Dialogue Breakdown Detection Based on Nonlinguistic Acoustic Information
Motoki Abe, Takashi Tsunakawa, Masafumi Nishida, Masafumi Nishimura · 2018
Chat dialogue systems suffer phenomena called “dialogue breakdown” when the system fails to control the dialogue properly and responds to the user with an inappropriate return. Dialogue breakdown is obstacle to realizing natural dialogue between the system and user, so it is necessary for systems to detect it quickly and recover. The previous research mainly focused on text dialogue. In this research, we focused on spoken dialogue and tried to detect dialogue breakdown by using acoustic information not included in text dialogue. We compared its performance with a detector using only linguistic information. As a result, the detector using acoustic information performed as well or better than the one using linguistic information. We found that the dialogue breakdown detector using linguistic information could not behave well in the case of redundant utterances, which is a unique feature of spoken dialogue.