We aim to build dialogue agents that optimize the dialogue strategy, specifically through learning the dialogue model components from dialogue data. In this paper, we describe our current research on automatically learning dialogue strategies in the healthcare domain. We go through our systematic approach of learning dialogue model components from data, specifically user intents and the user model, as well as the agent reward function. We demonstrate our experiments on healthcare data from which we learned the dialogue model components. We conclude by describing our current research for automatically learning dialogue features that can be used in representing dialogue states and learning the reward function.
展开▼