Research On Knowledge Graph Representation Learning Based On Entity Attribute Information

Currently,Knowledge Graph(KG)as a structural representation of knowledge has attracted great attention from researchers.Knowledge Graph Representation Learning(KRL)aims to learn an accurate knowledge representation with semantic information.The current KRL methods only consider the structural information and other semantic information(such as description information,image information,etc.)in KGs,and the application of entity attribute information is insufficient and incomplete.In addition,attribute information is usually complex,heterogeneous,diverse and difficult to utilize.How to use entity attribute information to improve the effect of entity representation is a challenging task.In order to solve the above problems and make use of attribute information reasonably,this paper proposes a knowledge graph representation learning model Duet Entity Representation Learning Model(DERL),which can combine the structure information and attribute information of entities to improve the overall effect of knowledge representation jointly.In order to encode entity attribute information reasonably,this paper designs an attribute encoder Entity Attribute Encoder(EAE)in DERL,which can reasonably encode complex attribute information to capture entity attribute semantic information.EAE contains two encoding components: an attribute type encoding component and an attribute value encoding component,which are used to encode entity attribute type information and entity attribute value information respectively.In addition,this paper embeds an attribute attention mechanism in the DERL model to distinguish the importance of different attribute types and attribute values to entities.In order to use entity attribute information to solve the Zero-Shot problem,this paper proposes a representation learning model Zero-Shot Representation Learning Model(ZSRL),which can use entity attribute information to learn the representation of new entities in KGs to solve the two problems in Zero-Shot scenario:(1)the newly added entities in KGs cannot be directly represented;(2)a compatible and accurate representation cannot be learned for the newly added entities.In this paper,we design an attribute encoder Entity Attribute Encoder Bi-GRU(EAEBG)for the ZSRL model,which can effectively encode the attribute information of new entities for entities' representations.In this paper,we design a fusion scoring function for the ZSRL model to further train the entity's structure-based representation and the entity's attributebased representation to improve the compatibility of the entity representation and further improve the effectiveness of the ZSRL model on the Zero-Shot problem.In this paper,the effectiveness of the DERL model and ZSRL model is verified on the real datasets DWY100 K and FB24 K.By comparing with the experimental results of the current popular models,the DERL model has a maximum performance improvement of 14.3% on the Knowledge Graph Completion(KGC)task,the ZSRL model has a maximum performance improvement of 28.6% on the Knowledge Graph Completion task based on Zero-Shot scenario.The experimental results prove that the representation learning model proposed in this paper can use the attribute information of entities to learn an accurate and rich semantic knowledge representation,and can combine the structure information of entities to improve the overall effect of KRL.This not only proves the scientificity and validity of DERL model and ZSRL model but also proves that EAE and EAEBG have excellent coding ability.In this paper,the use of attribute information to solve the Zero-Shot problem is proposed for the first time,which provides a new solution for the Zero-Shot problem in KRL.