Winnie: Task-Oriented Dialog System with Structure-Aware Contrastive Learning and Enhanced Policy Planning

Kaizhi Gao, Tianyu Wang, Zhongjing Ma, Suli ZouAAAI 2024 · DOI

TL;DR

This paper introduces Winnie, an end-to-end Task-Oriented Dialog (TOD) system that improves performance by incorporating dialog structure awareness through supervised contrastive learning and enhancing policy planning with an auxiliary system act classification task. Winnie achieves state-of-the-art results on several benchmark datasets without requiring additional pre-training corpora.

What Problem Does This Paper Address?

Task-Oriented Dialog (TOD) systems aim to help users accomplish specific goals, like booking a hotel or making a reservation. Traditional TOD systems often use a modular pipeline (understanding, state tracking, policy planning, response generation), which can lead to error propagation. While end-to-end models have emerged, they often struggle to fully leverage the inherent structure of dialogs (user utterances, belief states, database results, system acts, and responses) and effectively integrate policy planning. Existing datasets also often lack intermediate states, making it difficult to pre-train models that understand this structure. This paper addresses the challenge of building a more robust and efficient end-to-end TOD system that can better capture dialog structure and improve policy decision-making.

Methods

Winnie builds upon a T5 encoder-decoder architecture. Its key innovations are:

  • Dialog-Level Transformer Layer: To capture the structural information within a dialog, a Transformer layer is added to process utterance representations. This layer focuses on modeling the long-range dependencies and interrelationships between different turns in a conversation, complementing the semantic information captured by the base T5 encoder.
  • Supervised Contrastive Learning (SCL): SCL is employed to differentiate representations of utterances belonging to different dialog categories while encouraging cohesion among utterances of the same category. This helps the model better understand the dialog structure by learning to distinguish between semantically similar and dissimilar conversational elements.
  • System Act Classification (SAC) as an Auxiliary Task: To improve policy planning, an auxiliary task of classifying system acts is introduced during fine-tuning. This explicit injection of policy information helps the model capture more detailed system act annotations and accelerates convergence, leading to better decision-making.

The model combines these components with a residual connection to ensure the dialog-level Transformer learns structural information without overwriting the semantic information from the encoder.

Datasets

The proposed Winnie model is evaluated on three well-established task-oriented dialog benchmarks:

  • MultiWOZ2.2: A large-scale dataset covering seven domains, known for its complex ontology and diverse language styles.
  • CamRest676: A restaurant-domain dataset.
  • Stanford In-Car Assistant: A dataset for in-car assistant tasks like calendar scheduling and navigation.

The paper uses standard splits for training, validation, and testing for these datasets.

Key Insights and Ideas

  • Dialog Structure is Crucial: The paper strongly emphasizes that simply fine-tuning pre-trained language models on dialog data isn't enough. Explicitly modeling the structure of the conversation (how turns relate, what information is conveyed) is key to improving performance.
  • SCL for Structure: Supervised Contrastive Learning proves to be an effective mechanism for learning representations that capture dialog structure. By pushing apart dissimilar utterances and pulling together similar ones, it helps the model discern the nuances of conversational flow.
  • Policy Planning Boost: Adding system act classification as an auxiliary task significantly improves the model's ability to make correct policy decisions, which in turn leads to better response generation. This suggests that explicit guidance on system actions is more beneficial than relying solely on end-to-end generation.
  • Efficiency and Simplicity: Winnie achieves state-of-the-art results without requiring additional pre-training corpora or complex data augmentation, making it a practical and efficient approach.

Key Results

Winnie achieves state-of-the-art performance on all three evaluated datasets:

  • MultiWOZ2.2: Winnie achieves a combined score of 103.4, outperforming the previous state-of-the-art model (MTTOD) by 3.2 points and GALAXY by 3.2 points. It also shows significant improvements in inform and success rates.
  • CamRest676: Winnie achieves a combined score of 118.3, surpassing GALAXY by 1.0 point. It also achieves a perfect match rate of 100.0.
  • In-Car Assistant: Winnie also sets a new state-of-the-art on this dataset, being the only model to achieve top performance across inform rate, success rate, BLEU score, and combined score.

Ablation studies confirm that both SCL and SAC individually improve performance, but their combination synergistically enhances the overall end-to-end dialog modeling.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper notes that the performance of SCL is sensitive to the temperature hyperparameter ($\tau$), and finding the optimal value is crucial. While they identify 15.0 as optimal for MultiWOZ2.2, this might vary across datasets.
  • Future Work: The authors suggest that their approach of incorporating SCL and SAC can be extended to other TOD systems. They also imply that further exploration into how to best leverage dialog structure and policy information could lead to even greater improvements. The success of Winnie suggests that future research could focus on more sophisticated ways to integrate structural and policy learning into end-to-end dialog models.

Summary updated

BibTeX
@inproceedings{gao2024winnie,
  author = {Kaizhi Gao and Tianyu Wang and Zhongjing Ma and Suli Zou},
  title = {{Winnie: Task-Oriented Dialog System with Structure-Aware Contrastive Learning and Enhanced Policy Planning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i16.29758},
}
Figure