Adaptive Clustering for Coreference Resolution with Deterministic Rules and Web-Based Language Models

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 11–19, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Adaptive Clustering for Coreference Resolution with Deterministic Rules and Web-Based Language Models Razvan C. Bunescu School of EECS Ohio University Athens, OH 45701, USA bunescu@ohio.edu Abstract We present a novel adaptive clustering model for coreference resolution in which the expert rules of a state of the art deterministic sys- tem are used as features over pairs of clus- ters. A significant advantage of the new ap- proach is that the expert rules can be eas- ily augmented with new semantic features. We demonstrate this advantage by incorporat- ing semantic compatibility features for neutral pronouns computed from web n-gram statis- tics. Experimental results show that the com- bination of the new features with the expert rules in the adaptive clustering approach re- sults in an overall performance improvement, and over 5% improvement in F1 measure for the target pronouns when evaluated on the ACE 2004 newswire corpus. 1 Introduction Coreference resolution is the task of clustering a sequence of textual entity mentions into a set of maximal non-overlapping clusters, such that men- tions in a cluster refer to the same discourse entity. Coreference resolution is an important subtask in a wide array of natural language processing prob- lems, among them information extraction, question answering, and machine translation. The availabil- ity of corpora annotated with coreference relations has led to the development of a diverse set of super- vised learning approaches for coreference. While learning models enjoy a largely undisputed role in many NLP applications, deterministic models based on rich sets of expert rules for coreference have been shown recently to achieve performance rivaling, if not exceeding, the performance of state of the art machine learning approaches (Haghighi and Klein, 2009; Raghunathan et al., 2010). In particular, the top performing system in the CoNLL 2011 shared task (Pradhan et al., 2011) is a multi-pass system that applies tiers of deterministic coreference sieves from highest to lowest precision (Lee et al., 2011). The PRECISECONSTRUCTS sieve, for example, creates coreference links between mentions that are found to match patterns of apposition, predicate nomina- tives, acronyms, demonyms, or relative pronouns. This is a high precision sieve, correspondingly it is among the first sieves to be applied. The PRONOUN- MATCH sieve links an anaphoric pronoun with the first antecedent mention that agrees in number and gender with the pronoun, based on an ordering of the antecedents that uses syntactic rules to model dis- course salience. This is the last sieve to be applied, due to its lower overall precision, as estimated on development data. While very successful, this de- terministic multi-pass sieve approach to coreference can nevertheless be quite unwieldy when one seeks to integrate new sources of knowledge in order to improve the resolution performance. Pronoun reso- lution, for example, was shown by Yang et al. (2005) to benefit from semantic compatibility information extracted from search engine statistics. The seman- tic compatibility between candidate …

Content truncated to 500 words