SymPath: Symmetric Path Retrieval with Multi-Grained Semantic Alignment for Knowledge Graph Question Answering

Multi-hop knowledge graph question answering (KGQA) faces practical deployment hurdles: large language models suffer from high computational overhead and hallucination risks, while graph-based methods rely on entity-centric subgraphs that adapt poorly to dynamic graphs. To address these issues, we propose SymPath, a lightweight symmetric path retrieval method that decouples neural semantic retrieval from symbolic graph execution. Leveraging architectural symmetry, SymPath employs a parameter-shared 22-million-parameter dual-tower encoder that maps natural-language questions and candidate relation paths into a unified, symmetric semantic space, eliminating the parameter redundancy and representation drift of asymmetric encoders. A multi-grained alignment mechanism combines global mean pooling with token-level MaxSim to accurately discriminate multi-hop paths, which are then executed symbolically without free-form generation. Consequently, the model adapts to updated graphs without retraining. Under the gold topic entity protocol, experiments on WebQSP and ComplexWebQuestions show that SymPath achieves 67.8% and 60.6% F1, respectively, and reaches 61.7% Hits@1 under random entity sampling, outperforming a 110-million-parameter baseline. Furthermore, SymPath requires only ∼1 GB GPU memory and achieves low per-batch latency (300 ms on GPU; 530 ms on CPU). These results demonstrate a competitive accuracy–efficiency trade-off for resource-constrained KGQA.