Quaternion-Driven Knowledge Graph Completion

Rapid development on big data and artificial intelligence inspires a research upsurge in knowledge graph(KG)related technologies for smarting the search function on search engine.Large-scale general domain KGs are mostly constructed-manually or semi-automatically,and are far from a complete state.Knowledge graph completion(KGC)is an essential mechanism for the completeness of KGs in dynamic evolution process.Based on the unique characteristics of quaternions,quaternion groups,dynamic dual quaternions and capsule networks,this paper focuses on the most basic and core knowledge completion in the field of KGs,and explores and researches effective methods and key technologies aiming at the problem of KGC.The main research contents are as follows:1.To solve the deficiency on the ability to modeling relationship,which results in the loss on inter-entities semantic relationship.By studying relational patterns of KG,quaternion rotation operator and spatial geometric meaning of quaternion,Qua R a KGC method using quaternion as relational rotation is proposed,combining the advantages that quaternion representation is very suitable for smooth rotation and spatial transformation parameterization in vector space.Qua R maps entities and relations of KG to quaternion vector space,and defines each relation as a rotation from head entity to tail entity.Theorem proving verifies that Qua R can effectively model and reason symmetry,anti-symmetry,inversion and composition patterns.The experimental results show that Qua R is effective.2.To solve the insufficiency on mining the attribute information of each dimension of triples,which results in the loss of semantic information of triples.By studying the network structure of convolutional neural network and capsule network,combining the advantages of capsule network,more feature information can be encoded and retained in the whole network,Cap S-Qua R a quaternion-embedded capsule network KGC method is proposed on the basis of Qua R.Cap S-Qua R takes the training results of Qua R with strong relational modeling ability as the input of the optimized capsule network in this paper.After a series of operations such as convolution,reorganization,dynamic routing and inner product of the capsule network,triplet score is obtained to judge whether the triplet is correct or not,so as to supplement KG.The experimental results show that Cap S-Qua R has good performance and high accuracy.3.To overcome the insufficiency on modeling multi-hop combinational relationships,resulting in the loss of semantic information of some combinational relationships.By studying the group theory in relational embedding,the characteristics of quaternion group and combination relationship in KG,according to the corresponding relationship between group theory and relationship patterns,and using the characteristics of quaternion group,Quat GE a KGC method based on quaternion group is proposed.Quat GE uses the notation "axis-angle" to model the rotation operation of relationships in the quaternion group space.The experimental results show that Quat GE improves the accuracy of link prediction tasks compared with similar methods,especially in modeling complex combinational relationship patterns.4.To overcome the vulnerability on the feature interaction between entities and relations,which results in the loss of complex semantic connections between them.By studying the spatial structure of dual quaternions and combining the advantages of dual quaternions that can represent arbitrary rotation and translation in space,Dual DE a KGC method based on dynamic dual quaternions is proposed.Firstly,Dual DE defines the dual quaternion representation of relationship rotation and translation.Then,the dynamic strategy in dual quaternion space is designed.Finally,the dynamic mapping mechanism is used to construct the entity transfer vector and relationship transfer vector,and the embedding position of entity vectors in the dual quaternion space is continuously adjusted according to the dual quaternion multiplication rules.The complex relationships such as one-to-many,many-to-one and many-to-many are dynamically constructed,which enhances the feature interaction ability between triple elements.The experimental results show that Dual DE improves the experimental accuracy compared with similar methods,especially in modeling complex relationship types such as one-to-many,many-to-one and many-to-many.This paper studies and analyzes the main problems existing in KGC,proposes quaterniondriven KGC methods,and proves the effectiveness of these methods through theoretical analysis and experiments.It provides a certain theoretical basis and practical application reference for further discussion and research on the key technologies of KGC.