Coreference Systems Based on Kernels Methods

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 961–968 Manchester, August 2008 Coreference Systems based on Kernels Methods Yannick Versley SFB 441 University of T¨ubingen versley@sfs.uni-tuebingen.de Alessandro Moschitti DISI University of Trento moschitti@disi.unitn.it Massimo Poesio DISI University of Trento massimo.poesio@unitn.it Xiaofeng Yang Data Mining Department Institute for Infocomm Research xiaofengy@i2r.a-star.edu.sg Abstract Various types of structural information - e.g., about the type of constructions in which binding constraints apply, or about the structure of names - play a central role in coreference resolution, often in combi- nation with lexical information (as in ex- pletive detection). Kernel functions ap- pear to be a promising candidate to capture structure-sensitive similarities and com- plex feature combinations, but care is re- quired to ensure they are exploited in the best possible fashion. In this paper we propose kernel functions for three subtasks of coreference resolution - binding con- straint detection, expletive identification, and aliasing - together with an architec- ture to integrate them within the standard framework for coreference resolution. 1 Introduction Information about coreference relations–i.e., which noun phrases are mentions of the same entity–has been shown to be beneficial in a great number of NLP tasks, including information extraction (McCarthy and Lehnert 1995), text planning (Barzilay and Lapata 2005) and sum- marization (Steinberger et al. 2007). However, the performance of coreference resolvers on unrestricted text is still quite low. One reason for this is that coreference resolution requires a great deal of information, ranging from string matching to syntactic constraints to semantic knowledge to discourse salience information to c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. full common sense reasoning (Sidner 1979; Hobbs 1978, 1979; Grosz et al. 1995; Vieira and Poesio 2000; Mitkov 2002). Much of this information won’t be available to robust coreference resolvers until better methods are found to represent and encode common sense knowledge; but part of the problem is also the need for better methods to encode information that is in part structural, in part lexical. Enforcing binding constraints –e.g., ruling out Peter as antecedent of him in (1a) requires recognizing that the anaphor occurs in a particular type of construction (Chomsky 1981; Lappin and Leass 1994; Yang et al. 2006) whose exact definition however has not yet been agreed upon by linguists (indeed, it may only be definable in a graded sense (Sturt 2003; Yang et al. 2006)), witness examples like (1b). Parallelism effects are a good example of structural information inducing preferences rather than constraints. Recognizing that It in examples such as (1c,d) are expletives requires a combination of structural information and lexical information (Lappin and Leass 1994; Evans 2001). But some sort of structure also underlies our interpretation of other types of coreference: e.g., knowledge about the structure of names certainly plays a role in recognizing that BJ Habibie is a possible antecedent for Mr. Habibie. (1) a. John thinks that Peter hates him. b. John hopes that Jane is speaking only to …

Content truncated to 500 words