Benchmarking Explainable Argumentation Dialogue via Freeman's Theory

Argumentative dialogue involves structured exchanges of claims and supporting evidence, yet progress in building effective dialogue systems is limited by the scarcity of high-quality datasets. To address this, we introduce CMV-AD, a baseline dataset derived from the ChangeMyView corpus, designed for modeling structured argumentative interactions. We further propose FTCoT, a Freeman’s Theory-based chain-of-thought framework that enhances interpretability and reasoning in dialogue generation. FTCoT represents each dialogue turn with a structured quadruple: dialogue summary, user argument, assistant argument, and response reasoning. We construct FTCoT using large language models, leveraging their capabilities in reasoning and data annotation. Extensive automatic and human evaluations demonstrate the effectiveness of FTCoT in improving both the interpretability and quality of generated responses.