Enhancing knowledge base question answering with AMR-driven subgraph retrieval

Knowledge Base Question Answering (KBQA) aims to answer natural language questions using factual triples from a knowledge graph (KG). Mainstream Graph Neural Network (GNN)-based methods rely on subgraph retrieval to reduce reasoning complexity. However, existing retrieval techniques, which predominantly depend on shallow similarity matching between question text and KG relations, often fail to bridge the semantic gap between linguistic diversity and structured knowledge. This leads to inconsistent subgraph quality and poor interpretability, thereby hindering QA performance. To address this fundamental challenge, we propose a novel framework driven by Abstract Meaning Representation (AMR). Our method parses questions into AMR graphs that preserve core semantics, from which we extract canonical predicate-argument structures to serve as high-precision filters for pruning and refining an initially retrieved noisy subgraph. Furthermore, we leverage the AMR structure to inform the GNN's adjacency matrix and to generate highly interpretable reasoning chains. Experiments on the WebQSP and CWQ benchmarks demonstrate that our approach not only significantly improves subgraph quality and retrieval performance but also achieves competitive KBQA results.