Abstract
Robust dialogue belief tracking is a key component in maintaining goodquality dialogue systems. The tasks that dialogue systems are trying to solveare becoming increasingly complex, requiring scalability to multi domain,semantically rich dialogues. However, most current approaches have difficultyscaling up with domains because of the dependency of the model parameters onthe dialogue ontology. In this paper, a novel approach is introduced that fullyutilizes semantic similarity between dialogue utterances and the ontologyterms, allowing the information to be shared across domains. The evaluation isperformed on a recently collected multi-domain dialogues dataset, one order ofmagnitude larger than currently available corpora. Our model demonstrates greatcapability in handling multi-domain dialogues, simultaneously outperformingexisting state-of-the-art models in single-domain dialogue tracking tasks.