Research On Few-shot Knowledge Graph Completion Method Based On Representation Learnin

With the development of Internet technology and the continuous improvement of artificial intelligence technology,human knowledge can be widely disseminated in the network.In order to integrate the messy and diverse knowledge in an orderly manner,researchers have proposed the concept of Knowledge Graph(KG).Once the knowledge graph was proposed,it attracted widespread attention from academia and industry.However,with the continuous development of the knowledge graph,its existing problems have gradually been exposed,that is,there are missing data in the knowledge graph,and these missing data seriously affect the integrity of the knowledge graph is guaranteed.The emergence of knowledge graph completion technology has solved this problem well,which helps to rich the knowledge in the knowledge graph,improve the integrity and quality of the knowledge graph,and promote the development of artificial intelligence,but the traditional knowledge graph completion method usually requires a large amount of data for training,and its performance is not good in the case of sparse data.Therefore,the knowledge graph completion method that can achieve better results with a few number of samples is particularly important.Although there has been some research work in the field of few-shot knowledge graph completion at home and abroad,there are still some problems to be solved:(1)The existing few-shot knowledge graph completion methods cannot distinguish neighbors well when dealing with complex relationships.(2)In the process of aggregating entity neighbors,these methods often introduce task-independent noise information,which has a bad influence on the aggregation effect.This paper conducts research on the above two problems in the task of few-shot knowledge graph completion,and has achieved the following results:(1)A few-shot knowledge graph completion method combined with type-aware attention is proposed.This method can make full use of the entity neighbor information.Firstly,the type-aware neighbor encoder is used to learn the implicit type information contained in the entity neighbor,and the type-aware attention is obtained to enhance the entity representation;secondly,the Transformer encoder is used to capture the different meanings of the task relationship;finally The reference set is aggregated by the joint matching prototype network to obtain the reference set representation and perform entity prediction.The method is experimentally verified by entity prediction tasks on two public datasets,NELL and Wiki.The experimental results show that the method can effectively improve the performance of few-shot knowledge graph completion by learning more abundant entity neighbor information.(2)A few-shot knowledge graph completion method incorporating structural space information is proposed.This method first performs k-means clustering on the entities,and then calculates the cosine similarity and Jaccard similarity between entities to obtain the neighbor set of the entity in the structural space,and finally combines the entity space neighbors and the semantic neighbors in the knowledge graph,to further optimize the weight distribution when the entity neighbors are aggregated,and provide a more valuable entity representation for subsequent entity prediction.Finally,the experimental verification is carried out on the NELL and Wiki datasets,and the method further achieves comparable results.(3)Based on the above two research points,an interactive knowledge graph completion prototype system based on B/S architecture and Django framework was built,which visualizes the knowledge graph and realizes the knowledge graph completion function,enabling users to interactively way,select the head entity and relationship of the fact to be predicted which interested in,and output the prediction result through the model.