Attentive conversational agent with internal state transition for multiple users

Abstract

In this chapter, we discuss how to add emotion to virtual agents and how this influences human-agent communication. Firstly, we conducted a preliminary experiment to investigate which kind of emotions we human beings represent during human to human communication. How do we represent them, and how they change. As a consequence of this analysis, we constructed a state transition model, where the agent has emotions as internal states. We change them according to the particular situation. To realize this model, we thought that the method of machine learning and especially the method of Support Vector Machine(SVM), would be useful. We implemented an agent using that model, and conducted an experiment. We did questionnaire survey and a GNAT test to evaluate the performance of the implemented agent. After analyzing these results combined together, we concluded that our model worked well when representing the natural nonverbal behaviors, which should convey agent's present emotion. We couldn't get a satisfactory result in the aspect of verbal behavior. Further research is necessary especially on the verbal aspect. In addition to this, we also plan to consider a model of users' emotion during the conversation and then integrate it into the model. © Springer-Verlag Berlin Heidelberg 2010.