Effectiveness of Domain Adaptation in Japanese Predicate-Argument Structure Analysis

Effectiveness of Domain Adaptation in Japanese Predicate-Argument Structure Analysis Mizuki Sango, Hitoshi Nishikawa, Takenobu Tokunaga Tokyo Institute of Technology / W8-73, 2-12-1 Ookayama, Meguro, Tokyo 152-8552, Japan {sango.m.ab@m, hitoshi@c, take@c}.titech.ac.jp Abstract This paper proposes introducing domain adap- tation into Japanese predicate-argument struc- ture (PAS) analysis. Our investigation of a Japanese balanced-corpus revealed that the distribution of argument types is different across text media, particularly the difference is significant when the argument is exophoric. The past Japanese PAS analysis research has disregarded this tendency. We start with an RNN-based PAS analyzer as a baseline, extending it by introducing three kinds of domain-adaptation techniques and their com- binations. The evaluation experiments using a Japanese balanced-corpus (BCCWJ) con- firmed that the domain-adaptation is effective for improving the performance of the Japanese PAS analysis. 1 Introduction Predicate-argument structure (PAS) analysis is the task to identify the argument for each case of the target predicate. As it is a fundamental analysis for various natural language processing (NLP) applica- tions, the PAS analysis has been one of the most active research areas in NLP. In discourse-oriented languages like Japanese, the target language of the present study, arguments are often omitted from the sentence Kayama (2003). Those omitted arguments are considered as zero-pronouns or exophora. (1) meiru-o kaite okuttayo. yondene. mailACC wrotev1 sentv2 readv3/imperative I wrote a mail to you and sent it to you. Read it. For instance, example (1) has three predicates (v1, v2 and v3) and one explicit argument (mail). predicate\ case NOM ACC DAT v1: wrote [writer] mail [reader] v2: sent [writer] (mail) [reader] v3: read [reader] ((mail)) none Table 1: PAS analysis result for example (1) The PAS analysis result of example (1) looks like Table 1, where the elements enclosed with square brackets are exophoric, the round bracketed is an intra-sentential zero-anaphora and the double round bracketed is an inter-sentential zero-anaphora. The accusative argument of v1, “meiru-o (mail)”, is ex- plicitly marked by the case marker “o” and has a dependency relation with v1, which is indicated by a bare noun, i.e. without any bracket. Although the Japanese PAS analysis is similar to the semantic role labeling (SRL) (Zhou and Xu, 2015; He et al., 2017), it also involves anaphora res- olution for zero-pronouns and exophora to identify the argument for every case of the predicate, which corresponds to filling the bracketed elements in Ta- ble 1. We also find omitted arguments in other pro-drop languages such as Chinese, Turkish, and some null-subject languages in the Romance lan- guages (Iida and Poesio, 2011; Rello et al., 2012; Chen and Ng, 2016; Yin et al., 2017). The past Japanese PAS analysis utilizes vari- ous features obtained from the morphological and syntactic analysis (Matsubayashi and Inui, 2017; Hayashibe et al., 2011; Imamura et al., 2014; Shi- bata et al., 2016; Ouchi et al., 2015; Yoshikawa et al., 2013; Taira et al., 2008). The recent approach in- cludes the end-to-end approach that does not require any intermediate analysis (Ouchi et al., 2017). PACLIC 32 …

Content truncated to 500 words