Research On Relation Aware Knowledge Representation Learning Algorithm

The concept of Knowledge Graph was proposed by Google in 2015,aims to organize human cognition into a structured knowledge system.It is the basis for artificial intelligence technology to realize understanding,memory and reasoning.In recent years,it has been widely used in information retrieval,knowledge reasoning,auxiliary decision-making and other fields.As the basis for tasks such as knowledge graph completion,recommendation,and question answering.Knowledge Representation Learning(KRL)’s expressive ability largely affects the performance of such downstream tasks.Therefore,deep research on KRL is of great significance for the construction and application of knowledge graphs.In recent years,in order to realize the feature extraction of knowledge graph by KRL,researchers have done a lot of related work and made good progress,but there are still some limitations.With the widening of applications,knowledge graphs constructed for practical applications generally contain a large amount of relations.However,reviewing the existing KRL related works,most of them only model the entities and structures of knowledge graphs,which lack consideration of relations.In addition,the vast majority of existing KRL methods are end-to-end models,which can perform well on specific downstream tasks based on labeled data.And knowledge representations obtained by such methods lack generality.KRL constructs a representation space according to the features of the data itself,and different knowledge graphs do not share weight matrices,so they are mapped to different representation spaces.When applying knowledge representation to multi-knowledge graph tasks,there is the problem of representation space isolation,which reduces task performance.Aiming at the limitations of the existing methods,we propose a relation-aware Knowledge Representation Learning method(R-KRL).RKRL constructs a contrastive loss based on the idea of contrastive learning,constructs contrastive samples from two views of neighbor entities and adjacency relationships,and realizes the end-to-end decoupling of knowledge representation learning.Through experimental evaluation on a variety of KRL downstream tasks,our method is able to effectively represent relations while improving training efficiency and generality of knowledge representation.Aiming at the problem of representation space isolation when uses the R-KRL in multi-knowledge graph task,we propose a Cycle Consistency Based Knowledge Graph Domain Adaptation method(CC-KGDA).CCKGDA introduces cycle consistency loss to design an auxiliary reconstruction task,closes the loop to compare the difference between the original sample and the reconstructed sample,so as to map the two knowledge graphs into a unified representation space.And meanwhile introduces the knowledge graph negative sampling method to improve the performance.Through the experimental evaluation on the entity alignment,the method can effectively compensate for the loss caused by the problem of representation space isolation and achieve better task performance.The methods we proposed realize the end-to-end decoupling of KRL by introducing the idea of contrastive learning,thus improve the generality of knowledge representation and the efficiency of model training.Besides,we realized the knowledge graph domain adaptation by introducing the idea of cycle consistency,so the isolation problem of representation space in different knowledge graphs is solved,and the performance of knowledge representation learning on multi-knowledge graph tasks is optimized.In this thesis,we elaborated the idea and design details of the proposed methods,and verified their effectiveness by experiments.Our methods play an important role in promoting knowledge graph mining and completion tasks and solving the difficulties of knowledge representation learning tasks,and has obvious practical application value.