Coreference or Not: A Twin Model for Coreference Resolution

Proceedings of NAACL HLT 2007, pages 73–80, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Coreference or Not: A Twin Model for Coreference Resolution Xiaoqiang Luo IBM T.J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598, U.S.A. {xiaoluo}@us.ibm.com Abstract A twin-model is proposed for coreference res- olution: a link component, modeling the coref- erential relationship between an anaphor and a candidate antecedent, and a creation com- ponent modeling the possibility that a phrase is not coreferential with any candidate an- tecedent. The creation model depends on all candidate antecedents and is often expensive to compute; Therefore constraints are imposed on feature forms so that features in the cre- ation model can be efficiently computed from feature values in the link model. The pro- posed twin-model is tested on the data from the 2005 Automatic Content Extraction (ACE) task and the proposed model performs bet- ter than a thresholding baseline without tuning free parameter. 1 Introduction Coreference resolution aims to find multiple mentions of an entity (e.g., PERSON, ORGANIZATION) in a document. In a typical machine learning-based coref- erence resolution system (Soon et al., 2001; Ng and Cardie, 2002b; Yang et al., 2003; Luo et al., 2004), a statistical model is learned from training data and is used to measure how likely an anaphor 1 is corefer- ential to a candidate antecedent. A related, but often overlooked, problem is that the anaphor may be non- coreferential to any candidate, which arises from sce- narios such as an identified anaphor is truly generic and 1In this paper, “anaphor” includes all kinds of phrases to be resolved, which can be named, nominal or pronominal phrases. there does not exist an antecedent in the discourse con- text, or an anaphor is the first mention (relative to pro- cessing order) in a coreference chain. In (Soon et al., 2001; Ng and Cardie, 2002b), the problem is treated by thresholding the scores re- turned by the coreference model. That is, if the max- imum coreference score is below a threshold, then the anaphor is deemed non-referential to any candidate an- tecedent. The threshold approach does not model non- coreferential events directly, and is by no means the op- timal approach to the problem. It also introduces a free parameter which has to be set by trial-and-error. As an improvement, Ng and Cardie (2002a) and Ng (2004) train a separate model to classify an anaphor as either anaphoric or non-anaphoric. The output of this clas- sifier can be used either as a pre-filter (Ng and Cardie, 2002a) so that non-anaphoric anaphors will not be pre- cessed in the coreference system, or as a set of features in the coreference model (Ng, 2004). By rejecting any anaphor classified as non-anaphoric in coreference res- olution, the filtering approach is meant to handle non- anaphoric phrases (i.e., no antecedent exists in the dis- course under consideration), not the first mention in a coreference chain. In this paper, coreference is viewed as a process of sequential operations on …

Content truncated to 500 words