Antecedent Selection Techniques for High-Recall Coreference Resolution

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 496–505, Prague, June 2007. c⃝2007 Association for Computational Linguistics Antecedent Selection Techniques for High-Recall Coreference Resolution Yannick Versley versley@sfs.uni-tuebingen.de SFB 441 / Seminar f¨ur Sprachwissenschaft Universit¨at T¨ubingen Abstract We investigate methods to improve the re- call in coreference resolution by also trying to resolve those definite descriptions where no earlier mention of the referent shares the same lexical head (coreferent bridging). The problem, which is notably harder than iden- tifying coreference relations among men- tions which have the same lexical head, has been tackled with several rather different ap- proaches, and we attempt to provide a mean- ingful classification along with a quantita- tive comparison. Based on the different mer- its of the methods, we discuss possibilities to improve them and show how they can be ef- fectively combined. 1 Introduction Coreference resolution, the task of grouping men- tions in a text that refer to the same referent in the real world, has been shown to be beneficial for a number of higher-level tasks such as information ex- traction (McCarthy and Lehnert, 1995), question an- swering (Morton, 2000) and summarisation (Stein- berger et al., 2005). While the resolution of pronominal anaphora and tracking of named entities is possible with good accuracy, the resolution of definite NPs (having a common noun as their head) is usually limited to the cases that Vieira and Poesio (2000) call direct coreference, where both coreferent mentions have the same head. The other cases, called coreferent bridging by Vieira and Poesio1, are notably harder because the number of potential candidates is much 1Because bridging (in the sense of Clark, 1975, or Asher and larger when it is no longer possible to rely on surface similarity. To overcome the limit of recall that is encoun- tered when only relying on surface features, newer systems for coreference resolutions (Daum´e III and Marcu, 2005; Ponzetto and Strube, 2006; Versley, 2006; Ng, 2007, inter alia) use lexical semantic in- formation as an indication for semantic compati- bility in the absence of head equality. Most cur- rent systems integrate the identification of discourse- new definites (i.e., cases like “the sun” or “the man that Ben met yesterday”, which are definite, but not anaphoric) with the antecedent selection proper, which implies that the gain obtained for new features is dependent on the feature’s usefulness both in find- ing semantically related mentions and for the use in detecting discourse-new definites. One goal of this paper is to provide a better under- standing of these information sources by comparing proposed (and partly new) approaches for resolv- ing coreferent bridging by separately considering the task of antecedent selection (i.e., presupposing that discourse-new markables have been identified beforehand). Although state of the art methods for modular discourse-new detection (Uryupina, 2003; Poesio et al., 2005) do not achieve near-perfect accu- racy for discourse-new detection, the results we give for antecedent selection represent an upper bound on recall and …

Content truncated to 500 words