Probabilistic Ontology Trees for Belief Tracking in Dialog Systems

We introduce a novel approach for robust belief tracking of user intention within a spoken dialog system. The space of user intentions is modeled by a probabilistic extension of the underlying domain ontology called a probabilistic ontology tree (POT). POTs embody a principled approach to leverage the dependencies among domain concepts and incorporate corroborating or conflicting dialog observations in the form of interpreted user utterances across dialog turns. We tailor standard inference algorithms to the POT framework to efficiently compute the user intentions in terms of m-best most probable explanations. We empirically validate the efficacy of our POT and compare it to a hierarchical frame-based approach in experiments with users of a tourism information system.