RESUMO
In response to the limited capability of extracting semantic information in knowledge graph completion methods, we propose a model that combines spatial transformation and attention mechanisms (STAM) for knowledge graph embedding. Firstly, spatial transformation is applied to reorganize entity embeddings and relation embeddings, enabling increased interaction between entities and relations while preserving shallow information. Next, a two-dimensional convolutional neural network is utilized to extract complex latent information among entity relations. Simultaneously, a multi-scale channel attention mechanism is constructed to enhance the capture of local detailed features and global semantic features. Finally, the surface-level shallow information and latent information are fused to obtain feature embeddings with richer semantic expression. The link prediction results on the public datasets WN18RR, FB15K237 and Kinship demonstrate that STAM achieved improvements of 8.8%, 10.5% and 6.9% in the mean reciprocal rank (MRR) evaluation metric compared to ConvE, for the respective datasets. Furthermore, in the link prediction experiments on the hydraulic engineering dataset, STAM achieves better experimental results in terms of MRR, Hits@1, Hits@3 and Hits@10 evaluation metrics, demonstrating the effectiveness of the model in the task of hydraulic engineering knowledge graph completion.
