Knowledge Graph Completion Model Based On Representation Learning

Knowledge graph is a structured and semantic representation of the real word in the form of triples and models the connection between everything in the form of graph structure.Knowledge graph is an important branch of artificial intelligence.However,due to the limitation of knowledge acquisition and the continuous addition of knowledge,the existing knowledge graphs are all incomplete,so it is necessary to complete the knowledge graphs.Knowledge graph completion is simply to infer the missing part or implicit knowledge by using existing triples.Representation learning embeds entities and relations into a lowdimensional continuous vector space with semantic constraints,maintaining the inherent structure of knowledge graph while improving the interpretability of the model.With the development of representation learning,there are fruitful results in the research on knowledge graph completion.This thesis studies knowledge graph completion models based on representation learning.In summary,there are three contributions in this thesis:First,most existing models only focus on the structural information of single triple and ignore the importance information.This thesis proposes an Imp Rotat E model integrating importance information.It thinks that the importance of different information should be distinguished.Triplet importance has two parts.Entity importance is defined by Page Rank algorithm with improved transfer probability.Relational importance is correlated with entity importance and average number of mappings on both ends.The results of the link prediction experiment show that Imp Rotat E achieves better completion performance.Second,in view of the problems of polysemy and complex relationships in knowledge graphs,a knowledge graph completion method IQuat E is proposed.Most models are fixed for the representation vectors of entities and relations and seem encountering bottleneck when facing the challenges.IQuat E is a quaternion embedding model with dynamic transformation,which adds a layer of dynamic change vector before entity representation and relation representation.The experimental results on multiple datasets show the validity of the modeling method.Compared with Quat DE which has as many parameters as model in this thesis,the computation is reduced and the effect is improved.Third,most models,whether embeds in a real space or a complex number space,ignore component semantic information of entity,and the triples are independent in the training optimization.Referring to the idea of Simpl E,this thesis proposes a model SQuatE to enhance the entity representation.Two sets of vectors are used to distinguish the entity as the head entity and the tail entity.At the same time,the reverse relations are constructed to establish reverse triples to assist training.The performances of SQuatE are verified by different experiments.