Supervised Models for Coreference Resolution

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 968–977, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Supervised Models for Coreference Resolution Altaf Rahman and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {altaf,vince}@hlt.utdallas.edu Abstract Traditional learning-based coreference re- solvers operate by training a mention- pair classifier for determining whether two mentions are coreferent or not. Two in- dependent lines of recent research have attempted to improve these mention-pair classifiers, one by learning a mention- ranking model to rank preceding men- tions for a given anaphor, and the other by training an entity-mention classifier to determine whether a preceding clus- ter is coreferent with a given mention. We propose a cluster-ranking approach to coreference resolution that combines the strengths of mention rankers and entity- mention models. We additionally show how our cluster-ranking framework natu- rally allows discourse-new entity detection to be learned jointly with coreference res- olution. Experimental results on the ACE data sets demonstrate its superior perfor- mance to competing approaches. 1 Introduction Noun phrase (NP) coreference resolution is the task of identifying which NPs (or mentions) re- fer to the same real-world entity or concept. Tra- ditional learning-based coreference resolvers op- erate by training a model for classifying whether two mentions are co-referring or not (e.g., Soon et al. (2001), Ng and Cardie (2002b), Kehler et al. (2004), Ponzetto and Strube (2006)). Despite their initial successes, these mention-pair models have at least two major weaknesses. First, since each candidate antecedent for a mention to be resolved (henceforth an active mention) is considered inde- pendently of the others, these models only deter- mine how good a candidate antecedent is relative to the active mention, but not how good a candi- date antecedent is relative to other candidates. In other words, they fail to answer the critical ques- tion of which candidate antecedent is most prob- able. Second, they have limitations in their ex- pressiveness: the information extracted from the two mentions alone may not be sufficient for mak- ing an informed coreference decision, especially if the candidate antecedent is a pronoun (which is se- mantically empty) or a mention that lacks descrip- tive information such as gender (e.g., Clinton). To address the first weakness, researchers have attempted to train a mention-ranking model for determining which candidate antecedent is most probable given an active mention (e.g., Denis and Baldridge (2008)). Ranking is arguably a more natural reformulation of coreference resolution than classification, as a ranker allows all candidate antecedents to be considered simultaneously and therefore directly captures the competition among them. Another desirable consequence is that there exists a natural resolution strategy for a ranking approach: a mention is resolved to the candidate antecedent that has the highest rank. This con- trasts with classification-based approaches, where many clustering algorithms have been employed to co-ordinate the pairwise coreference decisions (because it is unclear which one is the best). To address the second weakness, researchers have investigated the acquisition of …

Content truncated to 500 words