Graph-Cut-Based Anaphoricity Determination for Coreference Resolution

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 575–583, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Graph-Cut-Based Anaphoricity Determination for Coreference Resolution Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 vince@hlt.utdallas.edu Abstract Recent work has shown that explicitly iden- tifying and filtering non-anaphoric mentions prior to coreference resolution can improve the performance of a coreference system. We present a novel approach to this task of anaphoricity determination based on graph cuts, and demonstrate its superiority to com- peting approaches by comparing their effec- tiveness in improving a learning-based coref- erence system on the ACE data sets. 1 Introduction Coreference resolution is the problem of identifying which noun phrases (NPs, or mentions) refer to the same real-world entity in a text or dialogue. Accord- ing to Webber (1979), coreference resolution can be decomposed into two complementary tasks: “(1) identifying what a text potentially makes available for anaphoric reference and (2) constraining the can- didate set of a given anaphoric expression down to one possible choice.” The first task is nowadays typ- ically formulated as an anaphoricity determination task, which aims to classify whether a given men- tion is anaphoric or not. Knowledge of anaphoric- ity could improve the precision of a coreference sys- tem, since non-anaphoric mentions do not have an antecedent and therefore do not need to be resolved. Previous work on anaphoricity determination can be broadly divided into two categories (see Poe- sio et al. (2004) for an overview). Research in the first category aims to identify specific types of non- anaphoric phrases, with some identifying pleonas- tic it (using heuristics [e.g., Paice and Husk (1987), Lappin and Leass (1994), Kennedy and Boguraev (1996)], supervised approaches [e.g., Evans (2001), M¨uller (2006), Versley et al. (2008)], and distribu- tional methods [e.g., Bergsma et al. (2008)]), and others identifying non-anaphoric definite descrip- tions (using rule-based techniques [e.g., Vieira and Poesio (2000)] and unsupervised techniques [e.g., Bean and Riloff (1999)]). On the other hand, research in the second cat- egory focuses on (1) determining the anaphoricity of all types of mentions, and (2) using the result- ing anaphoricity information to improve coreference resolution. For instance, Ng and Cardie (2002a) train an anaphoricity classifier to determine whether a mention is anaphoric, and let an independently- trained coreference system resolve only those men- tions that are classified as anaphoric. Somewhat sur- prisingly, they report that using anaphoricity infor- mation adversely affects the performance of their coreference system, as a result of an overly conser- vative anaphoricity classifier that misclassifies many anaphoric mentions as non-anaphoric. One solu- tion to this problem is to use anaphoricity infor- mation as soft constraints rather than as hard con- straints for coreference resolution. For instance, when searching for the best partition of a set of mentions, Luo (2007) combines the probabilities re- turned by an anaphoricity model and a coreference model to score a coreference partition, such that a partition is penalized …

Content truncated to 500 words