Abstract
Goal-oriented document-grounded dialogue systems typically follow a "find then talk" pipeline, consisting of separate Knowledge Identification (KI) and Response Generation (RG) modules. This decoupled approach often suffers from error propagation and fails to model the interdependence between identifying relevant information and formulating a response. We propose UniGDD, a unified generative framework that treats both sub-tasks as a single sequence-to-sequence problem. By integrating prompt-connected multi-task learning and a novel linear temperature scheduling mechanism for attention, UniGDD achieves state-of-the-art performance on the Doc2Dial benchmark. Notably, our framework exhibits superior robustness in data-scarce scenarios, outperforming pipeline baselines by significant margins.
Introduction
In goal-oriented dialogue, such as automated customer support, systems must consult external documents (e.g., policy manuals or FAQs) to provide accurate answers. The traditional approach treats this as a two-step pipeline: first, identifying the relevant snippet (KI), and second, generating a natural response (RG).
The primary drawback of this pipeline is error propagation: if the KI module selects an incorrect snippet, the RG module cannot produce a valid answer. Furthermore, these models ignore the synergy between the tasks; the context of a potential response can help pinpoint relevant knowledge, and vice versa. UniGDD addresses these issues by unifying KI and RG into a single generative process.
Methodology
UniGDD leverages a pre-trained T5 model to generate a concatenated sequence containing both the grounding knowledge and the final agent response.
Unified Generation and Multi-Task Learning
The core innovation lies in treating the problem as a sequence-to-sequence task where the target output is structured as <grounding> [Extracted Knowledge] <agent> [Generated Response]. To enhance the model's ability to distinguish between extraction and composition, we implement Prompt-Connected Multi-Task Learning (PCMTL). This involves three tasks guided by specific prompts:
- Main Task: Generate both grounding knowledge and the response.
- Auxiliary Task 1: Focus solely on generating grounding knowledge.
- Auxiliary Task 2: Focus solely on generating the response.
Linear Temperature Scheduling (LTS)
Supporting documents often contain significant noise. To help the model focus on relevant sections, we introduce Linear Temperature Scheduling (LTS) in the decoder's cross-attention mechanism. By gradually decreasing the softmax temperature during training, the attention distribution becomes increasingly "sharp," forcing the model to ignore irrelevant text and zero in on the "gold" knowledge.
Experimental Evaluation
We evaluated UniGDD on the Doc2Dial dataset, a large-scale benchmark featuring over 44,000 dialogue turns grounded in government service documents.
Performance and Robustness
UniGDD established new state-of-the-art results for both sub-tasks. In Knowledge Identification, it achieved an Exact Match (EM) of 66.9 and an F1 score of 77.5. In Response Generation, it reached 42.9 BLEU, significantly outperforming the DIALKI+BART pipeline (25.8 BLEU).
A key strength of the unified approach is its performance in low-resource settings. When trained on only 1/32 of the available data, UniGDD-base maintained a substantial lead over the best pipeline baselines, highlighting the efficiency of joint learning when data is scarce.
Conclusion
UniGDD demonstrates that unifying knowledge identification and response generation into a single generative framework effectively mitigates error propagation and leverages task interdependencies. While the model handles long documents well via LTS, future work may explore its scalability to massive technical manuals and its generalizability to other grounded dialogue domains. Human evaluations further confirm that UniGDD produces more relevant and informative responses compared to traditional pipeline architectures.
Summary updated