Communicative Listener Feedback in Human-Agent Interaction: Artificial Speakers Need to Be Attentive and Adaptive

In human dialogue, listener feedback is a pervasive phenomenon that serves important functions in the coordination of the conversation, both in regulating its flow, as well as in creating and ensuring understanding between interlocutors. This make feedback an interesting mechanism for conversational human-agent interaction. In this paper we describe computational models for an attentive speaker' agent is able to (1) interpret the feedback behaviour of its human interlocutors by probabilistically attributing listening-related mental states to them; (2) incrementally adapt its ongoing language and behaviour generation to their needs; and (3) elicit feedback from them when needed. We present a semi-autonomous interaction study, in which we compare such an attentive speaker agent with agents that either do not adapt their behaviour to their listeners' needs, or employ highly explicit ways of ensuring understanding. The results show that human interlocutors interacting with the attentive speaker agent provided significantly more listener feedback, felt that the agent was attentive to, and adaptive to their feedback, attested the agent a desire to be understood, and rated it more helpful in resolving difficulties in their understanding.