Few-Shot Knowledge Graph Completion Based On Attention Mechanism

Knowledge graph is a semantic network used to describe various entities,concepts and the relations among them existing in the real world.It developed from natural language processing,which can be used to query complex associative information at a higher level.Generally speaking,knowledge graphs in the real-world are sparse and incomplete,thus deriving the task of knowledge graph completion.Previous methods of knowledge graph completion mainly model and infer complex relations for 1-N,N-1 and N-N.The relations in knowledge graphs,which generally have many relations but have few corresponding triples show long-tailed distributions.Meanwhile,the relations in knowledge graphs with more relations but fewer corresponding triples generally meet long-tail distribution,which leads to unsatisfactory results of knowledge graph completion.The few-shot learning is thus introduced into knowledge graph completion models.The few-shot knowledge graph completion method,which predicts the missing entities based on incomplete triples with only a small number of corresponding entity pairs,not only relaxes the data volume requirement,but also solves the problem of long-tail distribution of knowledge graph relations.However,the fact is that different neighboring entities and their corresponding neighboring relations have strong and weak influences on entity and relationship representations,and their contributions are also different,and the treatment of the same weights necessarily affects the accuracy of entity/relation representations and thus the correctness of the triplet complementation.In addition,facing the scenario of sparse neighborhoods,the models utilizing multi-hop neighbor to enhance the semantic representation of few-shot relations may amplify the noise among the neighbor information,which will reduce the accuracy of knowledge graph completion.In order to solve the problems above,this paper proposes a few-shot knowledge graph completion based on the attention mechanism named as FKGCA.The model mainly applicable to few-shot relations scenarios,utilizes the Attention mechanism and Graph Neural Network method.It consists of three modules.The Neighborhood Encoder is used to learn the neighbor representations of entity through the Attention mechanism.The Relation Aggregator undertakes the neighbor representations of relation learning through Transformer.The Matching Processor works for matching triples through MTransH to complete the task of few-shot knowledge graph completion.The main contributions of this paper are obtained as follows:1)In this paper,the typical models of knowledge graph completion is analyzed,and the method of constructing a few-shot knowledge graph completion model is studied based on the Attention mechanism and Graph Neural Network method.The purpose is to obtain the relational embedding with neighborhood information so that the model can achieve accurate and efficient prediction.2)The graph attention network is applied to fuse the one-hop neighborhood information of entities to enhance the semantic representation of entities,while introducing a gating mechanism to reduce the impact of noise from sparse neighborhoods.3)Transformer is implemented to build the Relation Aggregator used to aggregate the information of neighborhood,which is conducive to obtaining relational representation and achieving accurate representation of few-shot relations.4)Experiments are conducted on two datasets in this paper.The results show that the models proposed outperform the traditional baseline algorithm,and obtain optimal and suboptimal results compared with the FKGC model as well.The superiority of FKGCA model proposed for FKGC is proved.