A Fully-Lexicalized Probabilistic Model for Japanese Zero Anaphora Resolution

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 769–776 Manchester, August 2008 A Fully-Lexicalized Probabilistic Model for Japanese Zero Anaphora Resolution Ryohei Sasano∗ Graduate School of Information Science and Technology, University of Tokyo ryohei@nlp.kuee.kyoto-u.ac.jp Daisuke Kawahara National Institute of Information and Communication Technology dk@nict.go.jp Sadao Kurohashi Graduate School of Infomatics, Kyoto University kuro@i.kyoto-u.ac.jp Abstract This paper presents a probabilistic model for Japanese zero anaphora resolution. First, this model recognizes discourse en- tities and links all mentions to them. Zero pronouns are then detected by case struc- ture analysis based on automatically con- structed case frames. Their appropriate antecedents are selected from the entities with high salience scores, based on the case frames and several preferences on the relation between a zero pronoun and an antecedent. Case structure and zero anaphora relation are simultaneously de- termined based on probabilistic evaluation metrics. 1 Introduction Anaphora resolution is one of the most important techniques in discourse analysis. In English, def- inite noun phrases such as the company and overt pronouns such as he are anaphors that refer to pre- ceding entities (antecedents). On the other hand, in Japanese, anaphors are often omitted and these omissions are called zero pronouns. We focus on zero anaphora resolution of Japanese web cor- pus, in which anaphors are often omitted and zero anaphora resolution plays an important role in dis- course analysis. Zero anaphora resolution can be divided into two phases. The first phase is zero pronoun detec- tion and the second phase is zero pronoun resolu- tion. Zero pronoun resolution is similar to coref- 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. * Research Fellow of the Japan Society for the Promotion of Science (JSPS) erence resolution and pronoun resolution, which have been studied for many years (e.g. Soon et al. (2001); Mitkov (2002); Ng (2005)). Isozaki and Hirao (2003) and Iida et al. (2006) focused on zero pronoun resolution assuming perfect pre-detection of zero pronouns. However, we consider that zero pronoun detection and resolution have a tight rela- tion and should not be handled independently. Our proposed model aims not only to resolve zero pro- nouns but to detect zero pronouns. Zero pronouns are not expressed in a text and have to be detected prior to identifying their an- tecedents. Seki et al. (2002) proposed a proba- bilistic model for zero pronoun detection and res- olution that uses hand-crafted case frames. In order to alleviate the sparseness of hand-crafted case frames, Kawahara and Kurohashi (2004) in- troduced wide-coverage case frames to zero pro- noun detection that are automatically constructed from a large corpus. They use the case frames as selectional restriction for zero pronoun resolution, but do not utilize the frequency of each example of case slots. However, since the frequency is shown to be a good clue for syntactic and case structure analysis (Kawahara and Kurohashi, 2006), we con- sider the frequency also can benefit zero pronoun detection. Therefore we propose …

Content truncated to 500 words