Joint Determination of Anaphoricity and Coreference Resolution using Integer Programming

Proceedings of NAACL HLT 2007, pages 236–243, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Joint Determination of Anaphoricity and Coreference Resolution using Integer Programming Pascal Denis and Jason Baldridge Department of Linguistics University of Texas at Austin {denis,jbaldrid}@mail.utexas.edu Abstract Standard pairwise coreference resolution systems are subject to errors resulting from their performing anaphora identifi- cation as an implicit part of coreference resolution. In this paper, we propose an integer linear programming (ILP) for- mulation for coreference resolution which models anaphoricity and coreference as a joint task, such that each local model in- forms the other for the final assignments. This joint ILP formulation provides f- score improvements of 3.7-5.3% over a base coreference classifier on the ACE datasets. 1 Introduction The task of coreference resolution involves impos- ing a partition on a set of entity mentions in a docu- ment, where each partition corresponds to some en- tity in an underlying discourse model. Most work treats coreference resolution as a binary classifica- tion task in which each decision is made in a pair- wise fashion, independently of the others (McCarthy and Lehnert, 1995; Soon et al., 2001; Ng and Cardie, 2002b; Morton, 2000; Kehler et al., 2004). There are two major drawbacks with most sys- tems that make pairwise coreference decisions. The first is that identification of anaphora is done implic- itly as part of the coreference resolution. Two com- mon types of errors with these systems are cases where: (i) the system mistakenly identifies an an- tecedent for non-anaphoric mentions, and (ii) the system does not try to resolve an actual anaphoric mention. To reduce such errors, Ng and Cardie (2002a) and Ng (2004) use an anaphoricity classi- fier –which has the sole task of saying whether or not any antecedents should be identified for each mention– as a filter for their coreference system. They achieve higher performance by doing so; how- ever, their setup uses the two classifiers in a cascade. This requires careful determination of an anaphoric- ity threshold in order to not remove too many men- tions from consideration (Ng, 2004). This sensi- tivity is unsurprising, given that the tasks are co- dependent. The second problem is that most coreference sys- tems make each decision independently of previous ones in a greedy fashion (McCallum and Wellner, 2004). Clearly, the determination of membership of a particular mention into a partition should be condi- tioned on how well it matches the entity as a whole. Since independence between decisions is an unwar- ranted assumption for the task, models that consider a more global context are likely to be more appropri- ate. Recent work has examined such models; Luo et al. (2004) using Bell trees, and McCallum and Well- ner (2004) using conditional random fields, and Ng (2005) using rerankers. In this paper, we propose to recast the task of coreference resolution as an optimization problem, namely an integer linear programming (ILP) prob- lem. This framework has several properties that make it highly suitable for addressing the …

Content truncated to 500 words