Joint Event Trigger Identification and Event Coreference Resolution with Structured Perceptron

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2074–2080, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Joint Event Trigger Identification and Event Coreference Resolution with Structured Perceptron Jun Araki and Teruko Mitamura Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {junaraki,teruko}@cs.cmu.edu Abstract Events and their coreference offer use- ful semantic and discourse resources. We show that the semantic and dis- course aspects of events interact with each other. However, traditional approaches ad- dressed event extraction and event coref- erence resolution either separately or se- quentially, which limits their interactions. This paper proposes a document-level structured learning model that simultane- ously identifies event triggers and resolves event coreference. We demonstrate that the joint model outperforms a pipelined model by 6.9 BLANC F1 and 1.8 CoNLL F1 points in event coreference resolution using a corpus in the biology domain. 1 Introduction Events convey semantic information such as who did what to whom where and when. They also corefer to each other, playing a role of discourse connection points to form a coherent story. These aspects of events have been already utilized in a wide variety of natural language processing (NLP) applications, such as automated population of knowledge bases (Ji and Grishman, 2011), topic detection and tracking (Allan, 2002), question an- swering (Bikel and Castelli, 2008), text summa- rization (Li et al., 2006), and contradiction detec- tion (de Marneffe et al., 2008). This fact illustrates the importance of event extraction and event coref- erence resolution. Those semantic and discourse aspects of events are not independent from each other, and in fact often work in interactive manners. We give two examples of the interactions: (1) British bank Barclays had agreed to buy(E1) Spanish rival Banco Zaragozano for 1.14 billion euros. The combination(E2) of the banking operations of Barclays Spain and Zaragozano will bring together two complementary businesses. (2) The Palestinian Authority condemned the attack(E3), saying it(E4) would divert international sympathy away from the far higher Palestinian civilian death toll. E1 corefers to E2, and E3 does to E4. E2 is more abstract than E1, and has less evidence of being an event. E4 is a pronoun, and thus may seem to refer to an entity rather than an event. Thus, E2 and E4 are relatively difficult to be recognized as events by themselves. However, event coref- erence E1-E2, which is supported primarily by E2’s participants Barclays and Zaragozano shared with E1, helps determine that E2 is an event. The same logic applies to E3 and E4. On the other hand, previous works typically rely on a pipelined model that extracts events (e.g., E1 and E3) at the first stage, and then resolves event corefer- ence at the second stage. Although this modularity is preferable from development perspectives, the pipelined model limits the interactions. That is, the first stage alone is unlikely to detect E2 and E4 as events due to the difficulties described above. These missing events make it impossible for the second stage …

Content truncated to 500 words