Japanese Semantic Role Labeling with Hierarchical Tag Context Trees

In this paper we describe that the hierarchical tag context tree (HTCT) approach improves the accuracy of semantic role labeling on Japanese text. In Japanese language there are functional multiword expressions such as no-tame-ni and yotte that have potential to designate semantic relations between a predicate and its arguments. Since these expressions come to the end part of each argument, the performance of the CRF-based semantic role labeler can be improved by taking into account the last morphemes of each argument as features. We apply our proposed system to the annotated corpus of semantic role labels on a balanced Japanese corpus. The experimental results show that the CRFbased labeler with features extracted by HTCT approach outperforms the normal CRF-based labeler.