Chinese Event Coreference Resolution: Understanding the State of the Art

International Joint Conference on Natural Language Processing, pages 822–828, Nagoya, Japan, 14-18 October 2013. Chinese Event Coreference Resolution: Understanding the State of the Art Chen Chen and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {yzcchen,vince}@hlt.utdallas.edu Abstract Given the relatively small amount of work on event coreference resolution, our under- standing of the task is arguably fairly lim- ited. This makes it difficult to determine how to improve an event coreference re- solver. We seek to gain a better under- standing of the state of the art in event coreference resolution by performing the first publicly available analysis of a Chi- nese event coreference resolver. 1 Introduction Event coreference resolution is the task of deter- mining which event mentions in a text refer to the same real-world event. Compared to entity coref- erence (the task of determining which entity men- tions in a text refer to the same real-world entity), there is much less work on event coreference. Given the lesser amount of work on event coref- erence, our understanding of the task is arguably fairly limited. Specifically, while it is not sur- prising that the performance of an event corefer- ence resolver depends heavily on the quality of the output of its preprocessing components, it is not clear to what extent the noise inherent in the out- put of each preprocessing component is limiting the performance of a resolver. Note that this is- sue is more serious for event coreference than for entity coreference: since an event coreference re- solver lies towards the end of the information ex- traction pipeline, it has to rely on the noisy output produced by more upstream components than its entity counterpart. The lack of understanding of this issue makes it difficult to identify the compo- nents that need most attention. Moreover, when analyzing the errors made by a resolver, it is not clear which types of errors can be fixed by im- proving the resolution algorithm and which ones can be fixed by improving the preprocessing com- ponents. This makes it difficult to precisely de- termine how to improve the resolution algorithm. Taken together, these two issues make it difficult to understand how an event coreference resolver should be improved. Our goal in this paper is to better understand the state of the art in event coreference and provide di- rections for further work on this task by addressing the aforementioned issues, which are summarized by the following two questions. First, to what ex- tent is the noise inherent in the output of each of its upstream components limiting its performance? Second, what are the major types of errors that are attributable to (and therefore can be fixed by im- proving) the resolution algorithm? We address these two questions by presenting a systematic analysis of a state-of-the-art Chinese event coreference resolver. Our decision to focus on Chinese can be attributed in part to the lack of publicly available results on Chinese event coref- erence resolution. In …

Content truncated to 500 words