Dialogue state tracking is the component that remembers what a conversation has established so far. At every turn it updates its picture of what the user wants, and everything downstream, deciding what the system should do and what it should say, reads from that picture.

The difficulty comes from scale rather than from any single conversation. Assistants like Siri, Alexa and Google Assistant sit in front of a large and growing catalogue of services and APIs, each with its own vocabulary of things a user might specify. A tracker that learned one service's slots has to cope with services it was never trained on.

Attending over candidates is a sensible response to that. Rather than treating the state as a fixed set of fields to fill, the model considers the candidate values actually available in context, which is what allows a new service to be handled without a new model.