Proposal of a Numeric Model of Mutual Evaluation between the Participants in Dyadic Active Listening Conversation for Developing Listener Agents

Our project is aiming to develop an active listener agent who provides mental care to its users. This is a task even difficult for humans and thus there exists an issue: whether it is possible to build such an agent. In addition, numeric metrics representing the state of the active listening conversation are required for the decision making of the agent. We propose a model representing the dynamics of the attitude and mood of the two participants of active listening conversation and verified the possibility of building an active listener agent with empirical results. We analyzed the corpus collected in a human-human dyadic conversation experiment from three view points, the speaker (potential user), the listener (the role that the agent will play), and an observer who did not participate the conversation (who is supposed to be able to observe the conversation from a more objective view). Encouraging results that suggest the possibility of the development of an active listener agent were found in the analysis: the attitude of the listener can have an influence on the speaker’s mood, the third person can detect speaker’s attitude, and the mood of the speaker can be potentially observed by another person. Finally, we show that the subjective evaluation on the speakers’ attitude can be estimated by low-level features like smiles, head nods, and speech state.