Adapting a virtual advisor’s verbal conversation based on predicted user preferences: a study of neutral, empathic and tailored dialogue

Virtual agents that improve the lives of humans need to be more than user-aware and adaptive to the user’s current state and behavior. Additionally, they need to apply expertise gained from experience that drives their adaptive behavior based on deep understanding of the user’s features (such as gender, culture, personality, and psychological state). Our work has involved extension of FAtiMA (Fearnot AffecTive Mind Architecture) with the addition of an Adaptive Engine to the FAtiMA cognitive agent architecture. We use machine learning to acquire the agent’s expertise by capturing a collection of user profiles into a user model and development of agent expertise based on the user model. In this paper, we describe a study to evaluate the Adaptive Engine, which compares the benefit (i.e., reduced stress, increased rapport) of tailoring dialogue to the specific user (Adaptive group) with dialogues that are either empathic (Empathic group) or neutral (Neutral group). Results showed a significant reduction in stress in the empathic and neutral groups, but not the adaptive group. Analyses of rule accuracy, participants’ dialogue preferences, and individual differences reveal that the three groups had different needs for empathic dialogue and highlight the importance and challenges of getting the tailoring right.