Modeling Brief Alcohol Intervention Dialogue with MDPs for Delivery by ECAs

This paper describes the design of a multimodal spoken dialogue system using Markov Decision Processes (MDPs) to enable embodied conversational virtual health coach agents to deliver brief interventions for lifestyle behavior change in particular excessive alcohol consumption. Its contribution is two fold. First, it is the first attempt to-date to study stochastic dialogue policy optimization techniques in the health dialogue domain. Second, it provides a model for longer branching dialogues (in terms of number of dialogue turns and number of slots) than the usual slot filling dialogue interactions currently available (e.g. tourist information domain). In addition, the model forms the basis for the generation of a richly annotated dialogue corpus, which is essential for applying optimization methods based on reinforcement learning.