Parasocial Consensus Sampling: Combining Multiple Perspectives to Learn Virtual Human Behavior

Virtual humans are embodied software agents that should not onlybe realistic looking but also have natural and realistic behaviors. Traditional virtual human systems learn these interactionbehaviors by observing how individuals respond in face-to-facesituations (i. e. , direct interaction). In contrast, this paperintroduces a novel methodological approach called parasocialconsensus sampling (PCS) which allows multiple individuals tovicariously experience the same situation to gain insight on thetypical (i. e. , consensus view) of human responses in socialinteraction. This approach can help tease apart what isidiosyncratic from what is essential and help reveal the strength ofcues that elicit social responses. Our PCS approach has severaladvantages over traditional methods: (1) it integrates data frommultiple independent listeners interacting with the same speaker, (2) it associates probability of how likely feedback will be givenover time, (3) it can be used as a prior to analyze and understandthe face-to-face interaction data, (4) it facilitates much quickerand cheaper data collection. In this paper, we apply our PCSapproach to learn a predictive model of listener backchannelfeedback. Our experiments demonstrate that a virtual humandriven by our PCS approach creates significantly more rapportand is perceived as more believable than the virtual human drivenby face-to-face interaction data.

Authors

• Lixing Huang
• Louis-Phillippe Morency
• Jonathan Gratch

Keywords

No keywords are indexed for this paper.