Joint Learning for Event Coreference Resolution

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 90–101 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1009 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 90–101 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1009 Joint Learning for Event Coreference Resolution 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 While joint models have been developed for many NLP tasks, the vast majority of event coreference resolvers, including the top-performing resolvers competing in the recent TAC KBP 2016 Event Nugget De- tection and Coreference task, are pipeline- based, where the propagation of errors from the trigger detection component to the event coreference component is a ma- jor performance limiting factor. To ad- dress this problem, we propose a model for jointly learning event coreference, trig- ger detection, and event anaphoricity. Our joint model is novel in its choice of tasks and its features for capturing cross-task interactions. To our knowledge, this is the first attempt to train a mention-ranking model and employ event anaphoricity for event coreference. Our model achieves the best results to date on the KBP 2016 En- glish and Chinese datasets. 1 Introduction Within-document event coreference resolution is the task of determining which event mentions in a text refer to the same real-world event. Compared to entity coreference resolution, event coreference resolution is not only much less studied, but it is arguably more challenging. The challenge stems in part from the fact that an event coreference re- solver typically lies towards the end of the stan- dard information extraction pipeline, assuming as input the noisy outputs of its upstream compo- nents. One such component is the trigger detection system, which is responsible for identifying event triggers and determining their event subtypes. As is commonly known, trigger detection is another challenging task that is far from being solved. In fact, in the recent TAC KBP 2016 Event Nugget Detection and Coreference task, trigger detection (a.k.a. event nugget detection in KBP) is deliberately made more challenging by focus- ing only on detecting the 18 subtypes of triggers on which the KBP 2015 participating systems’ performances were the poorest (Mitamura et al., 2016). The best-performing KBP 2016 system on English trigger detection achieved only an F-score of 47 (Lu and Ng, 2016).1 Given the difficulty of trigger detection, it is conceivable that many errors will propagate from the trigger detection component to the event coref- erence component in any pipeline architecture where trigger detection precedes event corefer- ence resolution. These trigger detection errors could severely harm event coreference perfor- mance. For instance, two event mentions could be wrongly posited as coreferent if the underlying triggers were wrongly predicted to have the same subtype. Nevertheless, the top-performing sys- tems in the KBP 2016 event coreference task all adopted the aforementioned pipeline architecture (Liu et al., 2016; Lu and Ng, …

Content truncated to 500 words