New HyCE-RAG Improves Multi-hop Question Answering with Hypergraphs.

Traditional Retrieval-Augmented Generation (RAG) systems often struggle with complex multi-hop questions because they primarily rely on semantic similarity, failing to capture intricate relationships between facts across multiple documents. Graph-based RAGs offer an improvement but are limited by pairwise connections, which cannot fully represent higher-order associations involving multiple entities and contexts. A new framework, HyCE-RAG, addresses this by employing hypergraphs to model these complex relationships. It constructs a query-aware evidence hypergraph, where entities, relations, and contextual evidence are linked through hyperedges. This allows for confidence propagation and guided evidence assembly, selecting and ranking evidence paths based on multiple factors like semantic relevance, entity connectivity, and reliability. By providing language models with these structured evidence chains, HyCE-RAG significantly enhances the faithfulness and interpretability of reasoning. Experimental results on several benchmarks demonstrate its superior performance over existing RAG and graph-based RAG methods in terms of answer accuracy, context relevance, and faithfulness, suggesting hypergraph-based evidence organization is a promising direction for advanced question answering.