Temporal Knowledge Graph Completion Based on Entity Multi-encoding and Temporal Awareness

Temporal knowledge graph completion (TKGC) has attracted significant attention. Previous works have led to great advances. However, most of methods still have the issue of one-sided acquisition of entity features and insufficient utilization of temporal facts from recent time steps. In this paper, we propose a model based on entity multi-encoding and temporal awareness (EmtE) to address these challenges. Entity multi-encoding aims to introduce entity slice feature encoding at the specific timestamp and time-aware entity feature which integrate time embeddings and static features of entity, thus deeply exploring the pluralistic information of entities in the temporal knowledge graph. Simultaneously, to better perceive the timing information, the model utilizes GRU in conjunction with a weight decay strategy to capture the temporal features of the historical facts. Experiments on the ICEWS14, ICEWS05-15, and GDELT datasets demonstrate that our model generally outperforms baseline models in terms of MRR, Hits@1, Hits@3 and Hits@10, effectively improving the performance of temporal knowledge graph completion.