Research On Knowledge Graph Embedding Algorithm Based On Modeling Relation Patern

The knowledge graph is a multi-relational directed graph composed of fact triples,which can structurally represent the rich knowledge in the real world and thus is widely used in some downstream tasks,such as recommendation systems,knowledge answering,and dialogue systems.However,due to the complexity of the real world,existing knowledge graphs still lack a large number of fact triples.The knowledge graph embedding method has been proposed as a technique to effectively complement the knowledge graph.It maintains the underlying meaning and structure of existing graphs while simplifying the operation by mapping the entities and relations of each triple in the knowledge graph onto a continuous vector space.The ability of knowledge graph embedding to effectively reason about missing facts based on existing knowledge lies mainly in the model’s ability to model multiple relational patterns.Intuitively,this thesis argues that if the knowledge graph embedding model can capture as many relational patterns as possible,it will effectively improve the representation capability of the model.To address this view,this thesis proposes a simple and effective model that models as many existing relational patterns as possible.The main work and innovations in this thesis are as follows.Firstly,a model LinBiasE is proposed.By extending the translational distance model,it can improve the representation capability while it can effectively model symmetric/antisymmetric,inversion,commutative/non-commutative composition relation patterns,and also solve complex relation problems(1-to-1,1-to-many,many-to-1,many-to-many).With an improved self-adversarial negative sampling mechanism to optimize the model,the model can better learn various embedding representations and reduce the risk of overfitting.The results on three benchmark datasets show that the LinBiasE model can effectively model various relational patterns and improve the representation capability,and it has good competitiveness among existing models.Next,an improved model LinBiasH is proposed.While LinBiasE can effectively model multiple relational patterns,it also sacrifices a large amount of memory space because it requires high-dimensional vectors to provide flexibility for modeling relational patterns.In contrast,hyperbolic spaces can model hierarchical data well while maintaining lowdimensional vectors.LinBiasH learns low-dimensional embeddings while preserving the underlying hierarchical structure by extending the model LinBiasE into hyperbolic spaces.The results on two benchmark datasets in low-dimensional conditions show that the LinBiasH model can greatly improve the performance of the LinBiasE model.Subsequent experiments also demonstrate that the LinBiasH model is competitive even in high-dimensional conditions.