Dialogue Breakdown Analysis in a Multi-Agent Non-Task-Oriented Conversation

Abstract

In this study, we analyze utterances that cause dialogue breakdown in multi-agent non-task-oriented conversation using a Large Language Model (LLM). Although simulations of human behavior using agents based on large-scale language models have been proposed, it is crucial to analyze the content of the dialog, which is one of the key factors in achieving more reliable simulations. We investigate the tendency of LLM agents' utterances to cause breakdowns in non-task-oriented dialogues. We annotate the causes of failure for the dialogue between the two agents, given the theme and the attributes of the agents. As a result of our analysis of the typology of utterances that lead to dialogue breakdowns, we found speech errors specific to LLM agents.