Improving the performance of chat-oriented dialogue systems via dialogue breakdown detection

Abstract

Dialogue breakdown detection is a technique used for identifying inappropriate utterances in dialogue systems that has attracted increased attention, especially in chat-oriented dialogue systems. Although it is generally assumed that dialogue breakdown detection avoids generating system responses that then cause difficulties in continuing the given dialogue, this has yet to be verified experimentally or theoretically. In this paper, we apply the dialogue breakdown detection technique to generate responses for a chat-oriented dialogue system and experimentally verify that performance is improved by measuring the extent to which dialogue breakdown is avoided. Our experimental results show that dialogue breakdown detection indeed is able to improve the appropriateness of system responses; however, short, simple, and dull responses tend to increase when using this technique.