UniGDD: A Unified Generative Framework for Goal-Oriented Document-Grounded Dialogue

Chang Gao, Wenxuan Zhang, Wai Lam

Introduction

Goal-Oriented Document-Grounded Dialogue (GDD) is a critical task in natural language processing, enabling automated systems to assist users by extracting and synthesizing information from technical documents, manuals, or policies. Traditionally, this task has been approached using a two-stage pipeline: Knowledge Identification (KI), where the system searches for relevant information within a document, and Response Generation (RG), where that information is transformed into a natural conversational response.

However, this pipeline architecture suffers from two primary flaws. First, it is susceptible to error propagation; if the KI component selects the wrong information, the RG component is fundamentally unable to produce a correct response. Second, it fails to capture the inherent interdependence between the two tasks. The phrasing of a response often depends on the specific nuances of the identified knowledge, and the requirements of the response should ideally guide the search for knowledge. To address these limitations, this paper introduces UniGDD, a unified generative framework that merges these traditionally separate steps into a single sequential process.

The UniGDD Framework

UniGDD moves away from the disconnected pipeline model in favor of a cohesive generative process. By treating knowledge identification and response generation as a single sequence-to-sequence task, the model can leverage mutual supervision between the two objectives.

1. Unified Sequence Generation

UniGDD utilizes a single encoder-decoder architecture (specifically based on T5). The model takes the dialogue history and the relevant document as input and generates a single output string. This string contains both the grounding knowledge and the final agent response, separated by special tokens:

[Input] -> <grounding> [Knowledge] <agent> [Response]

This format allows the model to learn the relationship between the source text and the final conversational output in a single pass, ensuring that the generated response is grounded directly in the identified knowledge.

2. Prompt-Connected Multi-Task Learning (PCMTL)

To ensure the model effectively learns both the "finding" and "writing" aspects of the task, the authors employ Prompt-Connected Multi-Task Learning. By using specific prompts—such as those instructing the model to generate only the grounding or only the response—the framework provides auxiliary supervision. Because these prompts share the same underlying structure as the main unified task, they act as "connective tissue," helping the model understand how individual sub-tasks contribute to the final output. This approach leverages the strengths of prompt-based learning to improve the model's flexibility and accuracy.

3. Linear Temperature Scheduling (LTS)

Documents used in GDD are often lengthy and contain significant amounts of irrelevant information. To help the model focus on the most pertinent sections, the authors introduced Linear Temperature Scheduling (LTS) for the attention mechanism. Early in the training process, the attention is "soft" or broad, allowing the model to explore the entire document. As training progresses, the "temperature" is linearly decreased, forcing the attention to become "sharper" and more concentrated on the specific spans required for the task. This mechanism effectively filters out document noise over time.

Experimental Results

The framework was evaluated on the Doc2Dial benchmark, a large-scale dataset containing over 4,800 dialogues based on technical documents. The results demonstrate that UniGDD significantly outperforms traditional pipeline models.

State-of-the-Art Performance

UniGDD-large achieved a new state-of-the-art (SOTA) on Doc2Dial, reaching 66.9 Exact Match (EM) for knowledge identification and 42.9 BLEU for response generation. In comparison, strong pipeline baselines using RoBERTa and T5 only achieved 40.7 BLEU, highlighting the benefits of the unified approach.

Robustness in Low-Resource Scenarios

One of the most significant findings is UniGDD's performance when training data is scarce. When trained on only 1/32 of the available data, UniGDD-base outperformed the best pipeline baseline by over 20 absolute points in EM and 10 points in BLEU. This suggests that the unified generative approach is much more efficient at learning the underlying patterns of document-grounded dialogue than disconnected pipeline systems, which require high-quality labels for every stage to function effectively.

Human Evaluation

Beyond automated metrics, human evaluators participated in blind tests to compare UniGDD's outputs with those of pipeline systems. The evaluators consistently rated UniGDD’s responses as more relevant and informative, confirming that the model's unified nature leads to more coherent and helpful dialogue.

Conclusion

UniGDD demonstrates that merging knowledge identification and response generation into a single generative task effectively mitigates the problem of error propagation and captures the interdependence of these sub-tasks. Through the use of Prompt-Connected Multi-Task Learning and Linear Temperature Scheduling, the model achieves superior performance, particularly in data-scarce environments. This research suggests that unified architectures are a promising direction for complex, multi-step NLP tasks that have traditionally relied on brittle pipeline systems.

Summary updated

BibTeX
@inproceedings{gao2022unigdd,
  author = {Chang Gao and Wenxuan Zhang and Wai Lam},
  title = {{UniGDD: A Unified Generative Framework for Goal-Oriented Document-Grounded Dialogue}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  url = {https://openreview.net/forum?id=xMjsM6nsjO},
}
Figure