Dynamic matching-prototypical learning for noisy few-shot relation classification

Few-shot relation classification (FSRC) targets at tackling the long-tail relation problem by identifying relations between entity pairs with few labeled instances. Most of exiting approaches are difficult to deal with text noise in the real world and complex scenes owing to the few labeled sample instances available for learning and the diversity of text expressions. Moreover, they are constrained by limited capacity to extract text features, thus ignoring the proximity of similar relations and the distinctiveness of different relations in the semantic space. Herein, we propose an innovative D ynamic M atching-Prototypical N etwork (DMPN), which equips with three mechanisms for noisy FSRC. In particular, DMPN adopts an I nstance-Guided M atching (IGM) module and a D imension-Guided A ttention (DGA) module to eliminate the deviation of prototype calculation, and dynamically adjusts the contributions of different instances and dimensional features in the support set. Additionally, a two-stage R elation-Aware T raining (RAT) method is designed to focus on noisy relations and improve the discrimination of relation representations. Extensive experiments conduct on the FewRel dataset demonstrate that our model outperform other competitive baselines. Specifically, in the noiseless scenario and under three different noise rate settings, DMPN achieves an average accuracy of 84.71%/80.81%/74.43%/66.81% respectively, which improves from 2.26% to 6.38% on average compared to all baseline methods.