Joint Case Argument Identification for Japanese Predicate Argument Structure Analysis

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 961–970, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Joint Case Argument Identification for Japanese Predicate Argument Structure Analysis Hiroki Ouchi Hiroyuki Shindo Kevin Duh Yuji Matsumoto Graduate School of Information and Science Nara Institute of Science and Technology 8916-5, Takayama, Ikoma, Nara, 630-0192, Japan { ouchi.hiroki.nt6, shindo, kevinduh, matsu }@is.naist.jp Abstract Existing methods for Japanese predicate argument structure (PAS) analysis identify case arguments of each predicate without considering interactions between the tar- get PAS and others in a sentence. How- ever, the argument structures of the pred- icates in a sentence are semantically re- lated to each other. This paper proposes new methods for Japanese PAS analysis to jointly identify case arguments of all predicates in a sentence by (1) modeling multiple PAS interactions with a bipar- tite graph and (2) approximately search- ing optimal PAS combinations. Perform- ing experiments on the NAIST Text Cor- pus, we demonstrate that our joint analysis methods substantially outperform a strong baseline and are comparable to previous work. 1 Introduction Predicate argument structure (PAS) analysis is a shallow semantic parsing task that identifies ba- sic semantic units of a sentence, such as who does what to whom, which is similar to semantic role labeling (SRL)1. In Japanese PAS analysis, one of the most prob- lematic issues is that arguments are often omitted in the surface form, resulting in so-called zero- pronouns. Consider the sentence of Figure 1. 1We use “PAS analysis” in this paper following previous work on Japanese PAS analysis. Figure 1: An example of Japanese PAS. The En- glish translation is “Because ϕi caught a cold, Ii skipped school.”. The upper edges are dependency relations, and the under edges are case arguments. “NOM” and “ACC” represents the nominative and accusative arguments, respectively. “ϕi” is a zero- pronoun, referring to the antecedent “watashii”. The case role label “NOM” and “ACC” respec- tively represents the nominative and accusative roles, and ϕi represents a zero-pronoun. There are two predicates “hiita (caught)” and “yasunda (skipped)”. For the predicate “yasunda (skipped)”, “watashii-wa (Ii)” is the “skipper”, and “gakko-wo (school)” is the “entity skipped”. It is easy to iden- tify these arguments, since syntactic dependency between an argument and its predicate is a strong clue. On the other hand, the nominative argument of the predicate “hiita (caught)” is “watashii-wa (Ii)”, and this identification is more difficult be- cause of the lack of the direct syntactic depen- dency with “hiita (caught)”. The original nomina- tive argument appears as a zero-pronoun, so that we have to explore the antecedent, an element re- ferred to by a zero-pronoun, as the argument. As the example sentence shows, we cannot use ef- fective syntactic information for identifying such arguments. This type of arguments is known as implicit arguments, a very problematic language 961 phenomenon for PAS analysis (Gerber and Chai, 2010; Laparra and Rigau, 2013). Previous work …

Content truncated to 500 words