A Focus-Relation Alignment-Based Dynamic State Representation Method for Multi-Hop Knowledge Graph Question Answering

Abstract

Guided by state information in the environment, reinforcement learning-based multi-hop knowledge graph question answering (KGQA) aims to discover reasoning paths that lead to the correct answer entity. Existing state observation methods typically encode the multi-hop question as a static vector, overlooking the fact that both the semantics of the question and its reasoning focus evolve dynamically across timesteps. This limitation is further compounded by the semantic gap between natural language queries and KG relations, which makes it difficult to align the evolving question focus with the correct reasoning relation and often leads to ineffective exploration and semantic drift. To address these challenges, we introduce a Focus-Relation Alignment (FRA) framework that provides a dynamic and adaptive state representation for multi-hop KGQA. The key idea is to bridge the evolving question focus with candidate KG relations, enabling the agent to capture the temporal progression of the query while grounding it in semantically consistent relations. By synchronizing the dynamic representation of multi-hop questions with focus-enhanced relation signals, our method empowers the agent to perceive not only where to explore next but also why that path is semantically meaningful. This high-level alignment mechanism effectively narrows the semantic gap, reduces spurious exploration, and yields more purposeful reasoning trajectories. Extensive experiments on three widely used datasets show that FRA achieves notable improvements in reasoning accuracy on multiple benchmarks. In particular, FRA outperforms prior methods by 2.55 points on Hits@1 on the PQ $\mathbf{2 H}$ dataset, demonstrating its effectiveness and potential over existing approaches.