Although knowledge graph can enhance the efficiency of knowledge understanding and application, its inherent accuracy and completeness problems make knowledge graph complementation a current research focus. In this paper, the Trans H algorithm is improved by fusing semantic information, simplifying the triad by constructing an information hyperplane, and introducing BERT word vectors, which effectively improves the training efficiency. Adopting the attention mechanism proposed by previous researchers, the semantic information and the parameter vectors of the original model are fused and inputted into the Attention structure, and the attention scores of the semantic information of the triad are calculated. In order to evaluate the performance of the joint model SI-KGC, the dataset and experimental environment are constructed, hyperparameters are set, and the evaluation results are obtained after experiments such as ternary classification. The SI-KGC model constructed in this paper improves the accuracy by 3.12% compared to TransD. Meanwhile, it has high accuracy in the ternary group classification task, with an average accuracy of 83.8% and 74.75% on the two datasets, respectively, with higher classification accuracy and superior performance. Comparing the performance of TransE and SI-KGC on the FB15K dataset, the performance gap between the two models is between 0.003 and 0.011, and SI-KGC has a complete modeling capability, which has the effect of improving the model performance.
