The "Etiquette Quotient"; Evaluating Social Skills in Conversational Avatars

Making avatars react appropriately in social interaction--to take offense when reasonable, to give deference where appropriate, etc.-- is a more fundamental need for believability and cost-effectiveness than is accuracy in appearance, especially for military applications such as cross-cultural training. We propose using a rich, universal theory of human-human "politeness" behaviors and the culture-specific interpretive frameworks for them (labeled "etiquette") from sociology, linguistics and anthropology to create a computational model of social behavior expectations. This model will link observable and inferred aspects of power and familiarity relationships, the degree of imposition of an act (all of which have implications for roles and intents) and the actor's character to produce politeness behaviors expectations. By using observations of politeness behaviors (or its lack), the same model permits inferences and updates about those attributes. In Phase I, we will refine and implement this model to provide a computational believability metric based on the delta between observed and expected politeness behaviors--an "Etiquette Quotient" (EQ)--of an actor in context. We will also validate and tune this measure in an experiment using avatars. In Phase II, we will expand this model with cross-cultural etiquette libraries and use it dynamically adapt avatars' social behavior.