Virtual humans are embodied software agents designed to simulate the appearance and social behaviors of humans, typically with the goal of facilitating natural interactions between humans and computers. They play an important role in the advancement of today's immersive virtual worlds, including domains such as virtual training (Swartout et al., 2006), education (Rowe et al., 2010), and health care (Bickmore et al., 2010).;One of the key challenges in creating virtual humans is giving them human-like nonverbal behaviors. There has been extensive research on analyzing and modeling human nonverbal behaviors. Some of them rely on results from observing and manually analyzing human behaviors, while others approach the problem by exploring advanced machine learning techniques on large amounts of annotated human behavior data. However, little attention has been paid to the "data" these systems learn from.;In this thesis, we propose a new methodology called Parasocial Consensus Sampling (PCS) to approach the problem of modeling human nonverbal behaviors from the "data" perspective. It is based on previous research on Parasocial Interaction theory (Horton & Wohl, 1956). The basic idea of Parasocial Consensus Sampling is to have multiple independent participants experience the same social situation parasocially (i.e. act "as if" they were in a real dyadic interaction) in order to gain insight into the typicality of how individuals would behave within face-to-face interactions.;First, we validate this framework by applying it to model listener backchannel feedback and turn-taking behavior. The results demonstrate that (1) people are able to provide valid behavioral data in parasocial interaction, (2) PCS data generates better virtual human behaviors and (3) can be used to learn better prediction models for virtual human. Second, we show that the PCS framework can help us tease apart the causalities of the variability of human behavior in face-to-face interactions. Such research work would be difficult to perform by traditional approaches. Moreover, PCS enables much larger scale and more efficient data collection method than traditional face-to-face interaction. Finally, we integrate the PCS-data driven models into a virtual human system, and compare it with a state-of-the-art virtual human application, the Rapport Agent (Gratch et al., 2007) in real interactions. Human subjects are asked to evaluate the performance of each agent regarding the correctness of the agents' behaviors, the rapport they feel during the interactions and the overall naturalness. The results suggest that the new agent predicts the timing of backchannel feedback and end-of-turn more precisely, performs more natural behaviors and thereby creates much stronger feeling of rapport between users and agents.
展开▼
