Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Disco-RAG: Discourse-Aware Retrieval-Augmented Generation

Dongqi Liu, Hang Ding, Qiming Feng, Jian Li, Xurong Xie, Zhucun Xue, Chengjie Wang, Jiangning Zhang, Yabiao Wang

TL;DR

Disco-RAG addresses the flat, unstructured handling of retrieved evidence in standard RAG by injecting discourse structure into the generation process. It builds intra-chunk RST trees to capture local coherence and inter-chunk rhetorical graphs to model cross-passage relations, complemented by a discourse-driven planning module that yields a generation blueprint. Empirical results across Loong, ASQA, and SciNews show consistent, significant gains over strong RAG baselines and previous SOTA methods without fine-tuning, with ablations confirming the critical role of both intra- and inter-chunk discourse components. The work demonstrates that explicit discourse modeling improves reasoning, coherence, and factuality in knowledge-intensive generation tasks, highlighting discourse structure as a key lever for advancing RAG systems.

Abstract

Retrieval-Augmented Generation (RAG) has emerged as an important means of enhancing the performance of large language models (LLMs) in knowledge-intensive tasks. However, most existing RAG strategies treat retrieved passages in a flat and unstructured way, which prevents the model from capturing structural cues and constrains its ability to synthesize knowledge from dispersed evidence across documents. To overcome these limitations, we propose Disco-RAG, a discourse-aware framework that explicitly injects discourse signals into the generation process. Our method constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence. These structures are jointly integrated into a planning blueprint that conditions the generation. Experiments on question answering and long-document summarization benchmarks show the efficacy of our approach. Disco-RAG achieves state-of-the-art results on the benchmarks without fine-tuning. These findings underscore the important role of discourse structure in advancing RAG systems.

Disco-RAG: Discourse-Aware Retrieval-Augmented Generation

TL;DR

Disco-RAG addresses the flat, unstructured handling of retrieved evidence in standard RAG by injecting discourse structure into the generation process. It builds intra-chunk RST trees to capture local coherence and inter-chunk rhetorical graphs to model cross-passage relations, complemented by a discourse-driven planning module that yields a generation blueprint. Empirical results across Loong, ASQA, and SciNews show consistent, significant gains over strong RAG baselines and previous SOTA methods without fine-tuning, with ablations confirming the critical role of both intra- and inter-chunk discourse components. The work demonstrates that explicit discourse modeling improves reasoning, coherence, and factuality in knowledge-intensive generation tasks, highlighting discourse structure as a key lever for advancing RAG systems.

Abstract

Retrieval-Augmented Generation (RAG) has emerged as an important means of enhancing the performance of large language models (LLMs) in knowledge-intensive tasks. However, most existing RAG strategies treat retrieved passages in a flat and unstructured way, which prevents the model from capturing structural cues and constrains its ability to synthesize knowledge from dispersed evidence across documents. To overcome these limitations, we propose Disco-RAG, a discourse-aware framework that explicitly injects discourse signals into the generation process. Our method constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence. These structures are jointly integrated into a planning blueprint that conditions the generation. Experiments on question answering and long-document summarization benchmarks show the efficacy of our approach. Disco-RAG achieves state-of-the-art results on the benchmarks without fine-tuning. These findings underscore the important role of discourse structure in advancing RAG systems.
Paper Structure (69 sections, 6 equations, 19 figures, 11 tables)

This paper contains 69 sections, 6 equations, 19 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Structure-Aware Retrieval-Augmented Generation
  4. Rhetorical Structure Theory for Text Generation
  5. Proposed Method
  6. Method Overview.
  7. Intra-Chunk RST Tree.
  8. Inter-Chunk Rhetorical Graph.
  9. Discourse-Driven Planning.
  10. Discourse-Guided RAG.
  11. Experimental Settings
  12. Evaluation Datasets.
  13. Automatic Metrics.
  14. Implementation Details.
  15. Selected Baselines.
  16. ...and 54 more sections

Figures (19)

  • Figure 1: Comparison between standard RAG and Disco-RAG. While standard RAG retrieves isolated chunks without structural links, Disco-RAG organizes evidence into discourse structures (trees & graphs). Here, S denotes Satellite (the supplementary part), and N denotes Nucleus (the core part).
  • Figure 2: The Disco-RAG pipeline: Starting from passage retrieval (providing context), then intra-chunk RST tree parsing (capturing local discourse), inter-chunk rhetorical graph construction (modeling global discourse), rhetorical planning (blueprint generation), and answer generation (producing the final output).
  • Figure 3: Performance comparison under varying chunk size (a), Top-$k$ value (b), and retrieval noise level (c).
  • Figure 4: Effect of structural perturbations on performance. Panels (a), (b), and (c) correspond to intra-chunk RST trees, inter-chunk rhetorical graphs, and discourse-aware plans, respectively. Each perturbation involves randomly altering or removing the relevant elements.
  • Figure 5: Case study comparing standard RAG and Disco-RAG on the query "When did The Lion King debut on Broadway?". Our method captures both the preview and official opening as well as the later relocation, while standard RAG gives only a vague year-based answer.
  • ...and 14 more figures