Socially Interactive Agent for Group Debate Mediation

Abstract

Political polarization makes constructive cross-partisan dialogue increasingly difficult, and human facilitators are subject to bias and cognitive limits when tracking evolving stances in multi-party debates. We propose a Socially Interactive Agent that acts as an active mediator by fusing non-verbal cues with real-time argument mining, for both active speakers and passive listeners. An Argument-related Theory of Mind module maintains and updates each participant’s belief and value profile, enabling targeted, empathetic interventions, optionally instantiated via Thomas-Kilmann conflict-management strategies. The research is organized in three studies: (i) the construction of a multi-party multilingual multimodal debate corpus with fine-grained annotation of stance, agreement, and engagement, in particular including passive listeners; (ii) the design and implementation of the ArgToM-augmented, tool-equipped SIA architecture; and (iii) a real-world deployment and evaluation quantifying the agent’s impact on group polarization and consensus gain relative to unmediated debates.