Research On Embedded Model For Knowledge Graph Completion

With the advent of the era of big data,the storage and representation of knowledge has become particularly important.Knowledge graph is one of the most effective ways of knowledge representation,which is widely used in intelligent applications such as intelligent search,smart question answering and personalized recommendation.Although the scale of existing knowledge graph is very large,it is still far from the complete state of knowledge.Actually,the completeness of knowledge graph will directly affect the performance of intelligent applications.To this end,knowledge graph completion technology has attracted much attention and has become a major research topic.Knowledge graph embedding aims to use continuous,dense and low-dimensional vectors to represent entities and relations in a knowledge graph,so as to implement knowledge reasoning and knowledge graph completion in the low-dimensional vector space.In recent years,the translation-based models show strong feasibility and robustness,and achieved state-of-the art performance in the task of knowledge graph completion.However,there are still some limitations in the existing translation-based models,such as the results of knowledge representation are inaccurate,and quality of the negative triplets is unsatisfactory.In this paper,we propose MvTransE and TransE-SNS models to overcome these shortcomings.The main research contents are as follows:(1)This paper proposes a multi-view learning embedding-based model,named MvTransE.The model generates multiple parallel subgraphs from semantic and structural perspective of entities.Then,we embed the original knowledge graph and generate subgraphs into global and local view spaces,respectively.Finally,we propose a multi-view fusion strategy which aims to integrate multi-view representations of relational facts.MvTransE solves two shortcomings of the existing model.Firstly,TransE,TransH and other models focus on elaborating the global representation of relational facts in a global perspective,while fail to learn various type of facts discriminatively.Particularly,it makes entities and relations congestion in embedding space,which reduces the precision of representing vectors regarding entities and relations.Secondly,puTransE adopts multiple parallel spaces to learn local facts,which impairs global facts of original knowledge graph,and thus reduces the ability to learn simple relational facts.Extensive experimental results show that MvTransE has achieved the most state-of-the-art performance.(2)This paper proposes a similarity negative sampling strategy for generating high-quality negative triplets.Firstly,the strategy divides all entities into multiple clusters by KMeans clustering algorithm.Then,be corresponding to each positive triplet,we replace the head entity by a negative entity from the cluster in which the head entity is located,and replace the tail entity in a similar approach.We combine similarity negative sampling strategy with TransE to get TransE-SNS.Similarity negative sampling strategy solves the problem that TransE generates extensive low-quality negative triplets in training.Extensive experimental results show that TransE-SNS significantly improves model performance compared to TransE.