Belief Tracking with Stacked Relational Trees

摘要: We describe a new model for Dialog State Tracking called Stacked Relational Tree, which naturally models complex relationships between entities across user utterances. It can represent multiple conversational intents and the change of focus them. Updates to are made by rule-based system in language tree regular expressions. also introduce probabilistic version that handle ASR/NLU uncertainty. show how parameters be trained from log data, showing gains on variety standard Belief Tracker metrics, measurable impact success rate an end-to-end dialog TV program discovery.