Knowledge graph completion aims to fill in missing information in knowledge graphs by addressing challenges such as data sparsity and complex relations. The data sparsity can be addressed by introducing richer contextual semantic features, while relational inference features can help infer complex relations. Current knowledge graph completion methods cannot simultaneously capture these two features. In this paper, we propose REKGC, a novel model for learning RE-coupled REpresentations for Knowledge Graph Completion. REKGC separates the learning of contextual semantic features and relational inference features, and then re-couples them to generate new representations. REKGC includes a contextual semantics module, a relational inference module, and a feature re-coupling module. REKGC ensures effective learning of both advantageous features. Experimental results demonstrate that by capturing and re-coupling two types of features, REKGC can significantly improve the performance of knowledge graph completion.