Abstract
A major bottleneck in training end-to-end task-oriented dialog systems is the lack of annotated data. To utilize limited training data more efficiently, we propose the Modular Supervision Network (MOSS), an encoder-decoder training framework that incorporates supervision from various intermediate dialog system modules, including natural language understanding (NLU), dialog state tracking (DST), dialog policy learning (DPL), and natural language generation (NLG). With only 60% of the training data, MOSS-all (utilizing supervision from all four modules) outperforms state-of-the-art models on the CamRest676 dataset. Furthermore, MOSS-all demonstrates even greater benefits in complex scenarios; with only 40% of the training data, it outperforms existing models on the LaptopNetwork dataset, a complex Chinese customer service dataset introduced in this work. The MOSS framework is highly flexible, accommodating varying levels of supervision at both the framework and model levels, making it ideal for real-world deployment.
Introduction
Building task-oriented dialog systems—such as those for restaurant reservations or technical support—traditionally involves a trade-off between modularity and end-to-end optimization. Traditional modular systems use a pipeline (NLU → DST → DPL → NLG) that is data-efficient and interpretable but suffers from error propagation and model mismatch. Conversely, end-to-end (E2E) systems map user input directly to responses, avoiding pipeline issues but requiring massive amounts of data and operating as "black boxes."
MOSS addresses these challenges by proposing a hybrid framework that is end-to-end trainable while incorporating rich, modular supervision to guide the learning process. This approach combines the interpretability of modular systems with the joint optimization benefits of E2E models.
The MOSS Architecture
The core of MOSS is a shared-encoder, multi-decoder architecture designed for flexibility and data efficiency.
- Shared Encoder: A bidirectional GRU processes the dialog history and current user utterance into a shared representation.
- Specialized Decoders: Instead of a single response decoder, MOSS employs four decoders corresponding to NLU, DST, DPL, and NLG.
- Hidden State Coupling: Unlike traditional pipelines connected by symbolic text, MOSS decoders pass hidden states to one another. For instance, the DST decoder's final hidden state serves as the prior for the DPL decoder. This ensures the system remains fully differentiable, allowing for joint optimization while mitigating the "broken telephone" effect of symbolic errors.
- Plug-and-Play Flexibility: MOSS can handle incomplete data. If a dataset lacks NLU annotations, that decoder can be omitted. It can also leverage raw, unannotated dialogs to specifically refine the NLG module.
Experimental Evaluation
The framework was evaluated on two benchmarks:
- CamRest676: A standard dataset for restaurant reservations.
- LaptopNetwork: A new, complex dataset featuring real-world Chinese conversations for troubleshooting laptop network issues. It includes 60 different dialog acts and technical, goal-driven interactions, making it significantly more challenging than standard benchmarks.
Key Results and Insights
MOSS demonstrates a significant "data discount," achieving superior performance with a fraction of the training data required by baseline models like TSCP.
- Data Efficiency: On the complex LaptopNetwork task, MOSS-all trained on only 40% of the data outperformed the state-of-the-art TSCP model trained on 100% of the data. On CamRest676, MOSS-all achieved a 0.259 BLEU score and 0.951 Entity Match rate.
- Supervision as Inductive Bias: Intermediate labels (intent, dialog state) act as a powerful guide. By learning these intermediate steps, the model internalizes the underlying logic of the task faster than pure E2E models.
- Interpretability: Because MOSS has individual decoders for NLU and DST, developers can inspect intermediate outputs to diagnose failures (e.g., determining if a wrong response was due to a tracking error or a policy error).
- Ablation Studies: Removing any single module of supervision (NLU or DPL) resulted in performance drops, confirming that each component of the modular roadmap contributes to the model's overall success.
Conclusion
MOSS provides a flexible and efficient framework for building task-oriented dialog systems. By leveraging modular supervision, it overcomes the data-hungry nature of E2E systems while maintaining differentiability. Future work will explore applying this framework to more open-ended domains and integrating reinforcement learning to further refine dialog policies in real-time.
Summary updated