A Multi-Scale Adaptive Retrieval-Augmented Generation Framework with Semantic and Structural Knowledge Graphs

Adaptive retrieval context range: Existing RAG methods struggle to determine an appropriate retrieval context range, where overly long contexts add redundancy and overly short contexts miss facts.

Complex multi-hop evidence integration: Existing RAG methods have limited performance in complex scenarios requiring multi-hop reasoning and cross-paragraph evidence integration.

Adaptive retrieval context range: Existing RAG methods struggle to determine an appropriate retrieval context range, where overly long contexts add redundancy and overly short contexts miss facts.

Complex multi-hop evidence integration: Existing RAG methods have limited performance in complex scenarios requiring multi-hop reasoning and cross-paragraph evidence integration.

Multi-scale adaptive RAG with semantic and structural KGs: The paper proposes a multi-scale adaptive retrieval-enhanced generation framework based on both semantic and structural knowledge graphs.

Granularity-aware structural KG slicing: A structural knowledge graph centered on chapters, paragraphs, and sentences supports flexible adjustable context slicing.

Entity-triple semantic KG cross-connections: A semantic knowledge graph models entities and relationships as triples and uses entity cross-connections to associate evidence across paragraphs.

Multi-scale adaptive RAG with semantic and structural KGs: The paper proposes a multi-scale adaptive retrieval-enhanced generation framework based on both semantic and structural knowledge graphs.

Granularity-aware structural KG slicing: A structural knowledge graph centered on chapters, paragraphs, and sentences supports flexible adjustable context slicing.

Improved HotpotQA retrieval and generation quality:

Effective for complex multi-hop QA:

Improved HotpotQA retrieval and generation quality: