Enhancing knowledge graph embedding with structure and semantic features

Knowledge graph embedding converts knowledge graphs based on symbolic representations into low-dimensional vectors. Effective knowledge graph embedding methods are key to ensuring downstream tasks. Some studies have shown significant performance differences among various knowledge graph embedding models on different datasets. They attribute this issue to the insufficient representation ability of the models. However, what representation ability knowledge graph embedding models possess is still unknown. Therefore, this paper first selects three representative models for analysis: translation and rotation models in distance models, and the Bert model in neural network models. Based on the analysis results, it can be concluded that the translation model focuses on clustering features, the rotation model focuses on hierarchy features, and the Bert model focuses on word co-occurrence features. This paper categorize clustering and hierarchy as structure features, and word co-occurrence as semantic features. Furthermore, a model that solely focuses on a single feature will lead to a lack of accuracy and generality, making it challenging for the model to be applicable to modern large-scale knowledge graphs. Therefore, this paper proposes an ensemble model with structure and semantic features for knowledge graph embedding. Specifically, the ensemble model includes a structure part and a semantic part. The structure part consists of three models: translation, rotation and cross. Translation and rotation models serve as basic feature extraction, while the cross model enhances the interaction between them. The semantic part is built based on Bert and integrated with the structure part after fine-tuning. In addition, this paper also introduces a frequency model to mitigate the training imbalance caused by differences in entity frequencies. Finally, we verify the effectiveness of the model through link prediction. Experiments show that the ensemble model has achieved improvement on FB15k-237 and YAGO3-10, and also has good performance on WN18RR, proving the effectiveness of the model.