Improving Hyper-Network-Based Knowledge Graph Completion with Linear Transformation

Knowledge graph is a data structure with strong expression ability, which has been widely used in both academic and production fields, but most of existing knowledge graphs suffer from incompleteness, various methods like ConvE, Hyper and InteraceE have been proposed to tackle with this problem. Combining with hyper-network-based convolution and linear transformation of entities, a knowledge graph completion method named HyperL is proposed in this paper, HyperL not only inherits the excellent information fusion and expression ability of convolution transformation, but also enhanced the expression ability of entities with linear transformation. The proposed method is evaluated on three standard datasets including FB15k-237, WN18RR and YAGO3-10, and the performances are compared with other benchmark methods. Experimental results shows that the MRR of HyperL on FB15k-237, WN18RR and YAG03 can reach 0.360, 0.468, 0.552, respectively, which indicate that HyperL is and effective knowledge graph completion method.