[5G3-OS-37b] OS-37

3:00 PM - 3:15 PM JST(6:00 AM - 6:15 AM UTC)

[5G3-OS-37b-05] Lightweight GraphRAG Retrieval for Unstructured Text via Entity Co-occurrence Graphs

〇Yudai Yamazaki¹, Anna Goshima¹, Shunichi Tahara¹, Kazuaki Furumai¹, Tatsuya Konishi¹, Kazushi Ikeda¹ (1. KDDI Research, Inc.)

Keywords:

AI,RAG,GraphRAG

Recent work on RAG has shown that graph-based retrieval methods (GraphRAG) can be effective for complex question answering. However, many GraphRAG approaches rely on external knowledge graphs or require costly preprocessing such as exploratory LLM calls and summary generation, which makes them expensive to apply to custom corpora. In this paper, we propose a lightweight GraphRAG retrieval method that does not depend on external knowledge graphs or summarization. Our method, Bubble Graph Preference (BGP), constructs a document--entity bipartite graph from entity co-occurrences extracted from an unstructured corpus, and leverages graph-based retrieval only when the vector-retrieval scores are uncertain. Experiments on three multi-hop QA benchmarks (MuSiQue, 2WikiMultihopQA, and HotpotQA) show that integrating vector retrieval with our lightweight graph retrieval improves the retrieval of gold documents. Compared with a standard vector-based RAG baseline, the proposed method increases gold-document coverage by up to 10.06 points and improves Recall@10 by up to 1.00 point.