KHL-RAG: Enhancing Retrieval-Augmented Generation via Dual-Graph Mechanism based on Rhetorical Structure Theory

ACL ARR 2026 January Submission5513 Authors

Keywords: Retrieval Augmented Generation, Heterogeneous Logical Paragraph Graph

Abstract: Retrieval-Augmented Generation has become a dominant paradigm for improving the reasoning reliability of large language models by grounding generation in external knowledge. However, existing approaches remain limited by unstable vector-based retrieval and the restricted coverage of entity-centric knowledge graphs. These limitations arise from the absence of paragraph-level discourse modeling. We propose KHL-RAG, a dual-graph RAG framework that integrates a knowledge graph with a heterogeneous logical paragraph graph grounded in Rhetorical Structure Theory. By explicitly modeling implicit discourse relations and adopting a Dual-Semantic Vector Filtering and query-aware aggregation strategy, KHL-RAG improves retrieval robustness and contextual completeness. Extensive experiments demonstrate that KHL-RAG consistently reduces hallucinations in complex reasoning tasks, achieving up to a 3.82\% improvement in answer correctness over state-of-the-art RAG baselines.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: Language Modeling,Generation

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 5513