Exploring Group Discussion with Conversational Agents Using Epistemic Network Analysis

Conversational agents—dialogue systems that provide learning support to students in real time—have shown promise in facilitating science discussion. To enact conversation norms, the appearances, personalities, or tones of the agents often resemble personas that students are familiar with, such as peers or mentors. This study uses epistemic network analysis (ENA) to explore how students interacted with two personas agent (a peer and an expert) in collaborative settings. Data came from chat logs of three student groups with low, mixed, and high ability. The groups interacted with both prototypes. The chat logs received qualitative codes for discussion types, including claim-making, reasoning, building on prior ideas, and responsiveness to the agents. ENA visualized the differences in discussion between groups and between the agent conditions. Overall, the higher-ability groups engaged in more claim-making in tandem with building on prior ideas when interacting with the peer agent, compared to the expert agent. Meanwhile, the low-ability group showed more syntheses of previous ideas when responding to the expert agent. Findings illuminate the design space for adapting agent personas to group settings to facilitate productive exchange.