Research On Knowledge Graph Completion Based On Convolutional Neural Network

Knowledge graph completion is the process of obtaining new facts by predicting the hidden relations between entities,thus making the knowledge graph more complete.Currently,embedding entities and relations in the knowledge graph into a continuous lowdimensional vector space has become an effective method for knowledge graph completion.Models such as ConvE,which applies convolutional neural networks,and KBAT,which applies graph convolutional neural networks,have achieved good performance in the knowledge graph completion task.However,models such as KBAT cannot model the bidirectional semantic relationships between entities and cannot obtain rich multi-hop neighborhood features,and models such as ConvE have difficulty in obtaining deep features of triads.To address the above problems,this paper proposes the multi-hop bidirectional attention model MBGAT and the residual-based convolutional neural network model Res DE by improving graph convolutional neural networks and convolutional neural networks to improve the accuracy of knowledge map completion.The main research of this paper is as follows:(1)Three categories of current knowledge graph completion methods are analyzed,including tensor decomposition-based knowledge graph completion methods,translation model-based knowledge graph completion methods,and deep learning-based knowledge graph completion methods.Their advantages and disadvantages are compared,and the problems existing in them are analyzed.(2)To address the problems of models such as KBAT,an encoder-decoder-based multi-hop bidirectional attention model MBGAT is proposed.The model can learn the information in the larger neighborhood of the entity,and obtain the bidirectional semantic relationship between the entities,which improves the richness of the information in the aggregated neighborhood of the entity,and improves the embedding quality of the entity and the relations.Experiments on the WN18 RR,FB15k-237 and NELL-995 public datasets show that the MBGAT model has improved performance in all metrics compared to the benchmark model,and improved the accuracy of knowledge graph completion.(3)Aiming at the problem that models such as ConvE only perform triple operations in shallow convolutional neural networks,a complete model Res DE based on residual convolutional neural networks is proposed.This model can use deep convolutional neural networks to extract the deep-level features of triples,and at the same time reduce the problem of information loss and gradient disappearance as the network layer deepens.Experiments on the WN18 RR and FB15k-237 public datasets with Res DE independently and as a decoder of the MBGAT model show that Res DE as an independent structure and as a decoder of the MBGAT model are superior to the benchmark model.