Few-Shot Knowledge Graph Completion Based On Neighborhood Information Fusion

The emergence of knowledge graph provides effective extraction,organization and integration of heterogeneous structural information,but knowledge graphs in real world are usually incomplete,and the introduction of new entities or new relations for knowledge graph completion is of great practicality and application value.Few-Shot knowledge graph completion task is gradually coming to the attention of researchers due to the phenomenon of long-tail distribution of relations.It predicts the missing entity according to incomplete triples,in which relations have a few relevant entity pairs.However,previous methods are limited by long-tail relations because of requiring substantial head and tail entity pairs for training,moreover,the majority of these methods could not effectively capture the relevance of entities and relations.And some algorithms for few-shot knowledge graph completion pay more attention to the role of entities in the neighborhood and ignore the contribution of neighbor relations when learning feature representations of the task relation.To overcome these drawbacks,we propose a few-shot knowledge graph completion model based on neighborhood information fusion named as FNIF.It is applicable to few-shot scenario inspired by the idea of meta-learning and consists of two modules for embedding and matching respectively.A neighbor relation-meta encoder is devised to describe relational relevance by using entity correlations,which could incorporate features of entities through the attention mechanism.As a result,we can capture continuous representations of relations being transferred to the query set.And the second module is a matching processor for optimizing relation embeddings and performing comparison and matching of queries and candidates.Considering the importance of query entities in the matching process,another few-shot knowledge graph completion model combining Transformer and convolutional neural network with application of meta-learning is proposed,which is called FTCN,it includes three parts: relationmeta learner,Transformer encoder,and the matching processor,respectively,to realize the function of encoding long-tailed relations based on information in reference set,updating query entities by combining reference information,and completing the matching of query entities,task relation-meta and candidate tail entities.The main contributions of this paper are as follows:(1)This paper analyzes typical models of knowledge graph completion and investigates how to apply the relevant techniques to few-shot problem.Since metalearning is applicable to few-shot tasks,this paper investigates how to construct a knowledge graph completion model on the basis of meta-learning.The acquired longtail relation embeddings are transferred from the reference domain to the query domain as migratable meta-information,and the training process is accelerated by using a gradient update strategy to achieve accurate and efficient prediction.(2)Graph attention network is applied to fuse the one-hop neighborhood information of entities,and an idea of transforming entity-level attention into relationlevel attention is proposed to obtain an effective representation of long-tail relations.(3)Transformer encoder is used to update query entity representations to achieve the fusion of information in reference set and query set efficiently,and to achieve more accurate representation of information interaction between task relations and query entities.(4)Extensive experiments reveal that our model performs better in comparison to several baselines on two datasets.In this paper,we conduct extensibility experiments on two standard datasets,scalability experiments are conducted on two datasets,and the experimental results show that both models outperform the two sets of baseline algorithms selected at the time of proposal.Time complexity and neighborhood attention of the model are analytically investigated,and it is demonstrated that the model guarantees the execution efficiency when achieving expected results.