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

In this paper we proposed the use of effectivefeatures and transfer leaning to improve theaccuracies of neural-network-based modelsfor accurate semantic role labeling (SRL) ofJapanese, which is an aggregated language.We first reveal that the final morphemes ineach argument, which have not been discussed in previous work on English SRLare effective features in determining semanticrole labels in Japanese. We then discuss thepossibility of using large-scale training corpora annotated with different semantic labelsfrom the target semantic labels by transferlearning on CNN, 3-LNN, and GRU models.The experimental results of Japanese SRLon the proposed models indicate that all ofthe neural-network-based models performedbetter with transfer learning as well as usingthe feature vectors of final moprhemes ineach argument.