Research On Knowledge Graph Reasoning Method Based On Deep Learning

With the rapid development of Internet technology,human society has accumulated a large amount of theoretical knowledge and practical experience.The knowledge graph is an effective means of knowledge representation for its highly condensed semantic network.Knowledge graphs can provide much convenience for many aspects of people’s lives,including search engines,question and answer systems,and so on.Although the scale of knowledge graphs is increasing exponentially,it still exists problems of sparseness and incompleteness.How to effectively mine the latent semantics of knowledge graphs and obtain potential facts is an urgent issue.At present,deep learning has made outstanding progress in the field of natural language,especially in the field of knowledge graph reasoning.This thesis mainly focuses on the problem of multi-semantic fusion and proposes a knowledge graph inference model based on multi-semantic neural network.The main contributions of this thesis are as follows:1.In order to solve the problem of single information types in mainstream knowledge graph reasoning models,this thesis proposes a Multi-semantic Deep Neural Network MSNN.MSNN uses message passing neural network and recurrent neural network to capture entity contextual information and relational path feature,then MSNN uses the attention mechanism to merges these semantics.The experimental results show that the MSNN model has achieved high performance in the current mainstream methods on the MRR,MR,Hit@1 and Hit@3 indicators of five datasets such as FB15k-237,WN18 RR,and NELL995.It proves that the use of two semantic information can improve the expressive ability of the model.2.In order to address the limitation of the fixed weights of propagate functions in traditional message passing neural network,this thesis proposes a relation-oriented convolutional message passing neural network in MSNN.With the efficient learning ability of the convolutional neural network,we calculate the weight of the entity and relation of the propagate function in a learnable way,which make the representation of the spread function more flexible.The experimental results show that the performance of the propagation function weight calculation scheme based on the convolutional network on the MRR,MR,Hit@1,Hit@3 indicators is higher than the traditional averaging scheme.3.In order to further improve the performance of MSNN network,this thesis designs a relation vector initialization scheme based on text embedding and proposes Text Enhanced-MSNN(TE-MSNN)to improve the expression ability of the initialization relation vector in the MSNN model.The representation of the relational text is obtained through the methods of text representation generation,text semantic filtering,and text representation dimensionality reduction.Finally,the representation is used as the initial relation vector of MSNN.4.In order to verify the performance of the TE-MSNN,this thesis constructs a new relational text enhancement dataset NELL995-text based on the NELL995.We crawl the description of the relation and the description of the entity type in each triple from the official website of the NELL system to generate the description text of each relationship.The experimental results show that TE-MSNN has a certain improvement compared with MSNN in MRR、MR、Hit@1、Hit@3.