Abstract
Dialogue State Tracking (DST) is a critical component in task-oriented dialogue systems, responsible for maintaining a representation of user goals throughout a conversation. However, existing DST models often suffer from error propagation—where incorrect predictions in early turns derail subsequent tracking—and struggle to model complex dependencies between different slots. This paper introduces the Dialogue State Distillation Network (DSDN), a framework that utilizes knowledge distillation to mitigate error propagation and an inter-slot contrastive learning module to capture dynamic slot correlations. Experimental results on the MultiWOZ benchmarks demonstrate that DSDN achieves state-of-the-art performance, particularly in long-turn dialogues.
Introduction
In task-oriented dialogue systems, the Dialogue State Tracker (DST) parses user utterances to update a set of (domain, slot, value) triplets, such as (hotel, area, north). Accurate tracking is essential for the system to provide relevant responses. Despite significant progress, two major challenges remain:
- Error Propagation (The Snowball Effect): Many state-of-the-art models rely on the predicted state from the previous turn to inform the current turn's prediction. If an error occurs early, it persists and often compounds, leading to a total failure in tracking the user's intent as the dialogue progresses.
- Slot Independence vs. Correlation: Traditional models often treat slots as independent entities. However, in real-world scenarios, slots are highly correlated (e.g., specifying a
hotel-nameoften implies a specifichotel-area). Capturing these "co-updates" typically requires complex, static graph structures that lack flexibility across different contexts.
Methodology
The proposed DSDN framework addresses these challenges through two innovative modules: Dialogue State Distillation and Inter-slot Contrastive Learning.
Dialogue State Distillation (DSDN)
To break the cycle of error propagation, the authors employ a knowledge distillation strategy. During training, a Teacher network is provided with the ground-truth ("golden") dialogue state from the previous turn. This allows the Teacher to learn optimal representations for predicting the current state. Simultaneously, a Student network is trained using only the dialogue context (utterances).
The Student is supervised not only by the ground-truth labels but also by a distillation loss that encourages its internal representations to match those of the Teacher. At inference time, only the Student network is used. By learning to mimic the Teacher's "informed" state, the Student can effectively utilize historical cues from the dialogue history without being vulnerable to the errors of its own previous predictions.
Inter-slot Contrastive Learning
To model slot dependencies dynamically, the authors introduce an inter-slot contrastive learning loss. Instead of relying on a fixed schema, the model learns to group related slots in the embedding space.
- Positive Samples: Slots that are updated simultaneously within the same turn are treated as positive pairs, and their representations are pulled closer together.
- Negative Samples: Slots that are not updated, or the same slot from different dialogue turns, are treated as negative pairs and pushed apart.
This approach allows the model to discover and exploit co-update patterns directly from the data, enhancing its ability to predict related slots accurately.
Experimental Results
The DSDN model was evaluated on the MultiWOZ 2.0 and 2.1 datasets, which are standard benchmarks for multi-domain DST.
Performance Comparison
As shown in the table below, DSDN achieves a new state-of-the-art Joint Goal Accuracy (Joint GA) of 59.25% on MultiWOZ 2.1, outperforming competitive baselines such as LUNA and PPTOD.
Robustness to Error Propagation
A key finding of the study is the model's performance over time. While most models see a sharp decline in accuracy as the number of dialogue turns increases, DSDN maintains significantly higher Joint GA in later turns. This confirms that the distillation approach effectively mitigates the "snowball effect" of error propagation, as the Student network learns to extract relevant state information directly from the dialogue history.
Ablation Study
Ablation experiments highlighted the importance of each component. Removing the Dialogue State Distillation module resulted in a 1.77% drop in Joint GA on MultiWOZ 2.0, while removing the Inter-slot Contrastive Learning module led to a 1.05% decrease. The combined effect of both modules provides the substantial performance gains observed, validating the dual-pronged approach to solving DST's core issues.
Conclusion
The Dialogue State Distillation Network (DSDN) represents a significant step forward in robust Dialogue State Tracking. By decoupling the model's reliance on previous predictions through distillation and learning dynamic slot correlations via contrastive learning, DSDN achieves superior performance and greater stability in long conversations. Future work may involve integrating this framework with larger-scale pre-trained language models to further enhance tracking capabilities in complex, open-ended domains.
Summary updated