Research And Implementation Of Small Sample Knowledge Graph Completion Based On GA

With the rapid development of big data and mobile internet industries,more and more heterogeneous network information resources need to be researched and developed.The emergence of knowledge graph(KG)provides effective technical support for the extraction,storage and utilization of heterogeneous network information resources.The knowledge graph plays an important role in many downstream applications.However,the knowledge graph is still incomplete.In order to make the knowledge graph cover more information,it is essential to add entities and relationships to the knowledge graph to complete the knowledge graph.Because there are a great quantity of long-tailed relationships in the knowledge graph,the knowledge graph completion method based on embedding needs to provide enough training examples for each relationship,so it performs poorly in the relationship with only a small number of associated triples.At present,many researchers have begun to use the Graph Convolution Network to encode the local graph structure in the few-shot knowledge graph.However,the existing works usually average the contribution of each neighbor node to the target entity,which limits the effect of link prediction to some extent.To solve this problem,this paper proposes a few-shot knowledge graph completion method based on Graph Attention Network(GAT).The main works of this paper are as follows:(1)In this paper,we introduce the graph attention network into the few-shot knowledge graph completion task using the matching metric method,and propose a few-shot knowledge graph completion method GFKGC based on GAT.This method captures the feature representation of the local neighborhood through the feature embedding method of GAT,assigns different weights to different neighbors through its attention mechanism,generates stronger feature representation,completes candidate entities to the query set and the reference set for similarity measurement,and obtains the similarity between the query set and the reference set through the matching network,select the candidate entity with the highest similarity measurement score as the completed tail entity to complete the knowledge graph.In this paper,we conducted link prediction experiments on two few-shot datasets,Nell-one and FB15k237-one,and analyzed the number of reference triplets in the reference set with the best effect for the completion of the few-shot knowledge graph.The experimental results show that the few-shot knowledge graph completion model based on GAT proposed in this paper can effectively predict the missing links in the few-shot knowledge graph.(2)Design and implement a knowledge graph question answering system for a specific domain.In the question answering system,we study the problem of missing answers when users retrieve answers in the system.For incomplete triples,we use the few-shot knowledge graph completion model GFKGC proposed in this paper to complete the triple,in order to improve the accuracy of knowledge question answering,and visually display the question answering effect before and after the completion of the knowledge graph to users.The system uses crawler and other technologies to collect data,and stores the data in the graph database Neo4 j to facilitate the user to query the answer.For the questions raised by users,first,the question is analyzed through the question preprocessing module,then the question triples are generated,and then the query statements are generated through the answer generation module to retrieve and obtain the answers in the knowledge graph.Finally,return the answer to the user.