This paper presents a discourse-aware hierarchical framework leveraging Rhetorical Structure Theory (RST) to overcome the limitations of existing approaches, which fail to capture discourse structures that facilitate human comprehension in long-document question answering systems. This framework transforms discourse trees into sentence-level representations and connects structural and semantic information using LLM-enhanced node representations. Its core innovations include three key elements: long-document-specific discourse parsing, LLM-based discourse relation node enhancement, and structure-based hierarchical retrieval. Experiments on the QASPER, Quality, and NarrativeQA datasets demonstrate consistent performance improvements over existing approaches, demonstrating that discourse structure integration significantly enhances question answering performance across a wide range of document types.
