Neural Network-Based Model for Japanese Predicate Argument Structure Analysis

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1235–1244, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Neural Network-Based Model for Japanese Predicate Argument Structure Analysis Tomohide Shibata and Daisuke Kawahara and Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {shibata, dk, kuro}@i.kyoto-u.ac.jp Abstract This paper presents a novel model for Japanese predicate argument structure (PAS) analysis based on a neural network framework. Japanese PAS analysis is chal- lenging due to the tangled characteristics of the Japanese language, such as case dis- appearance and argument omission. To unravel this problem, we learn selectional preferences from a large raw corpus, and incorporate them into a SOTA PAS anal- ysis model, which considers the consis- tency of all PASs in a given sentence. We demonstrate that the proposed PAS anal- ysis model significantly outperforms the base SOTA system. 1 Introduction Research on predicate argument structure (PAS) analysis has been conducted actively these days. The improvement of PAS analysis would benefit many natural language processing (NLP) applica- tions, such as information extraction, summariza- tion, and machine translation. The target of this work is Japanese PAS analy- sis. The Japanese language has the following char- acteristics: • head final, • free word order (among arguments), and • postpositions function as (surface) case markers. Japanese major surface cases are が (ga), を (wo), and に (ni), which correspond to Japanese post- positions (case markers). We call them nomina- tive case, accusative case, and dative case, respec- tively. In this paper, we limit our target cases to these three cases. Note that though they are sur- face cases, they roughly correspond to Arg1, Arg2, and Arg3 of English semantic role labeling based on PropBank. Japanese PAS analysis has been considered as one of the most difficult basic NLP tasks, due to the following two phenomena. Case disappearance When a topic marker は (wa) is used or a noun is modified by a relative clause, their case markings disappear as in the fol- lowing examples.1 (1) a. ジョンは John-TOP パンを bread-ACC 食べた。 ate → ジョンが John-NOM (John ate bread.) b. パンは bread-TOP ジョンが John-NOM 食べた。 ate → パンを bread-ACC (John ate bread.) (2) a. パンを 食べた ジョン を ... → bread-ACC ate John-ACC ジョンが (食べた) John-NOM (ate) (John, who ate bread, ...) b. ジョンが John-NOM 食べた パン が ...→ ate bread-NOM パンを (食べた) (ate) bread-ACC (Bread, which John ate, ...) In the example sentences (1a) and (1b), since a topic marker は is used, the NOM and ACC case markers disappear. In the example sentences (2a) and (2b), since a noun is modified by a relative clause, the NOM case of “ジョン” (John) for “食 べた” (eat) and ACC case of “パン” (bread) for “食べた” disappear. Argument omission Arguments are very often omitted in Japanese sentences. This phenomenon is totally different from English sentences, where the word order is fixed and pronouns are used con- 1In this paper, we use the following abbreviations: NOM (nominative), …

Content truncated to 500 words