Research Of Knowledge Representation Learning Algorithm For Knowledge Graph

Since it was proposed in 2012,knowledge graph has developed rapidly.As a new form of knowledge representation,it has become one of the three branches of artificial intelligence and an important knowledge pillar.Knowledge graph can effectively organize and represent knowledge.With the continuous development of intelligent applications,knowledge graph has been widely used in intelligent searc and personalized recommendation and other fields.Knowledge graph is a multi-relation directed graph,which usually represents knowledge in the form of network,nodes represent entities,and edges represent the relationship between entities.This kind of knowledge representation based on symbolic logic is characterized by easy to describe discrete and explicit knowledge and good interpretability,but there are still many continuous and invisible knowledge that cannot be described by symbols.Moreover,there are some problems such as low computational efficiency and sparse data.With the rapid development of deep learning,representation learning technology has achieved great success in the field of knowledge representation.Knowledge representation learning technology aims to map the entities and relations in the knowledge graph into a dense low-dimensional vector space.Although researchers have proposed different solutions for different problems,there are still some problems that need to be solved.On the one hand,in the knowledge graph,there may be multiple semantics for a relation,which requires a more finegrained division of the relation.On the other hand,the translation series model often only focuses on the structure of the triplet itself and represents each triplet in isolation,without considering other information outside the triplet,such as text description information and image information,which can be used as the semantic supplement of the triplet embedding structure vector.Based on the above two problems,this paper proposes the following two solutions:(1)In view of the existing methods can not well solve the multiple semantic problems of a relationship,this paper proposed Trans DTM model,introduced the idea of dynamic mapping matrix,and introduced two mapping matrices for the head entity and tail entity to solve the multiple semantic problems of the relationship.In order to change the original translation principle,the dynamic translation strategy is introduced to make the translation principle more flexible.In addition,during model training,a sampling strategy suitable for the model was selected to allow as many entities as possible in the dataset to participate in the training.Specifically,for 1-N relations,head entities are replaced with higher probability;For relationships with N-1,tail entities are replaced with higher probability.At the same time,the context of the relationship is used to improve the confidence of the triples.Finally,the experimental results on four datasets WN18,WN11,FB15 K and FB13 show that the Trans DTM model has better performance than other baseline models in the experimental tasks of link prediction and triplet classification.(2)Traditional representation learning models often only focus on the structure of the triplet itself and represent each triplet in isolation,without considering other information outside the triplet,such as text description information and image information,which can be used as the semantic supplement of the triplet embedding structure vector.Aiming at the above problems,in order to enhance the ability of knowledge representation,this paper proposes a knowledge representation learning model,ETLKRL,which integrates text description information and hierarchical type information.By integrating the structure of triples itself,text description and hierarchical type,it makes full use of information other than KG to improve the accuracy and interpretability of knowledge representation and knowledge reasoning.Problems such as insufficient utilization of external information are solved.In order to fuse text description information and hierarchical type information,convolutional neural network was used to introduce text description information and extract reliable feature information from text description.The attention mechanism based convolutional neural network was used to extract effective information from text sentences by assigning weights to related texts,so as to obtain a relationship vector representation with high discrimination and better semantic accuracy.A weighted hierarchical type encoder is used to construct hierarchical type projection matrices and relational specific type constraints are used to project entities.Experiments on four datasets,WN18,WN18 RR,FB15K,FB15K-237 and YAGO3-10,show that ETLKRL model has more obvious advantages in the two experimental tasks compared with the baseline model,and can deal with the problem of insufficient utilization of external information.And has a certain degree of extensibility.