A Query-Driven Graph Retrieval Framework with Adaptive Pruning for Multi-Hop Question Answering
Hao Wang, Lihang Feng, Tianyue Wang, Zhongyi Sun, He Li
Electronics
Problems Identified (4)
MHQA distributed evidence reasoning: Multi-hop question answering requires retrieving and reasoning over evidence distributed across multiple documents.
Flat RAG reasoning-chain weakness: Conventional flat retrieval-augmented generation approaches struggle to maintain coherent reasoning chains when implicit dependencies among entities and documents are involved.
MHQA distributed evidence reasoning: Multi-hop question answering requires retrieving and reasoning over evidence distributed across multiple documents.
Flat RAG reasoning-chain weakness: Conventional flat retrieval-augmented generation approaches struggle to maintain coherent reasoning chains when implicit dependencies among entities and documents are involved.
Proposed Solutions (5)
Query-driven dual-layer graph retrieval: The paper proposes a query-driven dual-layer graph retrieval framework for multi-hop question answering.
Heterogeneous graph joint entity-relation retrieval: The framework uses a unified heterogeneous graph of entities, relations, and supporting texts and dynamically constructs candidate subgraphs via joint entity and relation retrieval plus lexical retrieval signals.
Contrastive structural path scoring: Reasoning paths are refined by combining structural strength modeling with contrastive learning-based path scoring.
Query-adaptive evidence pruning: An adaptive pruning strategy regulates evidence scale according to query complexity and path score distributions.
Query-driven dual-layer graph retrieval: The paper proposes a query-driven dual-layer graph retrieval framework for multi-hop question answering.
Results (3)
Higher MHQA accuracy than baselines:
Improved complex multi-hop performance:
Structured query-adaptive evidence importance:
Research Domain
Multi-hop question answering / graph retrieval / retrieval-augmented generation