An Unsupervised Baseline For Dialogue Breakdown Detection Using Ouf-of-distribution Detection

For the last several decades, a focus of artificial intelligence work has been computer-based systems for interacting with humans. Such a system must ensure smooth communication to provide thoughtful answers by interacting with the system. A major critical issue is they can create an unwanted statement that results in a breakdown of dialogue, which degrades the overall interaction quality. Thus, it is incredibly beneficial to employ a system to detect whether a breakdown hampers the conversation flow. In this work, we propose an unsupervised dialogue breakdown detection technique that can be trained using only dialogue data and no other supervision. Our experiments show that using out-of-distribution detection methods can perform on par with competing supervised methods that use labeled English data. On the DBDC4 benchmark, it is the only unsupervised approach to our knowledge.