Multi-Task Learning for Japanese Predicate Argument Structure Analysis

Proceedings of NAACL-HLT 2019, pages 3404–3414 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3404 Multi-task Learning for Japanese Predicate Argument Structure Analysis Hikaru Omori Tokyo Metropolitan University omori-hikaru@ed.tmu.ac.jp Mamoru Komachi Tokyo Metropolitan University komachi@tmu.ac.jp Abstract An event-noun is a noun that has an argu- ment structure similar to a predicate. Re- cent works, including those considered state- of-the-art, ignore event-nouns or build a single model for solving both Japanese predicate ar- gument structure analysis (PASA) and event- noun argument structure analysis (ENASA). However, because there are interactions be- tween predicates and event-nouns, it is not suf- ficient to target only predicates. To address this problem, we present a multi-task learn- ing method for PASA and ENASA. Our multi- task models improved the performance of both tasks compared to a single-task model by shar- ing knowledge from each task. Moreover, in PASA, our models achieved state-of-the-art re- sults in overall F1 scores on the NAIST Text Corpus. In addition, this is the first work to employ neural networks in ENASA. 1 Introduction Japanese predicate argument structure analysis (PASA) examines semantic structures between the predicate and its arguments in a text. The identifi- cation of the argument structure such as “who did what to whom?” is useful for natural language pro- cessing that requires deep analysis of complicated sentences such as machine translation and recog- nizing textual entailment. PASA is a task targeted at predicates such as verbs and adjectives. How- ever, there are also many nouns that have event- related arguments in a sentence. We call these nouns that refer to events event-nouns, for exam- ple, a verbal noun (sahen nouns) such as houkoku “report” or a deverbal noun (nominalized forms of verbs) such as sukui “rescue.” Figure 1 shows examples of PASA and event- noun argument structure analysis (ENASA). In the NAIST Text Corpus (Iida et al., 2007), both predicates and event-nouns have one of three core (a) He reports the result to his boss. (b) His progress report was too short; hence, he got scolded by his boss. (c) He sent a brief note to his boss. Figure 1: Examples of PASA and ENASA. The edges denote dependency paths. case roles, nominative (NOM), accusative (ACC), and dative (DAT) as an argument. According to Iida et al. (2007), predicates have almost no ar- gument in the same bunsetsu1 phrase. However, in the case of event-nouns, approximately half of the accusative and dative arguments appear in the same bunsetsu phrase. Accordingly, although PASA and ENASA are semantically highly re- lated, they are syntactically different tasks. How- ever, most previous studies focused on predi- cates only; hence, there are few studies that focus 1Functional chunk in Japanese. It consists of one or more content words (noun, verb, adjective, etc.) followed by zero or more function words (postposition, auxiliary verb, etc.). A verb phrase in Japanese thus cannot bear noun arguments in the same bunsetsu. 3405 on event-nouns (Komachi et al., 2007; Taira et al., 2008). To …

Content truncated to 500 words