A Knowledge Graph Link Prediction Model with Combined 1D and 2D Convolutional Embeddings

Ming Shi,Jing Zhao
DOI: https://doi.org/10.1109/DSAA54385.2022.10032373
2022-01-01
Abstract:Traditional knowledge graph link prediction models use translation operations to obtain shallow knowledge of entities and relations, and some models utilize convolutional neural networks to extract nonlinear structures and deep knowledge of entities and relations. Linear translation models cannot mine and model deep and potential knowledge, while nonlinear convolutional neural network models give rise to the loss of most of the shallow knowledge. This paper proposes the 1D and 2D convolutional combined embedding model CombE to relieve these problems effectively. In our method, we leverage 2D convolution operations to facilitate the interaction between entities and relations to capture the potential knowledge. In order to enable total fusion of element-level embeddings, we perform vector multiplication operations on entity and relation embeddings to generate a 2D matrix. In addition, we use some 1D convolution filters over input embeddings to capture the surface and shallow knowledge. Finally, we add the results obtained from the two parts element-wise and save them. Experiments show that our method is superior to other recent methods.
What problem does this paper attempt to address?