Research On Knowledge Graph Representation Learning Algorithm Based On Neural Network

With the development of artificial intelligence and information technology,effective information needs to be extracted accurately from large amounts of data.As a structured representation of information,knowledge graphs have gradually entered the field of vision of researchers,making it easy to use existing data and mine new information.Knowledge graph representation learning which expresses entities and relationships of different types and backgrounds in the same semantic space without destroying structural features,plays a very important role in the construction,expansion and application of the knowledge graph.The representation learning algorithm model based on multiple feature extraction named MF＿Simpl E is proposed in the dissertation to address the problem that existing knowledge graph representation learning algorithms do not model many-to-one and one-to-many types of relationships with sufficient accuracy.MF＿Simpl E performs conversion reasoning on different types of relationships in the knowledge graph.In addition,it also uses neural networks to extract individual multiple features of entities and nodes,and calculates the similarity scores of the triples based on the features.We perform node prediction experiments on the MF＿Simpl E model on the FB15 K and WN18 datasets.The results show that MF＿Simpl E can achieve better results in representation learning than other models.However,the improvement of accuracy is limited in the improvement of accuracy in the complex type of knowledge graph.The representation learning model based on relational attention mechanism named RAGCN+MF＿Simpl E is proposed in the dissertation to improve the learning of algorithmic models on complex relational knowledge graphs.The graph convolutional neural network based on relational attention mechanism named RAGCN is proposed in the paper to extract structural features to obtain entity embedding vectors with structural features of the graph.RAGCN+MF＿Simpl E consists of an encoder named RAGCN and a decoder named MF＿Simpl E.We input the entity vector obtained by RAGCN and the initialized relationship vector into MF＿Simpl E to get the confidence scores of each triad.We uses RAGCN+MF＿Simpl E to perform triple classification experiments on the FB13 and WN11 data sets.Besides,the structural feature extraction comparative experiment and the head-to-tail entity prediction experiment were carried out on the data sets FB15K-237 and WN18 RR.Experimental results show that RAGCN can effectively extract the neighborhood features of nodes in the knowledge graph.Experimental results show that RAGCN can effectively extract the neighborhood features of nodes in the knowledge graph.Apart from this,RAGCN+MF＿Simpl E can achieve significant improvement in triple classification and node prediction tasks,and the improvement of the experimental indicators is very significant on FB15K-237 dataset with multiple relationships.MF＿Simpl E model proposed in this dissertation uses neural networks to extract individual characteristics of nodes,which can efficiently perform representation learning in a large-scale sparse knowledge graph.RAGCN+MF＿Simpl E is applicable to relational complex knowledge graphs,and it is proposed to obtain the structural features of nodes in knowledge graphs to improve the accuracy of representation learning.