Enhancing the convolution-based knowledge graph embeddings by increasing dimension-wise interactions

Knowledge graph embedding learns distributed low-dimensional representations for the elements in knowledge graphs, so that knowledge can be conveniently integrated into various tasks and smart systems. Recently, convolutional neural network has been introduced to embedding technique and obtained impressive achievements in link prediction task. ConvKB, a recently proposed method, captured the global dimension-wise interactions in facts with the convolutional filters. However, ConvKB failed to learn the local interactions between the entity and relation embedding. Moreover, rich interactions among feature maps are neglected in the existing convolutional embedding models. In this paper, based on ConvKB, we propose ConvD which models the local relationships in facts and integrates the cross-channel information based on the dimension-wise interactions to further improve the performance. From the experimental results, ConvD obtains scores that are 96% and 5% better than ConvKB on MRR and Hits@10 in the link prediction task. Furthermore, ConvD surpassed state-of-the-art baselines on WN18RR and achieved competitive results on FB15k-237 respectively.