Beyond Chunking: Discourse-Aware Hierarchical Retrieval for Long Document Question Answering

Researchers have introduced a novel discourse-aware hierarchical framework designed to enhance long-document question answering systems. Unlike traditional methods that treat texts as flat sequences or rely on heuristic chunking, this approach leverages Rhetorical Structure Theory (RST) to capture the natural discourse structures guiding human comprehension. The framework features three key innovations: language-universal discourse parsing for lengthy documents, Large Language Model (LLM)-enhanced node representations to bridge structural and semantic information, and structure-guided hierarchical retrieval. By converting discourse trees into sentence-level representations, the system effectively integrates structural context with semantic meaning. Extensive experiments across four diverse datasets demonstrate consistent performance improvements over existing approaches in multiple genres and languages. Furthermore, the framework exhibits strong robustness across various document types and linguistic settings, addressing significant limitations in current information retrieval technologies. This development marks a significant step forward in enabling AI systems to better understand and process complex, long-form textual data by mimicking human-like structural comprehension.