Research On Key Technologies Of Knowledge Graph Representation Learning For Multi-dimensional Features

With the rapid development of big data and artificial intelligence technology,knowledge graph has attracted extensive attention of researchers and industry.Knowledge graph representation learning,based on the research of word vector embedding in natural language processing,using low dimensional dense continuous vectors to represent massive knowledge in knowledge graph,has become a hot research field.Because the traditional static knowledge graph can not meet the needs of a large number of time sensitive facts,temporal knowledge graph and its representation learning came into being in recent years and achieved rapid development.By studying the problems existing in the representation and learning of existing knowledge graphs,this paper proposes two knowledge representation models for static and dynamic knowledge graphs: static knowledge graph representation and learning model based on entity-relation interaction and temporal knowledge graph representation and learning model based on quaternion vector space rotation,and analyzes their adaptability,advantages and disadvantages respectively,Its effectiveness is verified by experiments.The main work of this paper includes:Firstly,according to the characteristics of semantic interaction and influence between entities and relations of knowledge graph,a static knowledge graph representation learning method based on entity-relation interaction is proposed.By proposing a new entity-relation interaction mechanism,the interactive semantics of entity and relation can be fully learned from the context information in the process of representation learning.Based on the proposed entity-relation interaction mechanism,two interaction methods are proposed for training based on feed-forward neural network and convolution neural network respectively.Through the comparison of embedded visualization,the experiments show that the proposed interaction mechanism can obtain the embedded representation containing more semantic information in the knowledge graph representation learning,which proves its effectiveness.Secondly,according to the characteristics that entities and relations evolve with the development of time in temporal knowledge graph,a temporal knowledge graph representation learning model based on quaternion vector space rotation is proposed.By using the characteristics of the quaternion system proposed by Hamilton,both entities and relations are encoded as quaternion embedding,and the temporal-evolving entity embedding representation is modeled as rotation in quaternion vector space.It not only proves theoretically that the proposed method can model complex relation patterns such as time series evolution patterns,but also analyzes the abilitity of the model in modeling various relation patterns combined with cases.