关键词: 知识图谱问答, 分层强化学习, 复杂约束问题, 注意力机制, 维度灾难

Abstract: To address the issues of ignored constraint information and the curse of dimensionality in long reasoning paths for knowledge graph question answering (KGQA), a hierarchical reinforcement learning-based constrained KGQA model (KGQA-HRL) was proposed. First, the concept of hierarchical reinforcement learning was deeply integrated by decomposing triples in the knowledge graph, and a high-level policy along with a low-level policy was designed to mitigate the dimensionality challenge in reasoning paths. Then, to improve the accuracy of path selection, an attention-based action selection strategy was introduced, combined with an entity selection strategy incorporating constraint information, effectively narrowing the search space for reasoning. Additionally, a question update mechanism was embedded between the action selection and entity selection strategies, enabling secondary updates of the question at each reasoning step. Finally, a constraint set was constructed in the entity selection strategy, and constraint scores were calculated to incorporate constraint information from the question, enhancing the precision of entity selection. Experiments were conducted on four KGQA benchmark datasets to evaluate the performance of KGQA-HRL. The results demonstrate that KGQA-HRL achieves state-of-the-art performance across all datasets, with an average improvement of 3.1% over the previous best model (COPAR). Notably, it exhibits outstanding performance in complex three-hop query tasks (3.0% improvement on the PQ dataset and 2.1% on MetaQA), validating its superior reasoning capability.

Key words: knowledge graph question answering, hierarchical reinforcement learning, complex constrained problems, attention mechanism, curse of dimensionality