Inferring pragmatics from dialogue contexts in simulated virtual agent games

Abstract

Virtual agents in video games may conduct two types of interactions: physical and dialogical. While the former is recognized as gazes and gestures, which received significant attention, the latter is often simplified in simulated virtual agent games. However, dialogical interactions affect the mental states of individual agents, and the relations between them, therefore playing a more important role than physical interactions in games. An implemented dynamic Bayesian Network (DBN) based on speech acts is proposed to model the dialogical effects as dialogue contexts in different aspects, such as emotion states, social relations, and social roles. We adopt a scene in the famous movie Doubt that has 53 dialogue sentences as the test corpus and implement 21 types of speech acts in the experiments. The results indicate that, with our DBN model, agents have the ability of context awareness to infer indirect speech acts from given direct speech acts, and that this ability may assist agents to plan dialogues based on speech acts in future work. © 2012 Springer-Verlag.