JointRel: Joint semantic embedding with relational message passing for knowledge graph completion

Knowledge graph completion (KGC) aims to predict missing tail or head entities by learning graph information. However, most studies focus on the accurate expression of entity semantics, and ignore edge features. This results in a lack of balance in semantic expression of entities and relations, and limits the application of knowledge graphs. To this end, we propose a Joint semantic embedding method with Relational message passing (JointRel) by a dual-channel graph augmentation network, to explicitly learn both node and edge features. Specifically, in entity embedding learning, the neighbor information of the central node is learned by node-level and edge-level graph learning, and then the updated representation is obtained by aggregating these two kinds of neighbor information. Furthermore, in relational embedding learning, the relational context of entity pairs is taken as prior knowledge, and relational paths between them are combined by graph attention to obtain an updated representation of edge features. Experiments on four KGC datasets indicate JointRel's superiority, with MRR improvements of 4.2%, 0.2%, 3.4% and 24.8% over SOTA methods. JointRel provides a more balanced semantic representation framework for entities and relations, enhancing the stability and robustness of KGC, which is beneficial for downstream applications like question answering and recommendation systems.