Abstract
Dialogue State Tracking (DST) is a critical component of task-oriented dialogue systems, yet its reliance on large-scale human-labeled data presents a significant bottleneck. This paper introduces a novel dual-learning framework that treats DST as a sequence generation task rather than a classification problem. By pairing a state tracker with a dual utterance generator, the system can leverage unlabeled dialogues through a closed-loop feedback mechanism of reconstruction errors and reward signals. Experimental results on the MultiWOZ 2.1 benchmark demonstrate that this approach achieves state-of-the-art performance among non-pretrained models and maintains high accuracy even with minimal labeled data.
Introduction
Dialogue State Tracking (DST) serves as the "brain" of task-oriented chatbots, estimating user goals—such as finding a "cheap hotel" in the "center"—at every turn of a conversation. Traditionally, training these models requires massive amounts of human-labeled data, which is expensive and time-consuming to produce. Previous attempts to utilize unlabeled data often suffered from "reward sparsity," as standard classification-based DST models provide fragmented signals that are difficult for a dual-learning agent to interpret effectively.
Methodology
To overcome the limitations of traditional DST, the authors introduce two primary innovations: reframing DST as sequence generation and implementing a dual-learning loop.
DST as Sequence Generation
Instead of treating dialogue states as isolated triples, the model represents the state as a structured sequence (e.g., <hotel> <price> cheap <area> centre </hotel>). This allows the model to capture inter-slot relationships and utilize techniques from Neural Machine Translation (NMT), such as BLEU scores and Language Model (LM) probabilities, as external rewards.
The Dual Learning Framework
The system consists of a "primal" agent (the Tracker) and a "dual" agent (an Utterance Generator). These agents interact through two distinct loops:
- State-Context-State Loop: The generator creates a synthetic utterance from a given state; the tracker then attempts to reconstruct the original state from that utterance.
- Context-State-Context Loop: The tracker predicts a state from an unlabeled dialogue; the generator then tries to recreate the original user utterance based on that predicted state.
Coarse-to-Fine Tracker Architecture
The tracker utilizes a four-part architecture to handle sequence generation efficiently: a Context Encoder for dialogue history, a State Sketch Decoder to generate the structure (domains and slots), and a Sketch-Guided Decoder to fill in specific values. This "sketch-first" approach simplifies the learning process by separating structural decisions from content generation.
Experimental Results
The model was evaluated on MultiWOZ 2.1, the largest benchmark for multi-domain DST. The proposed "dual-DST" achieved a 49.88% joint goal accuracy, outperforming all other models that do not rely on massive pre-trained encoders like BERT.
Efficiency and Low-Resource Performance
Because the model generates the entire state sequence at once, it achieves an inference time complexity of O(1). Furthermore, the dual-learning framework acts as a powerful bootstrapping mechanism. As shown in the performance analysis, the model significantly outperforms standard semi-supervised baselines like pseudo-labeling when labeled data is scarce.
Even with only 20% of labeled data, the dual-learning approach provides a cohesive enough signal to reach performance levels comparable to models trained on much larger datasets, proving that closed-loop feedback is more effective than simple label estimation.
Conclusion
The dual-learning framework for DST effectively reduces the dependency on labeled data by leveraging the structural relationships within dialogue states. While the current model relies on delexicalized scripts, future work may involve integrating pre-trained models like BERT or GPT to further enhance nuance and accuracy. This research marks a significant step toward more efficient and scalable task-oriented dialogue systems.
Summary updated