Probabilistic Ontology Trees for Belief Tracking in Dialog Systems

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