Hybrid-TE: Hybrid Translation-Based Temporal Knowledge Graph Embedding

Knowledge graph embedding has become a promising method for knowledge graph completion. In this work, we propose Hybrid-TE, a hybrid translation-based temporal knowledge graph embedding, which combines two translation models TransD and HyTE for modeling both temporal and multirelational facts. Benefiting from two underlying models, Hybrid-TE first builds entity and relation embeddings in separate vector space for modelling multi-relational facts, and then explicitly learns time information by translational embedding on timespecific hyperplanes. We observe that a simple combination of two models does not lead to a satisfactory predictive precision. We therefore propose to project a triplet to all time-specific hyperplanes on which it is temporally valid. Besides, we also explore extra negative relation samplings that differ from positive samplings in relations. We conduct extensive experiments with real datasets on link prediction, relation prediction and temporal scope prediction. Experiments show significant improvements over previous time-insensitive or time-aware models.