Japanese Predicate Argument Structure Analysis Exploiting Argument Position and Type

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 201–209, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Japanese Predicate Argument Structure Analysis Exploiting Argument Position and Type Yuta Hayashibe Mamoru Komachi Yuji Matsumoto Graduate School of Information Science Nara Institute of Science and Technology 8916-5, Takayama, Ikoma Nara 630-0192, Japan { yuta-h, komachi, matsu }@is.naist.jp Abstract We propose an approach to Japanese pred- icate argument structure analysis exploit- ing argument position and type. In partic- ular, we propose the following two meth- ods. First, in order to use information in the sentences in preceding context of the predicate more effectively, we pro- pose an improved similarity measure be- tween argument positions which is more robust than a previous co-reference-based measure. Second, we propose a flex- ible selection-and-classification approach which accounts for the minor types of ar- guments. Experimental results show that our proposed method achieves state-of- the-art accuracy for Japanese predicate ar- gument structure analysis. 1 Introduction The goal of predicate-argument structure analy- sis is to extract semantic relations such as “who did what to whom” that hold between a predicate and its arguments constituting a semantic unit of a sentence. It is an important step in many Natu- ral Language Processing applications such as ma- chine translation, summarization and information extraction. Arguments are classified into three categories according to their positions relative to the predi- cates: intra-sentential arguments (those that have direct syntactic dependency with the predicates), zero intra-sentential arguments (those appearing as zero-pronouns but have their antecedents in the same sentence), and inter-sentential arguments (those appearing as zero-pronouns and their an- tecedents are not in the same sentence). We call them INTRA D, INTRA Z, and INTER respec- tively. Furthermore, we call these categories the argument types. While the analysis of INTRA D is comparatively easy, INTRA Z and INTER are more difficult. We consider that there are two rea- sons for this. The first reason is the poverty of features for argument identification compared to INTRA D. While for INTRA D we have important clues such as the function word or directly dependency rela- tion, we don’t for INTRA Z and INTER. The second reason is the limited amount of training examples. For example, in a Japanese newswire corpus, INTRA Z and INTER account for 30.5% and 12.4% of all the nominative (ga) cases, and 13.1% and 0.2% of all of the accusative (wo) cases (Iida et al., 2007). In this paper, in order to solve these problems we propose the following two methods exploiting argument position and type. First, we propose an improved similarity mea- sure between argument positions of two predicates that take semantically similar arguments. For ex- ample, someone possibly arrested can also surren- der him/herself, that is, objects of “arrest” and sub- jects of “surrender (oneself)” are occupied by se- mantically similar nouns. Gerber and Chai (2010) proposed analysis of English nominal predicates with this similarity to take discourse context into account. However, the similarity measure they …

Content truncated to 500 words