Jointly Extracting Japanese Predicate-Argument Relation with Markov Logic

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1125–1133, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Jointly Extracting Japanese Predicate-Argument Relation with Markov Logic Katsumasa Yoshikawa, Masayuki Asahara and Yuji Matsumoto Graduate School of Information Science, Nara Institute of Science and Technology 8916-5, Takayama-cho, Ikoma, Nara 630-0192, Japan {katsumasa-y,masayu-a,matsu}@is.naist.jp Abstract This paper describes a new Markov Logic approach for Japanese Predicate-Argument (PA) relation extraction. Most previous work built separated classifiers correspond- ing to each case role and independently identified the PA relations, neglecting de- pendencies (constraints) between two or more PA relations. We propose a method which collectively extracts PA relations by optimizing all argument candidates in a sentence. Our method can jointly consider dependency between multiple PA relations and find the most probable combination of predicates and their arguments in a sen- tence. In addition, our model involves new constraints to avoid considering inappro- priate candidates for arguments and iden- tify correct PA relations effectively. Com- pared to the state-of-the-art, our method achieves competitive results without large- scale data. 1 Introduction Predicate-argument (PA) relation extraction is one of the challenging problems in Natural Lan- guage Processing. The analysis extracts seman- tic information such as “who did what to whom”, which is often useful to various applications like information extraction, document summarization, and machine translation. Predicate-argument relation extraction is of- ten called semantic role labeling. In English, it has been researched on large corpora such as FrameNet (Fillmore et al., 2001) and Prop- Bank (Palmer et al., 2005). CoNLL Shared Task 2008 (Surdeanu et al., 2008) is a representative work of semantic role labeling based on these cor- pora. Japanese PA relation extraction is a kind of semantic role labeling but an argument is often called case. A typical example of Japanese PA re- Figure 1: Example of Predicate-Argument Struc- ture lation is shown in Figure 1. In this example, “�� � (went)” is a predicate and there are two argu- ments for the predicate, that is, a nominative case role (ga) is “� (He)” and a dative case role (ni) is “��� (library)”. In Japanese, Taira et al. (2008) and Imamura et al. (2009) tackled PA relation extraction on NAIST Text Corpus (Iida et al., 2007). They created three separated models corresponding to each of the case; ga (Nominative), wo (Accusative), and ni (Dative). Even though some English semantic role labeler apply global models, most of them solve problems on a per-predicate basis (Toutanova et al., 2008; Watanabe et al., 2010). In this work, we propose an approach to Japanese PA relation extraction on a per-sentence basis and utilize important depen- dencies between one PA relation and another in the same sentence. In order to use such dependencies as global constraints, we apply a Markov Logic ap- proach to Japanese PA relation extraction. In recent years, in English semantic role labeling, a Markov Logic model has achieved one of the state-of-the- art results (Meza-Ruiz and Riedel, 2009a). With global constraints between multiple …

Content truncated to 500 words