E$^{2}$GraphRAG: Dual-Layer Entity--Event Indexing and Retrieval for Graph-Augmented Generation

ACL ARR 2026 January Submission4983 Authors

Keywords: GraphRAG, knowledge graph, RAG

Abstract: Hierarchical graph-based retrieval-augmented generation(RAG) methods represent knowledge at different granularities across each layer, forming a structure with more natural semantic connections. However, this approach still faces two challenges: the multi-level structure is complex and grows increasingly abstract in content, while direct retrieval matching often fails to adequately decompose queries based on community structures. In this work, We introduce a hierarchical entity–event knowledge indexing pipeline that injects event nodes extracted from entity-linked evidence, producing a compact yet semantically rich graph for RAG.We propose an event-based query decomposition strategy that couples dense retrieval–based query grounding with hierarchical graph retrieval, including local entity alignment and global event retrieval for coherent evidence organization.Extensive experiments on UltraDomain show that E$^{2}$GraphRAG consistently achieves win rates ranging from 50.0\% to 99.2\% against NaiveRAG, GraphRAG, LightRAG, and HiRAG. Ablations validate the contributions of E$^{2}$GraphRAG.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: Dialogue and Interactive Systems

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 4983