Abstract
Dialogue State Tracking (DST) is a critical component of task-oriented dialogue systems, responsible for extracting user constraints and goals from conversation history. Traditional DST models often rely on complex, task-specific architectures that struggle with zero-shot transfer and cryptic schema definitions. This paper proposes a simplified, unified framework that frames DST as a text-to-text generation task using the T5 language model. By employing schema-driven prompting and natural language descriptions, the model achieves state-of-the-art performance on the MultiWOZ 2.2 benchmark without requiring specialized classifiers for different slot types.
Introduction
In task-oriented dialogue, the system must maintain a "dialogue state"—a set of slot-value pairs (e.g., (hotel, area, centre))—to track user requirements. Historically, DST research has focused on building specialized architectures to handle "categorical" slots (fixed vocabularies) and "non-categorical" slots (open-ended text) separately. These approaches often fail to generalize to new domains where slot names may be abbreviated or ambiguous. Furthermore, sequential decoding of the entire state can lead to error propagation in long dialogues.
Methodology
The authors introduce a generative approach based on the T5 (Text-to-Text Transfer Transformer) model, utilizing two key innovations: Independent Decoding and Schema-Driven Prompting.
Independent Decoding
Unlike sequential models that generate the entire state at once, the proposed model performs a separate inference pass for each slot in the schema. This "independent decoding" prevents errors in one slot from cascading into others and allows the model to focus its attention on a specific task-aware context.
Schema-Driven Prompting and NL Augmentation
For each slot, the input to the T5 encoder consists of the dialogue history concatenated with a prompt. A basic prompt includes the domain and slot name (e.g., [domain] train [slot] day). To further enhance understanding, the authors append natural language (NL) descriptions directly from the database documentation. For instance, a cryptic slot like ref is augmented with "reference number of the hotel booking."
This allows the model to leverage its pre-trained linguistic knowledge to map conversational cues to schema requirements. By treating all slots as text generation tasks, the model simplifies the pipeline, handling categorical and non-categorical slots identically. Including the list of possible values for categorical slots in the prompt acts as a soft constraint, guiding the generative decoder without requiring a hard-coded classifier.
Experimental Evaluation
The model was evaluated on three major benchmarks:
- MultiWOZ 2.2: The primary testbed, featuring over 10,000 dialogues across 8 domains with rich schema metadata.
- MultiWOZ 2.1: Used to test robustness against less complete schema descriptions.
- M2M (Machines Talking to Machines): A smaller dataset used to verify performance across varying conversation styles.
Results and Discussion
The proposed model achieved a new state-of-the-art Joint Goal Accuracy (JGA) of 57.6% on MultiWOZ 2.2, outperforming specialized models like TRADE and DS-DST.
Key findings include:
- Scaling Benefits: Performance improved significantly when moving from T5-small (60M parameters) to T5-base (220M parameters), particularly when NL descriptions were included.
- The Importance of Descriptions: Ablation studies showed that slot-level descriptions are more critical than domain-level descriptions. The model uses these descriptions as a semantic "bridge" to understand the meaning of slots.
- Annotation Quality: Error analysis revealed that over 53% of the model's "errors" were actually due to noise or incorrect labels in the original dataset, suggesting the model's underlying capability is even higher than the metrics indicate.
Conclusion
This work demonstrates that a general-purpose language model, when guided by well-constructed prompts and natural language schema descriptions, can outperform complex, task-specific DST architectures. While independent decoding increases inference latency, the gains in accuracy and simplicity make this a compelling approach for robust dialogue state tracking. Future work will explore domain adaptation and more efficient decoding strategies to mitigate latency.
Summary updated