Inconsistent dialogue responses and how to recover from them
This work addresses the issue of maintaining consistency in preferences, opinions, beliefs, and facts for chat systems, which is an incremental improvement in dialogue AI.
The paper tackles the problem of inconsistent dialogue responses in chat systems by developing a dataset covering inconsistency introduction, understanding, and resolution, and finds that this dataset aids in detection and resolution tasks, with models like ChatGPT performing well in resolution but struggling with detection.
One critical issue for chat systems is to stay consistent about preferences, opinions, beliefs and facts of itself, which has been shown a difficult problem. In this work, we study methods to assess and bolster utterance consistency of chat systems. A dataset is first developed for studying the inconsistencies, where inconsistent dialogue responses, explanations of the inconsistencies, and recovery utterances are authored by annotators. This covers the life span of inconsistencies, namely introduction, understanding, and resolution. Building on this, we introduce a set of tasks centered on dialogue consistency, specifically focused on its detection and resolution. Our experimental findings indicate that our dataset significantly helps the progress in identifying and resolving conversational inconsistencies, and current popular large language models like ChatGPT which are good at resolving inconsistencies however still struggle with detection.