Improving Japanese semantic-role-labeling performance with transfer learning as case for limited resources of tagged corpora on aggregated language

Improving Japanese semantic-role-labeling performance with transfer learning as case for limited resources of tagged corpora on aggregated language Takuya Okamura∗, Koichi Takeuchi∗, Yasuhiro Ishihara∗, Masahiro Taguchi†, Yoshihiko Inada‡, Masaya Iizuka§, Tatsuhiko Abo¶, Hitoshi Ueda¶ ∗Graduate School of Natural Science and Technology, Okayama University, Okayama, Japan t-okamura@s.okayama-u.ac.jp, koichi@cl.cs.okayama-u.ac.jp †Graduate School of Humanities and Social Science, Okayama University, Okayama, Japan ‡Graduate School of Education, Okayama University, Okayama, Japan §Institute for Education and Student Services, Okayama University, Okayama, Japan ¶Graduate School of Natural Science and Technology, Okayama University, Okayama, Japan Abstract In this paper we proposed the use of effective features and transfer leaning to improve the accuracies of neural-network-based models for accurate semantic role labeling (SRL) of Japanese, which is an aggregated language. We first reveal that the final morphemes in each argument, which have not been dis- cussed in previous work on English SRL are effective features in determining semantic role labels in Japanese. We then discuss the possibility of using large-scale training cor- pora annotated with different semantic labels from the target semantic labels by transfer learning on CNN, 3-LNN, and GRU models. The experimental results of Japanese SRL on the proposed models indicate that all of the neural-network-based models performed better with transfer learning as well as using the feature vectors of final moprhemes in each argument. 1 Introduction Several studies on semantic role labeling (SRL) have been conducted, mainly for English texts with a wide coverage from revealing syntactic and grammatical features that impact SRL deci- sions (Gildea and Jurafsky, 2002) to end-to-end models without syntactic inputs. This is because current rich language resources are related to anno- tated corpora with semantic roles such as PropBank (Kingsbury et al., 2002), FrameNet (Baker et al., 1998), and shared tasks of CoNLL 2005 (Carreras and M`arquez, 2005), 2009 (Hajiˇc et al., 2009), and 2012 (Pradhan et al., 2012). Instead of English SRL, most studies (Taira et al., 2008; Imamura et al., 2009; Sasano and Kurohashi, 2011; Hayashibe et al., 2011; Ouchi et al., 2015; Shibata et al., 2016; Ouchi et al., 2017; Matsubayashi and Inui, 2018) on SRL for Japanese texts, the grammar of which is quite different from English, have been focused on detecting three types of case-marker-based labels, i.e., nominative, accusative, and dative. This is because large-scale corpora annotated with these three labels have been developed (Kyoto Corpus (Kawahara et al., 2002) and NAIST Text Corpus (Iida et al., 2007)) and widely used1. There are, however, annotated cor- pora with semantic tags corresponding to semantic role labels (EDR corpus2, BCCWJ-PT (Takeuchi et al., 2015) and Japanese FrameNet (Ohara et al., 2011)) or semantic categories (GDA corpus3) that contain semantic relations as cause and location. Thus, by constructing a Japanese SRL system us- ing these resources, we can discuss the effective approaches and models for a situation without standard semantic-role-labeled corpora as well as effective features for Japanese SRL. We address the following two issues: what is an effective grammatical feature for Japanese SRL, and what is a model …

Content truncated to 500 words