [], "last": "Liao", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Grishman", "suffix": "" } ], "year": 2010, "venue": "Proceedings of ACL", "volume": "", "issue": "", "pages": "789--797", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shasha Liao and Ralph Grishman. 2010. Using doc- ument level cross-event inference to improve event extraction. In Proceedings of ACL, pages 789-797.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Nomlex: A lexicon of nominalizations", "authors": [ { "first": "Catherine", "middle": [], "last": "Macleod", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Grishman", "suffix": "" }, { "first": "Adam", "middle": [], "last": "Meyers", "suffix": "" }, { "first": "Leslie", "middle": [], "last": "Barrett", "suffix": "" }, { "first": "Ruth", "middle": [], "last": "Reeves", "suffix": "" } ], "year": 1998, "venue": "Proceedings of EU-RALEX", "volume": "98", "issue": "", "pages": "187--193", "other_ids": {}, "num": null, "urls": [], "raw_text": "Catherine Macleod, Ralph Grishman, Adam Meyers, Leslie Barrett, and Ruth Reeves. 1998. Nomlex: A lexicon of nominalizations. In Proceedings of EU- RALEX, volume 98, pages 187-193.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Event extraction as dependency parsing", "authors": [ { "first": "David", "middle": [], "last": "Mcclosky", "suffix": "" }, { "first": "Mihai", "middle": [], "last": "Surdeanu", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2011, "venue": "Proceedings of ACL", "volume": "", "issue": "", "pages": "1626--1635", "other_ids": {}, "num": null, "urls": [], "raw_text": "David McClosky, Mihai Surdeanu, and Christopher D. Manning. 2011. Event extraction as dependency parsing. In Proceedings of ACL, pages 1626-1635.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Name tagging with word clusters and discriminative training", "authors": [ { "first": "Scott", "middle": [], "last": "Miller", "suffix": "" }, { "first": "Jethran", "middle": [], "last": "Guinness", "suffix": "" }, { "first": "Alex", "middle": [], "last": "Zamanian", "suffix": "" } ], "year": 2004, "venue": "Proceedings of HLT-NAACL", "volume": "4", "issue": "", "pages": "337--342", "other_ids": {}, "num": null, "urls": [], "raw_text": "Scott Miller, Jethran Guinness, and Alex Zamanian. 2004. Name tagging with word clusters and discrim- inative training. In Proceedings of HLT-NAACL, volume 4, pages 337-342.", "links": null }, "BIBREF17": { "ref_id": "b17", "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.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Fast and robust joint models for biomedical event extraction", "authors": [ { "first": "Sebastian", "middle": [], "last": "Riedel", "suffix": "" }, { "first": "Andrew", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1--12", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sebastian Riedel and Andrew McCallum. 2011a. Fast and robust joint models for biomedical event extrac- tion. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 1- 12.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Robust biomedical event extraction with dual decomposition and minimal domain adaptation", "authors": [ { "first": "Sebastian", "middle": [], "last": "Riedel", "suffix": "" }, { "first": "Andrew", "middle": [], "last": "Mccallum", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the BioNLP Shared Task 2011 Workshop", "volume": "", "issue": "", "pages": "46--50", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sebastian Riedel and Andrew McCallum. 2011b. Ro- bust biomedical event extraction with dual decom- position and minimal domain adaptation. In Pro- ceedings of the BioNLP Shared Task 2011 Work- shop, pages 46-50.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "A markov logic approach to bio-molecular event extraction", "authors": [ { "first": "Sebastian", "middle": [], "last": "Riedel", "suffix": "" }, { "first": "Hong-Woo", "middle": [], "last": "Chun", "suffix": "" }, { "first": "Toshihisa", "middle": [], "last": "Takagi", "suffix": "" }, { "first": "Jun'ichi", "middle": [], "last": "Tsujii", "suffix": "" } ], "year": 2009, "venue": "Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task", "volume": "", "issue": "", "pages": "41--49", "other_ids": {}, "num": null, "urls": [], "raw_text": "Sebastian Riedel, Hong-Woo Chun, Toshihisa Takagi, and Jun'ichi Tsujii. 2009. A markov logic approach to bio-molecular event extraction. In Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task, pages 41-49.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Semi-supervised relation extraction with large-scale word clustering", "authors": [ { "first": "Ang", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Ralph", "middle": [], "last": "Grishman", "suffix": "" }, { "first": "Satoshi", "middle": [], "last": "Sekine", "suffix": "" } ], "year": 2011, "venue": "Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "521--529", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ang Sun, Ralph Grishman, and Satoshi Sekine. 2011. Semi-supervised relation extraction with large-scale word clustering. In Proceedings of the 49th An- nual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 521-529.", "links": null } }, "ref_entries": { "FIGREF1": { "num": null, "uris": null, "text": "Example notation with s = 3, m = 2.", "type_str": "figure" }, "FIGREF2": { "num": null, "uris": null, "text": "Decoding algorithm for event extraction. z \u2022l means appending label l to the end of z. During test, lines 4-5 & 13-14 are omitted.", "type_str": "figure" }, "FIGREF3": { "num": null, "uris": null, "text": "Attack and y has only one \"Attacker\"", "type_str": "figure" }, "FIGREF4": { "num": null, "uris": null, "text": "Illustration of global features (4-6) in", "type_str": "figure" }, "FIGREF5": { "num": null, "uris": null, "text": "(c) shows an example of feature (6) for the following sentence:(4) Barry Diller resigned as co-chief executive of Vivendi Universal Entertainment.", "type_str": "figure" }, "FIGREF6": { "num": null, "uris": null, "text": "Training curves on dev set.", "type_str": "figure" }, "TABREF0": { "num": null, "type_str": "table", "content": "", "html": null, "text": "Local features." }, "TABREF1": { "num": null, "type_str": "table", "content": "
", "html": null, "text": "" }, "TABREF2": { "num": null, "type_str": "table", "content": "
", "html": null, "text": "Global features." }, "TABREF3": { "num": null, "type_str": "table", "content": "
EventProbability
Attack0.34
Die0.14
Transport0.08
Injure0.04
Meet0.02
", "html": null, "text": "" }, "TABREF4": { "num": null, "type_str": "table", "content": "", "html": null, "text": "Top 5 event subtypes that co-occur with Attack event in the same sentence." }, "TABREF5": { "num": null, "type_str": "table", "content": "
Beam size12481632
Training time (sec) 993 2,034 3,982 8,036 15,878 33,026
Harmonic mean57.6 57.758.658.057.857.8
% of invalid updates2 Table 4: 1 4 beam size 8 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.4516 local+global local32
", "html": null, "text": "Comparison of training time and accuracy on the dev set." }, "TABREF8": { "num": null, "type_str": "table", "content": "
MethodsTrigger F 1Arg F 1
Ji and Grishman (2008)67.342.6
cross-doc Inference
Ji and Grishman (2008)59.736.6
sentence-level
MaxEnt classifiers64.7 (\u21931.2) 33.7 (\u219310.2)
Joint w/ local63.7 (\u21932.0) 35.8 (\u219310.7)
", "html": null, "text": "Overall performance with gold-standard entities, timex, and values. \u2020 beyond sentence level." } } } }