Presentation Information
[3Yin-A-64]Analyzing User Simulators for Persuasion Dialogue with User Cognitive Styles
〇Yotaro Watanabe1, Kise Yoshida2, Yuka Saito2, Meguru Takasaki1, Shoji Moriya2, Reina Akama2,3, Jun Suzuki2 (1. PKSHA Technology Inc., 2. Tohoku University, 3. NINJAL)
Keywords:
Dialogue Simulation,Persuasive Dialogue,Cognitive Style
In persuasive dialogue research, user simulation based on large language models (LLMs) plays an important role in evaluating persuasive dialogue agents. Prior studies have mainly incorporated personality traits and decision-making styles as persona information, while cognitive styles related to information processing have received limited attention. In this study, we introduce three cognitive styles, Need for Cognition, Need for Closure, and Maximization, into an LLM-based user simulator and analyze their effects in donation persuasion dialogues. Experimental results indicate that incorporating cognitive styles influences dialogue behavior and structure. In particular, user simulators with cognitive-style information tend to spend more turns in conditional examination phases and exhibit increased diversity in dialogue intents. Persuasion outcomes also show a tendency toward conditional acceptance rather than immediate agreement.
Correlation analysis further reveals significant positive correlations between the modeled values of Need for Cognition and Need for Closure and their corresponding dialogue behaviors.
Correlation analysis further reveals significant positive correlations between the modeled values of Need for Cognition and Need for Closure and their corresponding dialogue behaviors.