Temporal Knowledge Graph Completion Based on Product Space and Contrastive Learning of Commonsense

Abstract Static knowledge graph completion (KGC) has made significant progress in the field of artificial intelligence. However, knowledge is time-sensitive, thus the introduction of Temporal Knowledge Graph Completion (TKGC) is necessary to accurately reflect its dynamic changes. Current TKGC methods often overlook the semantic similarity between entities and entity-relation pairs, which leads to a loss of rich semantic information in temporal knowledge graphs. Furthermore, using a single embedding space limits these methods' ability to model various temporal patterns and rich semantic information. To address this issue, we propose a commonsense-based contrastive learning mechanism and a product space approach. Specifically, we incorporate commonsense information to learn time-invariant event representations and integrate relationship and temporal information into different embedding spaces to comprehensively capture semantic changes. Through these methods, we can effectively compare representations of the same entity at different time points, reducing the semantic distance between entities and entity-relation pairs, thereby significantly improving the performance of the completion task. Experimental results show that our method significantly enhances the performance of TKGC on multiple public datasets.