Knowledge Link Inference of Graph Structure Based on Holographic Model

Abstract

For current knowledge link inference methods, whether it is traditional translation models, semantic matching models, or convolutional neural network models, it is impossible to obtain rich semantic information. This paper mainly uses a pre-training layer based on the holographic model, and combines the knowledge structure to perform the knowledge link inference. Firstly, the pre-training layer is used as the model initialization. Secondly, the graph structure encoder layer not only combines the information of entities and relationships directly related to the current entity, but also considers the information including multi-hop neighbor nodes and auxiliary relationships. Finally, ConvKB is used as a decoder to score the triples. The model is evaluated on two benchmark datasets WN18RR and FB237, that is slightly better than the previous embedding models on some indicators.