Research On Representation Learning Algorithms For Knowledge Graphs

Knowledge graph is a structured information encoding of rich relationships between entities,which aims to organize real-world facts into computer-readable structures in the form of triples.The representation of triples is(h,r,t),h represents the head entity,t represents the tail entity,and the relationship between h and t is represented as r.It can be seen that this kind of organization makes the knowledge graph show strict logical rules and scattered symbolic characteristics,which is limited in many fields of application.For this reason,a representation learning method oriented to knowledge graph is proposed,which encodes each element(entity and relationship)in the knowledge graph into a continuous low-dimensional vector space to obtain computable vector,such as TransE,TransH,TransR and other representation learning model based on translation.Despite the success of these translation-based models,there are still some problems.On the one hand,these models mostly focus on the semantic features represented by the independent triples,ignoring the fact that the entity itself is an individual with comprehensive attributes.In fact,for different relationships,attention should be paid to the different attribute representations of the entity itself.On the other hand,the knowledge graph follows a certain human cognitive thinking mode when constructing,and retains the semantic hierarchical structure between entities.However,the traditional translation model ignores the size of the semantic category between entities when performing representation learning,so the semantic level of the knowledge graph cannot be expressed well.To solve the above two problems,this paper introduces entity attributes and semantic level information,and proposes the following two solutions:(1)This paper proposes a representation learning algorithm based on PartOf relationship – TransP.By analyzing the representation attributes of the entities in the knowledge graph,it is found that the head entity and tail entity of the triple under the PartOf relationship do not belong to the same entity type,and the head entity is a component of the tail entity,and the attribute difference is obvious.Therefore,the model first encodes each entity into a sphere,and uses the relative position relationship between the spheres to model the PartOf relational triples.Secondly,design the loss function training model according to the Euclidean distance used in other translation models.Finally,the sensitivity of marginal parameters is analyzed during the training process.(2)A representation learning algorithm,Trans?isA,which integrates TransP and semantic level is proposed.Based on the research of TransP,inspired by the transitivity of the relationship,it is believed that the semantic hierarchical relationship in the knowledge graph should retain the transitivity feature during representation learning,so as to ensure the subordination characteristics of the semantic category.Therefore,the model firstly models the semantic relationship of entities according to the transitive judgment condition of the binary relationship.At the same time,in order to better combine the attributes of entities and the semantic relations between entities,the core algorithm of the TransP model is incorporated,which effectively improves the knowledge representation ability of the Trans?isA model.On WN18 and WN18 RR,the above two models are verified by experiments.Firstly,the performance of the models is improved in the link prediction experiment.Secondly,the test indicators in the triple classification experiment are improved by about 10%,which demonstrates the representation learning capabilities of the two models.