An NLP Interface for Social AI Agents in The Resistance: Avalon

Abstract

Social deduction games such as Avalon present a unique challenge for AI agents. To discover the hidden roles of others, players must employ indirectness and deception in their communication. DeepRole, an Avalon-playing AI agent created by MIT researchers in 2019, can communicate through in-game actions but is unable to communicate in natural language. We have created Avalocution, a bot that enhances DeepRole with one-way bot-to-human natural language utterances informed by DeepRole's internal knowledge representation. We hypothesized that our natural language interface would produce direct and indirect communication, exhibit human- like behavior, and provide a positive gameplay experience for human players. We collected survey data from research participants who played Avalon against Avalocution agents, and the survey data supports our hypotheses. We conclude that adding Avalocution's simple one-way utterance generation model to DeepRole's existing decision-making framework captures the nuance of communication required in Avalon while providing an excellent gameplay experience.