Event Coreference Resolution with Non-Local Information

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 653–663 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 653 Event Coreference Resolution with Non-Local Information Jing Lu and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {ljwinnie,vince}@hlt.utdallas.edu Abstract Existing event coreference resolvers have largely focused on exploiting the information extracted from the local contexts of the event mentions under consideration. Hypothesizing that non-local information could also be useful for event coreference resolution, we present two extensions to a state-of-the-art joint event coreference model that involve incorporating (1) a supervised topic model for improving trigger detection by providing global context, and (2) a preprocessing module that seeks to improve event coreference by discarding un- likely candidate antecedents of an event men- tion using discourse contexts computed based on salient entities. The resulting model yields the best results reported to date on the KBP 2017 English and Chinese datasets. 1 Introduction Event coreference resolution is the task of deter- mining the event mentions in a document that refer to the same real-world event. One of its major challenges concerns error propagation: since the event coreference resolution component typically lies towards the end of the standard information extraction pipeline, the performance of an event coreference resolver can be adversely affected by errors propagated from its upstream components. The upstream component that has the largest im- pact on event coreference performance is arguably trigger detection. Recall that the goal of a trigger detector is to identify event triggers and assign an event subtype to each of them. Failure to detect triggers could therefore limit the upper bound on event coreference performance. To address error propagation, one way that has been shown to be effective for a variety of NLP tasks is to develop joint models, which allow cross- task output constraints to be learned from anno- tated training data. For event coreference, a learner can easily learn, for instance, that two coreferent event mentions must have the same event subtype, thereby allowing event coreference to influence trigger detection. Unfortunately, the vast major- ity of existing event coreference resolvers have adopted a pipeline architecture where trigger detec- tion precedes event coreference. In particular, joint models are both under-studied and under-exploited for event coreference given the usefulness they have demonstrated for other NLP tasks. One ex- ception is Lu and Ng’s (2017a) joint model, which jointly learns trigger detection and event corefer- ence and has achieved state-of-the-art results. As a structured conditional random field, the model em- ploys unary factors to encode the features specific for each task and binary/ternary factors to capture the interaction between each pair of tasks. The use of binary/ternary factors is a particularly appealing aspect of this model: it allows these cross-task in- teractions to be captured in a soft manner, enabling the learner to learn which combinations of values of the output variables are …

Content truncated to 500 words