A Dynamic Convolutional Network-Based Model for Knowledge Graph Completion

Knowledge graph embedding can learn low-rank vector representations for knowledge entities and relations, has been a main research topic completion. Several recent works suggest that convolutional neural network (CNN)-based models capture interactions between head relation embeddings, hence perform well on However, previous have ignored the different contributions of interaction features to experimental results. In this paper, we propose novel model named DyConvNE base Our uses dynamic convolution kernel because assign weights varying importance features. We also new method negative sampling, which mines hard samples as additional training. performed experiments data sets WN18RR FB15k-237, results show our is better than several other benchmark algorithms addition, used test when predicting Hits@1 values specific-relationship testing. This gives about 2% relative improvement over do not use in terms Hits@1.