ConvMB: Improving Convolution-Based Knowledge Graph Embeddings by Adopting Multi-Branch 3D Convolution Filters

Xiaobo Guo, Fali Wang, Neng Gao, Zeyi Liu, Kai Liu

Abstract: In this paper, we propose a novel embedding model, named ConvMB, for knowledge base completion. ConvE captures the potential semantic associations of the knowledge base by learning the local feature interactions among the same entry of the embeddings. ConvKB further makes full use of the global feature interactions among the same entries of the embeddings to get a better performance. CapsE gains improvements by considering the deep correlations of the feature interaction. To the best of our knowledge, the deep associations of the feature interactions have not been considered further by existing convolution-based knowledge graph embedding models from the point of acrossing entries. Our model ConvMB fills this gap by adopting a multi-branch 3D convolutional structure and outperform state-of-the-art models. In ConvMB, the embeddings of the head entity, relation and tail entity of a triplet are firstly reshaped as a 2D matrix respectively. Then these three 2D matrixes are stacked into a 3D matrix, which is fed into the convolutional layer. Multiple filters of different shapes in different branches are operated on the 3D matrix to generate different feature maps. The shape of filters in the first layer is 1 × 1 × 3, which can be used to capture the feature interactions among the same entries and can get the corresponding feature maps. The shape of filters in layers after the first one include 2 × 2 × τ, 3 × 3 × τ, 4 × 4 × τ, 5 × 5 × τ, τ is number of filters in each layer of each branch. These filters can further act on the stacked feature maps to get deep associations across different entries. The final feature maps produced by each branch are concatenated into a single feature vector representing the input triplet. The single feature vector is multiplied with a weight vector via a dot product to predict whether the triplet is correct or not. Experiments show that link prediction performance is significant improved by ConvMB, relative to previous SOTA embedding models on two benchmark datasets WN18RR and NELL-995.