Title: Structural Frameworks Enable Retrieval, Reranking, and Generation

Abstract: Document-grounded dialogue systems (DGDS) rely on external textual knowledge to address user queries within specific domains. Conventional approaches generally fragment documents into isolated passages for the purposes of retrieval and response synthesis. This methodology fails to adequately exploit internal document structures and often supplies insufficient contextual information for accurate knowledge selection and answer formulation. To resolve these limitations, this study introduces SF-Re2G. The proposed method first enhances passage representation by contrasting it with other segments from the same section, thereby boosting retrieval accuracy. Second, it employs a structure-aware reranker that capitalizes on the observation that multiple relevant passages for a single dialog turn are typically located in close proximity. Specifically, retrieved candidates are organized into subgraphs based on document hierarchy, allowing the reranker to adjust candidate scores by incorporating group-level data. Finally, the selected passages inform the response generation process, leveraging subgraph context to improve output quality. Evaluations on two DGDS datasets confirm the efficacy of our approach in both Chinese and English.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC