Communicating with multiple users for embodied conversational agents in quiz game context

Abstract

This paper presents two approaches to achieve attentiveness of a virtual quiz agent in the interactions with multiple users at the same time. One attempts to improve the agent with an utterance strategy to determine whether, when, and whom to suggest the users the availability of hints or to urge the users to answer. In the other one, a transition state model presenting the agent's attitude that allows the agent to perform meaningful idle motions is introduced. The transition model is obtained in a Wizard-of-Oz (WOZ) experiment and is learned with support vector machine (SVM). These two prototype systems are evaluated with an automatic attitude test, go/no-go task (GNAT) and regular questionnaires. From the joint results of the evaluation subject experiments, the direction of finding an appropriate triggering timing of the agent's behaviours is proved to be effective.