Pre-trained Model Enhanced Contrastive Knowledge Graph Completion

Current knowledge graphs, structured representations of real-world facts, no matter how large, typically contain only a portion of all possible facts. Link prediction, inferring missing facts based on existing ones, is one of the most critical tasks of knowledge graph completion (KGC). We address the KGC scientific issue, propose a pre-trained model enhanced contrastive KGC framework, and design and implement a corresponding KGC algorithm. On one hand, the pre-trained model enhanced contrastive KGC algorithm uses the precise semantic representation advantage of pre-trained models to embed entity descriptions, and based on these embedding vectors, entity clustering is carried out; On the other hand, entities are selected from within the clustered entity cluster to construct negative samples, and together with the original positive samples, the unsupervised advantages of contrastive learning are fully utilized to achieve efficient comparative KGC. Experimental results on two publicly available da-tasets WN18RR and FB15k237 show that the performance of our algorithm is not only better than that of most existing KGC algorithms, but also on the FB15k237 dataset, some metrics of our algorithm are over 2 percentage points higher than that of the currently optimal SimKGC algorithm. We conclude that selecting entities from the entity set with stronger semantic cohesion to build a negative sample set can make contrastive learning pay more attention to difficult negative samples similar to positive samples. By increasing the model loss during the training process, the fine-grained feature discrimination ability can be improved, thereby improving the performance of contrastive KGC.