A Model for Context-sensitive Interpretation of Communicative Feedback

A prevalent and important coordination mechanism in dialogue is communicative 'listener feedback' in form of short verbal-vocal signals (e.g., 'uh-huh', 'yeah', 'huh'), head gestures (e.g., nodding, wiggling) or facial expressions. Listeners' feedback signals are indications of their cogni-tive state and communicate whether listeners and speakers are in contact , whether listeners perceive, understand and accept the speakers' utterance as well as whether they agree to or which attitude they have towards the speakers' utterance (Allwood et al., 1992). Feedback signals are rich in their form – enabling a fine-grained expression of subtle differences in meaning –, multi-functional, and interact heavily with their dialogue context. Consequentially, feedback is only conventionalised to a certain degree. Speakers are nevertheless able to interpret communicative feedback, use it to reason about the lis-tener's cognitive state as well as their common ground, and adapt their language accordingly (Clark, 1996). Our objective is to model this ability of human speakers for artificial conversational agents (e.g., dialogue systems, embodied conversational agents), making them attentive to their users' needs as expressed in their feedback behaviour (Buschmeier & Kopp, 2011). Here we present an agent-centric cognitive model for the interpretation of feedback signals in their dialogue context. It uses features of the feedback behaviour observed by the agent as well as the agent's utterance , expectations and knowledge about the task to reason about an 'attributed listener state' (ALS; the agent's reconstruction of the user's cognitive state) and the grounding status of information. Using the framework of Bayesian networks, the model represents ALS and common ground probabilistically in terms of degrees of belief. Because of this, it is straightforward to deal with the uncertainties inherent in the observations, the form-meaning relation of a feedback signal and the user's behaviour.