[4J3-J-13] AI application: text applications and discourse

[4J3-J-13-01] Dialogue Breakdown Detection using Universal Transformer

〇Kenta Kuwahara¹, Hidefumi Ohmura¹, Kouichi Katsurada¹ (1. Tokyo University of Science)

Keywords:dialogue breakdown, dialogue system, machine learning

In recent years, a lot of dialogue systems have been developed at many institutions. However, most of these systems sometimes cause the problem of dialogue breakdown because of the systems’ inappropriate utterances, which makes it impossible to continue the conversation between a user and the system. In this paper, we propose a dialogue breakdown detection system using a universal transformer. Since our system can model the detailed structure of input sentence, it is expected to predict the dialogue breakdown accurately than the previous systems. To show its performance, we conducted an experiment in which our system is compared with a previous study using an RNN. The experimental result shows that our system can obtain a higher F-measure metrics than the previous method in the task of dialogue breakdown challenge.