Jeff Orkin, Deb Roy

Abstract

As a step toward simulating dynamic dialogue between agents and humans in virtual environments, we describe learning a model of social behavior composed of interleaved utterances and physical actions. In our model, utterances are abstracted as {speech act, propositional content, referent} triples. After training a classifier on 100 gameplay logs from The Restaurant Game annotated with dialogue act triples, we have automatically classified utterances in an additional 5,000 logs. A quantitative evaluation of statistical models learned from the gameplay logs demonstrates that semi-automatically classified dialogue acts yield significantly more predictive power than automatically clustered utterances, and serve as a better common currency for modeling interleaved actions and utterances.