Answers are grounded in this paper and cite their sources.
Dialogue state tracking (DST) is a core component of multi-turn task-oriented question answering, as it continuously maintains and updates the structured belief state from user interactions. In realistic conversational scenarios, however, DST often encounters unstable slot updates caused by domain switching, implicit reference, omitted information, and user self-correction across turns. These challenges are particularly evident when the model relies only on surface semantic matching while underutilizing the schema structure behind domain-slot-value definitions. More specifically, existing methods still lack an explicit mechanism for slot selection before update and often underexploit schema as a structural routing signal for belief state evolution. To address this issue, this paper proposes SMART-DST, a schema-guided multi-turn adaptive routing method for dialogue state tracking and slot updating. The proposed method treats schema not merely as a static ontology constraint, but as an active signal for routing dialogue understanding and belief state evolution. Specifically, SMART-DST first performs schema-aware state routing by modeling domain-slot associations, slot dependency relations, and value-type priors, enabling the model to focus on the most relevant subspace of the dialogue state under the current utterance. On this basis, a dual-stage slot update mechanism is introduced to decide whether a slot should be retained, modified, overwritten, or cleared before estimating its final value. This design reduces error propagation in long conversational contexts and improves the robustness of slot transitions under ambiguous user expressions. Meanwhile, the explicit incorporation of schema semantics also enhances the interpretability of state updates. Experimental results on multi-turn dialogue benchmarks show that SMART-DST consistently improves dialogue state tracking accuracy and slot update reliability over competitive baselines, for example achieving 64.7% joint goal accuracy on MultiWOZ 2.1 and 66.2% overall joint goal accuracy on SGD, compared with 61.5% and 63.0% obtained by the strongest baseline, respectively, especially in complex cases involving cross-domain shifts, ellipsis, and user self-revision.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: