Multi-Hop Reasoning of Temporal Knowledge Graph Based on Reinforcement Learning

The reasoning research of temporal knowledge graph has attracted many researchers' attention in recent years. Among the existing reasoning methods, scholars mainly focus on the research of multi-hop reasoning tasks under future timestamps. Multi-hop reasoning refers to making multiple jumps in the temporal knowledge graph to reason new information by using the relation between entities and timestamp. In the multi-hop reasoning task, it is necessary to consider the semantics of facts, which is often reflected in the semantic relation between entities or between entities and relations. Additionally, with the continuous evolution of timestamps, the historical state under different timestamps will have different effects on the reasoning results. Therefore, the current research on multi-hop reasoning of temporal knowledge graph faces two challenges: How to model the entity semantic hierarchy in temporal knowledge graphs. How to enhance the modeling ability of temporal sequence of historical states, so that we can obtain more accurate temporal evolutional information between entities. For the sake of addressing the above-mentioned two challenges, we put forward a neoteric reinforcement learning reasoning model. Specifically, firstly, we embed entities, relations and timestamps into polar coordinates by means of representation learning, and introduce the obtained entity embedding vectors into the knowledge graph reasoning model based on reinforcement learning, which can provide coded representations of the entities with hierarchical relations and help to provide more semantic information between facts. Then, we introduce a component called temporal information integration module based on self-attention network. It contribute to comprehend and employ temporal evolutionary patterns for more accurate reasoning. Experimental results prove that our proposed model performs better than other state-of-the-art methods on four public datasets.