Recent advances in large language models (LLMs) have significantly improved the response quality of dialogue systems. However, the issue of dialogue breakdowns where systems produce utterances that confuse users, such as those containing misinformation or lacking common sense still persists. Since such breakdowns can negatively affect users’ impressions of dialogue systems, it is essential to appropriately repair the flow of conversation. Nevertheless, no existing method can robustly handle a wide range of breakdown types. To address this issue, this study aims to develop a dialogue breakdown repair generation system that can robustly handle various types of dialogue breakdowns. Specifically, we train an LLM using the Dialogue Breakdown Repair Corpus, which contains repair utterances corresponding to diverse breakdown scenarios. As a result, we construct a model specialized in generating repair utterances for various breakdown types and demonstrate that it achieves higher accuracy than existing models.

Recent advances in dialogue systems have been remarkable; however, conversational breakdowns still occur, making it essential to develop appropriate repair strategies. Nevertheless, when a system breakdown actually occurs, it remains unclear how the system should perform the repair, and no corpus has been available to investigate this issue. To address this gap, we presented typical examples of system-induced dialogue breakdowns to crowd workers and collected their expected repair utterances toward the broken system. Each repair utterance was annotated with dialogue act tags, and we constructed a breakdown-repair corpus consisting of 3,990 utterances covering ten representative types of breakdowns. This corpus includes breakdown cases across diverse situations, allowing for the examination of various repair patterns. Furthermore, we also conducted a questionnaire on participants’ personal traits, creating a dataset that enables the investigation of repair strategies tailored to individual user characteristics. In this paper, we report an overview of the dataset and preliminary analysis results.