Coreference Resolution Using Competition Learning Approach

Coreference Resolution Using Competition Learning Approach Xiaofeng Yang*+ Guodong Zhou* Jian Su* Chew Lim Tan + *Institute for Infocomm Research, 21 Heng Mui Keng Terrace, Singapore 119613 +Department of Computer Science, National University of Singapore, Singapore 117543 *{xiaofengy,zhougd,sujian}@ i2r.a-star.edu.sg +(yangxiao,tancl)@comp.nus.edu.sg Abstract In this paper we propose a competition learning approach to coreference resolu- tion. Traditionally, supervised machine learning approaches adopt the single- candidate model. Nevertheless the prefer- ence relationship between the antecedent candidates cannot be determined accu- rately in this model. By contrast, our ap- proach adopts a twin-candidate learning model. Such a model can present the competition criterion for antecedent can- didates reliably, and ensure that the most preferred candidate is selected. Further- more, our approach applies a candidate filter to reduce the computational cost and data noises during training and resolution. The experimental results on MUC-6 and MUC-7 data set show that our approach can outperform those based on the single- candidate model. 1 Introduction Coreference resolution is the process of linking together multiple expressions of a given entity. The key to solve this problem is to determine the ante- cedent for each referring expression in a document. In coreference resolution, it is common that two or more candidates compete to be the antecedent of an anaphor (Mitkov, 1999). Whether a candidate is coreferential to an anaphor is often determined by the competition among all the candidates. So far, various algorithms have been proposed to deter- mine the preference relationship between two can- didates. Mitkov’s knowledge-poor pronoun resolution method (Mitkov, 1998), for example, uses the scores from a set of antecedent indicators to rank the candidates. And centering algorithms (Brennan et al., 1987; Strube, 1998; Tetreault, 2001), sort the antecedent candidates based on the ranking of the forward-looking or backward- looking centers. In recent years, supervised machine learning approaches have been widely used in coreference resolution (Aone and Bennett, 1995; McCarthy, 1996; Soon et al., 2001; Ng and Cardie, 2002a), and have achieved significant success. Normally, these approaches adopt a single-candidate model in which the classifier judges whether an antecedent candidate is coreferential to an anaphor with a con- fidence value. The confidence values are generally used as the competition criterion for the antecedent candidates. For example, the “Best-First” selection algorithms (Aone and Bennett, 1995; Ng and Cardie, 2002a) link the anaphor to the candidate with the maximal confidence value (above 0.5). One problem of the single-candidate model, however, is that it only takes into account the rela- tionships between an anaphor and one individual candidate at a time, and overlooks the preference relationship between candidates. Consequently, the confidence values cannot accurately represent the true competition criterion for the candidates. In this paper, we present a competition learning approach to coreference resolution. Motivated by the research work by Connolly et al. (1997), our approach adopts a twin-candidate model to directly learn the competition criterion for the antecedent candidates. In such a model, a classifier is trained based on the instances formed by an anaphor and a pair of …

Content truncated to 500 words