Research On Knowledge Representation Learning Method Based On Deep Embedding

Knowledge representation learning is an important direction in artificial intelligence research.Its goal is to convert natural language text or other forms of knowledge into a machine-recognizable form,so that the machine can use this knowledge to perform tasks such as logical deductions,decision-making,and responding to inquiries.In knowledge representation learning,it covers many fields such as semantics,logic,reasoning,etc.These technologies are one of the core technologies for realizing artificial intelligence.This thesis is devoted to the study of relevant embedding models for knowledge graph representation learning.These models can express the extracted knowledge in a concise and efficient vector form through a reasonable embedding method,which is convenient for computer processing and analysis.Simultaneously,it is fused with the knowledge graph to enhance both the expressive and reasoning capabilities of the knowledge graph.An important area of research in knowledge representation learning is the utilization of translation-based techniques for learning representations in knowledge graphs.The fundamental concept involves utilizing low-dimensional vectors to represent entities and relationships within a knowledge graph.These vectors are subjected to geometric transformations to gauge the semantic similarity and correlation between entities and relations.However,most models directly embed based on existing facts,ignoring the differences of different types of objects,and the processing of embedding vectors is relatively simple,resulting in insufficient expressiveness of vectors.In addition,most models have certain shortcomings when dealing with complex relations.This thesis makes a comprehensive analysis of the relevant methods of knowledge graph representation learning,proposes the following two solutions to the problems existing in the existing representation models,and explores the relevant applications of knowledge representation models:(1)Existing embedding models based on translational distance only consider different types of relations when modeling the embedding of entities and relations in independent entity space and relational space,ignoring entity types,and have shortcomings in dealing with reverse relations.Aiming at the above problems,this thesis proposes a multi-pattern deep embedded knowledge representation model based on entity-relation mapping matrix.In this model,first of all,the entity-relation mapping matrix is introduced,and the measurement of entity types is added on the basis of considering various types of relations;then,the concept of multi-modal deep embedding is introduced,and the forward translation geometric distance model,the reverse translation geometric distance model and the symmetric relation model are integrated at the same time,so as to eliminate the defects in processing the reverse relation and improve the processing efficiency of the symmetric relation.Experimental results show that the model’s ability to reason about missing information has been significantly improved,and its ability to handle complex relational patterns has also been significantly enhanced.Compared with other current state-of-the-art matching models,this model has significant advantages.(2)When entity and relation are mapped to complex vector space in the previous rotation-based embedding model,there is a lack of relation between the head entity vector and the tail entity vector in complex vector space,as well as the lack of expressive force of embedding vector and the low flexibility of rotation of entity and relation.To solve the above problems,this thesis proposes a deep convolutional quaternion-based inverse relational rotation embedding knowledge representation model.In this model,firstly,quaternions are used to expand the complex-valued space and introduce a more expressive hypercomplex number representation to model entities and relations;then,different embedding generation classes are used to generate embedding vectors for the head entity and tail entity,and on the basis of defining the relation as the rotation of the head entity to the tail entity in the space,a reverse relation vector is introduced to enhance the connection between the head entity and the tail entity;finally,the quaternion embedding is deeply processed using a convolutional neural network to improve the model link prediction accuracy.Compared with other baseline models,this model can represent entities and relations more reasonably,and can effectively model and reason complex relational patterns,improving the effectiveness of knowledge representation.(3)The multi-pattern deep embedding model based on entity relation mapping matrix and the inverse relationship rotation embedding model based on deep convolution quaternion proposed above are used to conduct experiments in combination with relevant medical research results data sets extracted from social platforms,aiming to explore the effectiveness of the two knowledge representation models in linking to predict the latest medical research results.The experimental findings demonstrate that the knowledge representation model proposed in this thesis exhibits exceptional performance in the medical recommendation link prediction task.Therefore,using this model to push the latest medical research results to doctors on social platforms has obvious advantages.