[2Win5] Poster session 2

[2Win5-42] Extended Strategy Design for Persuasive Dialogue Agents and an Automatic Evaluation Method Using Intention Shift

〇Yuto Nakano¹, Shinnosuke Nozue¹, Yotaro Watanabe³, Meguru Takasaki³, Reina Akama^1,2, Jun Suzuki^1,2 (1.Tohoku University, 2.Institute of Physical and Chemical Research, 3.PKSHA Technology Inc.)

Keywords:Dialogue System, Persuasive Dialogue, Strategy Design, Automatic Evaluation of Persuasive Dialogue, Chain of Thought

Persuasive dialogue agents play a crucial role in driving behavioral change, yet their effectiveness depends on users' intent and engagement levels. This paper explores a fine-grained framework for modeling persuasive states and an expanded set of persuasive strategies, leveraging insights from fields such as behavioral and social psychology. To systematically quantify subtle shifts in user intent, we introduce Average Intention Shift (AIS), which captures incremental intention changes over interactions. We evaluate our approach through a simulation-based study across multiple experimental conditions, showing significant improvements in metrics such as persuasion success rates and AIS. Notably, our approach proves particularly effective in influencing users with low initial intent, addressing a key challenge in persuasive AI. Moreover, our analysis reveals that the optimal persuasion strategy varies depending on users' initial intent levels, highlighting the necessity of adaptive, intent-aware persuasion models.