关键词: 时序知识图谱, 分级知识嵌入, 知识推理, 强化学习, 奖励塑形

Abstract: In response to the issues in temporal knowledge graph reasoning methods that fail to adequately capture semantic dependencies, temporal evolution information, and lack interpretability, a temporal knowledge graph reasoning method based on hierarchical knowledge embedding and reinforcement learning is proposed, named THKERL. THKERL consists of two key components： the hierarchical knowledge embedding model (HKEM) and the reinforcement learning reasoning model (RLRM). HKEM obtains more accurate knowledge graph feature representations through two levels of knowledge embedding： the subgraph level aims to model the semantic dependencies between concurrent facts in each knowledge graph, while the global graph level is primarily used to capture the dynamic feature information of entities as they evolve over time. On this basis, RLRM employs reinforcement learning, introducing a weighted action scoring mechanism to design the policy network, fully considering the relationship between query questions and reasoning paths for reward shaping, to achieve more reliable knowledge reasoning. To validate the effectiveness of the THKERL method, experiments are conducted on datasets such as ICEWS14, and the experimental results are compared and analyzed with mainstream temporal knowledge graph reasoning methods like TiTer. The experimental results indicate that THKERL achieves an average improvement of over 5.9 percentage points in Hits@[k] and over 6.8 percentage points in MRR in entity prediction tasks.

Key words: temporal knowledge graph, hierarchical knowledge embedding, knowledge reasoning, reinforcement learning, reward shaping