Knowledge Representation Learning With Semantic Interactions And Semantic Constrains

A knowledge graph is a directed graph composed of entities and relations.Knowledge graphs storing facts about the world have been popular in a variety of artificial intelligence tasks including web semantic analysis,dialogue systems,and recommendation systems,etc.Recently,knowledge representation learning utilizes low-dimensional dense real-valued vectors to represent entities and relations,so as to calculates semantic connections between entities and relations in low-dimensional vector spaces.Knowledge representation learning can effectively alleviate the problem of data sparsity,and realize the fusion of heterogeneous information,and then improve the performance of knowledge acquisition,fusion and inference.This paper focuses on the problems of modeling semantic association in knowledge representation learning,and conducts research work from the perspective of semantic interactions and semantic constrains.First of all,there are complicated semantic interactions between entities and relations in knowledge graphs.And most knowledge representation learning models exploit the semantic interaction between head entities(or tail entities)and relations,but ignore the semantic interaction between head entities and tail entities.In addition,entities in a triple of the knowledge graph have many similar attributes,but most knowledge representation learning models neglect the approximate constraints between the underlying features of entities.To deal with these issues,this paper separately proposes knowledge representation learning with semantic interactions and with semantic constrains.The main work of this paper is as follows:(1)This paper proposes a knowledge representation learning model integrating semantic interactions,which models semantic interactions between head entities(tail entities)and relations,and between head entities and tail entities,while transferring different semantics with forward and backward.Specifically,this model represents entities and relations as complex embeddings,and measures the semantic similarity between entities translated by relations,and thus models forward and backward semantic interactions.(2)This paper proposes a knowledge representation learning model with semantic constraints,which imposes semantic distance constraints and feature compression constraints on entity embedding vectors.This model restricts the geometric structure of embedding space with semantic distance constraints,in order to model the potential association of entities within a triplet.Through feature compression constraints,this paper retains the positive and negative attributes of entities and learns compact embedding representations.This paper conducts experiments on standard link prediction task to evaluate proposed models.Knowledge representation learning with semantic interactions and with semantic constrains achieve state-of-the-art performances,which further demonstrates that semantic interactions and semantic constrains can help improve the representation ability.