Reproduce “Beyond Chunk-Local Extraction: Cross-Chunk Graph Augmentation for GraphRAG” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaming Zhang, Yibo Zhao, Jing Yu, Jianxiang Yu, Xiang Li
CROSSAUG addresses the issue of incomplete GraphRAG indices by developing a GNN-guided, self-supervised approach to identify and incorporate cross-chunk relational structures from long documents. This method consistently improves question answering performance across multiple GraphRAG frameworks and benchmarks, such as boosting HippoRAG2's F1 score by +1.33 on MuSiQue and +1.09 on LiteraryQA.
Key insights
- Graph Neural Networks can effectively identify structurally incomplete regions within a knowledge graph, enabling targeted and efficient utilization of expensive LLM computation for relation extraction.
- Self-supervised graph corruption, by masking edges and deleting nodes, provides a scalable method to train GNNs for detecting graph missingness without requiring extensive manual annotation.
- Combining GNNs for structural guidance with LLMs for semantic extraction allows GraphRAG systems to overcome the limitations of local chunk processing and capture complex relations spanning multiple text segments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.