| With the acceleration of the global Internet development,a large amount of text data containing rich relation information is being generated.Automatic relation classification in text data becomes particularly important.Existing relation classification methods rely on large amounts of labeled training data,and are difficult to generalize on novel categories.Few-shot learning can quickly generalize on novel categories by learning a small number of labeled samples,which draw researchers' attentions.We studied the existing few-shot relation classification algorithms and analyzed their advantages and disadvantages.Graph Neural Networks(GNN)can pass information between samples of support set and query set,but it only models similarity base on a single aspect.Prototypical network is simple and efficient,but it does not consider the varying importance of support samples to different queries in the calculation of category prototypes.Meta learning method adjusts the parameters of the whole model to fit meta-task based on the support set,but is limited by the model size.Aiming at disadvantages of existing models,MGAPN,a Meta-info Graph Attention Prototypical Network is designed.Using multi-head product attention to calculate the massage passing between embedded vertexes enriches the relation representation between samples.Category prototypes are calculated dynamically on different query samples.Meta information produced by meta learning on classification weights is combined as label information.Finally,the intra-class inconsistency measurement in support set is integrated during training for further improvement of performance.Performance evaluations are conducted for our proposed model on different task settings in large-scale supervised few-shot relation classification datasets FewRel 1.0 and FewRel 2.0.Result shows that the performance of MGAPN is significantly better than existing models,and it surpasses the BERT-PAIR model when using bert-small as feature extractor. |
PDF Full Text Request