Knowledge Representation Learning Based On Structure And Semantics

In large-scale knowledge graphs,the ability of knowledge representation learning directly affects the effect of knowledge graphs in providing user portraits or performing intelligent reasoning.In recent years,the representation learning method,represented by deep learning has provided a new idea for the representation of knowledge,which transformed the structured knowledge graph into a low-dimensional vector to realize the distributed representation of entities and relationships.In the study of knowledge representation learning,integrating of multi-source heterogeneous information such as entity descriptions,entity information in the Internet,and entity information in other knowledge bases into triple structure information has become a new research direction.This method enriches the semantic information expressed by triples,but only the word-level and sentence-level relationship is considered in the fusion process,and no hidden relationship between entities is found.Through the study of entity description text in the knowledge base,it is found that the subject information of the entity description text is closer to the relationship between the entities,and can reveal the hidden associations between the entities.Therefore,this paper proposes a knowledge embedding model that combines entity describe topic,and enriches the structure information of triples by using the topic between description texts.The experiment on Freebase's standard benchmark data set FB15 K,and shows that the new model is significantly better than other models that incorporate entity description information,and outperforms the performance of the baseline model in both link prediction and entity classification.Translation-based model is one of the important directions of knowledge representation learning research.This type of model mainly focuses on the structural information of triples.Although the translation model is more efficient in memory and time,it cannot remove the influence of noise feature dimension during encoding.There are still some restrictions on encoding relation models.The improved translation model only studies one of the problems,and does not completely solve the defects of the translation model.Through the study of the translation model,the article proposes an improved translation model,which can solve the two major defects of the original model at the same time.The verification experiments on the sub-data sets FB15 K and WN18 under the articles Freebase and Word Net show that the new model is significantly superior to other translation-based embedded models,and has achieved excellent results in both link prediction and entity classification tasks.