What key advantage did treating DST as an MRC task provide over earlier fixed-ontology methods?

Answer

MRC models inherently handle open vocabularies by pointing directly to the answer span.

A significant limitation of earlier approaches, especially those relying on fixed sets of enumerated states or predefined ontologies, was their rigidity regarding possible slot values. The Machine Reading Comprehension (MRC) framing, particularly when implemented with attention-based neural networks, overcame this by allowing the model to point directly to the precise answer span within the conversation text. This capacity to deal with *open vocabularies* meant the model was not constrained to a predefined list of acceptable values, thus addressing the scalability and flexibility issues that plagued models dependent on fixed state definitions.

What key advantage did treating DST as an MRC task provide over earlier fixed-ontology methods?
inventionAIStatetrackingconversation