A Twin-Candidate Model of Coreference Resolution with Non-Anaphor Identification Capability

A Twin-Candidate Model of Coreference Resolution with Non-Anaphor Identification Capability Xiaofeng Yang1,2, Jian Su1, and Chew Lim Tan2 1 Institute for Infocomm Research, 21, Heng Mui Keng Terrace, Singapore, 119613 {xiaofengy, sujian}@i2r.a-star.edu.sg 2 Department of Computer Science, National University of Singapore, Singapore, 117543 {yangxiao, tancl}@comp.nus.edu.sg Abstract. Although effective for antecedent determination, the tradi- tional twin-candidate model can not prevent the invalid resolution of non-anaphors without additional measures. In this paper we propose a modified learning framework for the twin-candidate model. In the new framework, we make use of non-anaphors to create a special class of training instances, which leads to a classifier capable of identifying the cases of non-anaphors during resolution. In this way, the twin-candidate model itself could avoid the resolution of non-anaphors, and thus could be directly deployed to coreference resolution. The evaluation done on newswire domain shows that the twin-candidate based system with our modified framework achieves better and more reliable performance than those with other solutions. 1 Introduction In recent years supervised learning approaches have been widely used in corefer- ence resolution task and achieved considerable success [1,2,3,4,5]. Most of these approaches adopt the single-candidate learning model, in which coreference rela- tion is determined between a possible anaphor and one individual candidate at a time [1,3,4]. However, it has been claimed that the reference between an anaphor and its candidate is often subject to the other competing candidates [5]. Such information is nevertheless difficult to be captured in the single-candidate model. As an alternative, several researchers proposed a twin-candidate model [2,5,6]. Instead of directly determining coreference relations, this model would judge the preference between candidates and then select the most preferred one as the an- tecedent. The previous work has reported that such a model can effectively help antecedent determination for anaphors [5,6]. However, one problem exits with the twin-candidate model. For every encoun- tered NP during resolution, the model would always pick out a “best” candidate as the antecedent, even if the current NP is not an anaphor. The twin-candidate R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 719–730, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 720 X. Yang, J. Su, and C.L. Tan model itself could not identify and block such invalid resolution of non-anaphors. Therefore, to apply such a model to coreference resolution, some additional ef- forts have to be required, e.g., using an anaphoricity determination module to eliminate non-anaphors in advance [5], or using threshold to prevent the selection of a candidate if the confidence it wins other competitors is low [6]. In this paper, we explore how to effectively apply the twin-candidate model to the coreference resolution task. We propose a modified learning framework with the capability of processing non-anaphors. In the framework, we make use of non-anaphors to create training instances. This special class of instances would enable the learned classifier to identify the test instances formed by non-anaphors during resolution. Thus, the resulting model could avoid resolving a non-anaphor to a non-existent antecedent by itself, without specifying …

Content truncated to 500 words