Research On Knowledge Graph Completion Methods Based On Deep Learning And Tensor Decomposition

In recent years,knowledge graphs are widely applied to various fields of artificial intelligence systems by human beings,such as intelligent question answering systems,intelligent search engines and so on.The possibility of omissions in knowledge graphs is very high,because the methods of the knowledge graphs’ construction are not perfect,which may result in the loss of information on graphs.In order to improve the completeness of knowledge graphs,the research on the knowledge graph completion methods have attracted more and more attention.Many knowledge graph completion methods have been proposed,and knowledge graphs can be basically completed by these methods.But there are still some problems with these methods.For example,training data sampling process of some methods are so defective that may result in the low quality of the training set and restrict the upper limit of model accuracy.Complete semantic information cannot be assigned by these methods to vectors that can be understood by computers as well.And too much semantic information about the graphs is focused on some methods,and the structural information about graphs is ignored.In this paper,an in-depth study was made related to the above problems,and the main works are as follows:1.Due to the problem of low training sets quality,a triple negative sample sampling method named Uniform-Encoder based on the deep learning model Transformer-Encoder is proposed.Nowadays,the negative sample generation of training set in knowledge graph completion methods is mostly random replacement of head or tail entities,and a large number of "invalid triples" may be generated by these methods.The loss cannot be reduced during the training process,resulting in the failure of the model parameters to successfully converge.The purpose of the negative sample sampling method based on the deep learning model in this article is to improve the rationality of the sample.More reasonable negative samples can be screened out by the negative sample sampling method proposed in this paper.It is verified that the negative sample sampling method proposed in this paper can filter out reasonable negative samples as training data,the quality of training data and the accuracy of the knowledge graph completion model are improved as well.2.Due to the problem of insufficient feature extraction,a knowledge graph completion model based on XLNet pre-trained language model(KG-XLNet)is proposed in this paper.The XLNet pre-trained language model and the LSTM model are contained in the model’s feature extraction part.And the extracted features are passed into the classification and scoring algorithms for relationship prediction and link prediction of the knowledge graph.In addition,the integrated learning methods were incorporated into this model to improve the accuracy of predictions.3.Due to problem of the single-step reasoning based models cannot make full use of the structure information about the graphs,a knowledge graph completion model based on graph path and tensor decomposition is proposed in this paper,in which the tensor decomposition method of representation learning to obtain the vector representation of entities and relations was used,and aggregation of relational paths of the depth first path search algorithm and RNNCell to effectively utilize the structural information of the knowledge graph are realized in the model.Finally,experiments are carried out based on the negative sample sampling optimization method and two graph completion models.The experimental results show that the existing problems of the current knowledge graph completion methods are improved by the models and methods proposed in this paper to a certain extent.