Semantic Role Labeling for Japanese Using Span-Based Models

We propose a span-based model for Japanese Semantic Role Labeling (SRL) with deep neural networks. Most previous studies of semantic role labeling have been conducted in English, however, it is not clear that previously proposed models are effective in Japanese. Besides, most of the previous studies relating to Japanese SRL focus on the dependency-based SRL task. Thus, in this paper, we apply span-based models to a Japanese corpus NPCMJ-PT which is annotated with semantic role labels and has about 52,500 entries. The semantic roles are defined with 32 types of semantic role labels such as Arg0, Arg1, and ArgM-LOC that are similar to PropBank. In the span-based models, there are two primary modeling approaches. One is a model that estimates a semantic role label for each span, and the other is a model that estimates a span for each label. Thus, we apply the two types of span-based models to a Japanese SRL task. The experimental results demonstrate that the model, which estimates a span for each semantic role label, achieved the highest F1 score of 77.54, in comparison to the other model.