Search papers, labs, and topics across Lattice.
This paper introduces SelfGraphRAG, a novel framework that generates synthetic question-answer pairs from knowledge graph structures to train query-conditioned graph retrievers, addressing the challenge of insufficient labeled data for supervised graph retrieval. By leveraging the relational information inherent in knowledge graphs, SelfGraphRAG captures multi-hop paths and local neighborhoods, enhancing the retrieval precision and downstream reasoning capabilities of large language models. Experimental results demonstrate that SelfGraphRAG outperforms traditional embedding-based methods, highlighting the potential of knowledge graph structures as a source of effective supervision in RAG systems.
Synthetic question generation from knowledge graphs boosts retrieval precision and reasoning performance, even in the absence of labeled data.
Retrieval-augmented generation (RAG) improves large language models by incorporating external knowledge without retraining, but existing methods often underuse the relational structure encoded in knowledge graphs. Graph-based RAG can capture entity relationships, yet supervised graph retrieval typically requires labeled question-answer data that may not be available for newly constructed graphs. We address this limitation with SelfGraphRAG, a framework that generates question-answer pairs directly from knowledge graph structure and uses them to train a query-conditioned graph retriever. The generated questions capture multi-hop paths and local neighborhoods, providing relational supervision without manual annotation. Experiments on multi-hop question answering and classification benchmarks show that SelfGraphRAG improves retrieval precision and downstream reasoning performance over embedding-based baselines. These results suggest that knowledge graph structure can provide useful supervision for training graph retrievers when labeled data are unavailable.
