ResConvE: Deeper Convolution-Based Knowledge Graph Embeddings

Yongxu Long, Zihan Qiu, Dongyang Zheng, Zhengyang Wu, Jianguo Li, Yong Tang

Abstract: Link prediction on knowledge graphs (KGs) is an effective way to address their incompleteness. ConvE and InteractE have introduced CNN to this task and achieved excellent performance, but their model uses only a single 2D convolutional layer. Instead, we think that the network should go deeper. In this case, we propose the ResConvE model, which takes reference from the application of residual networks in computer vision, and deepens the neural network, and applies a skip connection to alleviate the gradient explosion and gradient disappearance caused by the deepening of the network layers. We also introduce the SKG-course dataset from Scholat for experiments. Through extensive experiments, we find that ResConvE performs well on some datasets, which proves that the idea of this method has better performance than baselines. Moreover, we also design controlled experiments setting different depths of ResConvE on FB15k and SKG-course to demonstrate that deepening the number of network layers within a certain range does help in performance improvement on different datasets.