Multi-Scale Convolutional Network for Knowledge Graph Completion

Knowledge graphs reflect various facts in the actual world through formal representations and are interconnected through topological structures. However, most knowledge graphs are incomplete. Knowledge graph completion infers unknown or implie
$d$
facts based on existing facts. ConvKB is a typical CNN-based model that captures the translation properties between embedding units of the same dimension through one-dimensional convolution. Since one-dimensional convolution can only receive restricted semantic information,
$w$
e use multi-scale convolution for capturing feature interactions between units of different dimensions. To further increase the number of interactions, we employ circular convolution. In our research, we introduce a simple and effective embedding model, named MCNKC, for knowledge graph completion. We demonstrate that MCNKC achieves better results than ConvKB on FB15K-237 and WN18RR.