The Restaurant Game: Learning Social Behavior and Language from Thousands of Players Online

摘要: We envision a future in which conversational virtual agents collaborate with humans games and training simulations. A representation of common ground for everyday scenarios is essential these if they are to be effective collaborators communicators. Effective can infer partner’s goals predict actions. communicators the meaning utterances based on semantic context. This article introduces computational model called Plan Network, statistical that encodes context-sensitive expected patterns behavior language, dependencies social roles object affordances. describe methodology unsupervised learning Network using multiplayer video game, visualization this network, evaluation learned respect human judgment typical behavior. Specifically, we Restaurant from data collected over 5,000 gameplay sessions minimal investment online (MIMO) role-playing game The Game. Our results demonstrate kind sense agents, have implications automatic authoring content future.