A Discriminative Approach to Japanese Zero Anaphora Resolution with Large-scale Lexicalized Case Frames

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 758–766, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP A Discriminative Approach to Japanese Zero Anaphora Resolution with Large-scale Lexicalized Case Frames Ryohei Sasano Precision and Intelligence Laboratory Tokyo Institute of Technology sasano@pi.titech.ac.jp Sadao Kurohashi Graduate School of Informatics Kyoto University kuro@i.kyoto-u.ac.jp Abstract We present a discriminative model for Japanese zero anaphora resolution that si- multaneously determines an appropriate case frame for a given predicate and its predicate-argument structure. Our model is based on a log linear framework, and exploits lexical features obtained from a large raw corpus, as well as non-lexical features obtained from a relatively small annotated corpus. We report the results of zero anaphora resolution on Web text and demonstrate the effectiveness of our approach. In addition, we also investigate the relative importance of each feature for resolving zero anaphora in Web text. 1 Introduction Zero anaphora resolution is the task of detect- ing and identifying the omitted arguments of a predicate. Since arguments are often omitted in Japanese, zero anaphora resolution plays an im- portant role in the analysis of Japanese predicate- argument structures. For example, in the follow- ing text: (i) Musuko-wa itazura-ga son-TOP mischief-NOMsukide like watashi-mo (φ-ni) te-wo yaiteiru. I φ-DAT hands-ACC burn (have difficulty) (My son likes mischief, so I have difficulty with φ.) the dative argument of the predicate ‘yaku1 (burn)’ has been omitted. The omitted element is called a zero pronoun, and in this example it refers to ‘musuko (son).’ Although most previous work has focused on zero anaphora in newspaper articles, 1‘Yaku’ is the original form of ‘yaiteiru.’ this paper aims to resolve zero anaphora in Web text, since this involves a wider range of writing styles and is thus considered to be a more practical setting. Reference resolution systems generally require a variety of information sources ranging from syntactic and discourse preferences to semantic preferences (Ng and Cardie, 2002; Haghighi and Klein, 2010). Since syntactic and discourse pref- erences are not word-specific, they can be learned from a relatively small annotated corpus. Seman- tic preferences, on the other hand, represent highly lexicalized knowledge, and hence it is difficult to learn these from a small annotated corpus. In some cases, knowledge of the relations between a predi- cate and its particular argument is insufficient, par- ticularly for zero anaphora resolution. For exam- ple, although the dative argument of the predicate ‘yaku (bake/burn)’ is generally filled by a disk, such as a CD or DVD, it is often filled by a per- son, such as ‘musuko (son),’ if the accusative ar- gument is filled by ‘te (hands)’ as in the example (i).2 Thus, knowledge of relations among a predi- cate and its multiple arguments is required to take such preferences into account. Sasano et al. (2008) exploited large-scale case frames that were automatically constructed from 1.6 billion Web sentences as such a lexical re- source, and proposed a probabilistic model for Japanese zero anaphora resolution. Their model …

Content truncated to 500 words