Global Mixup: Eliminating Ambiguity with Clustering

Data augmentation with Mixup has been proven an effective method to regularize the current deep neural networks. generates virtual samples and corresponding labels simultaneously by linear interpolation. However, one-stage generation paradigm use of interpolation have two defects: (1) The label generated sample is simply combined from original pairs without reasonable judgment, resulting in ambiguous labels. (2) Linear combination significantly restricts sampling space for generating samples. To address these issues, we propose a novel method, Global Mixup, based on global clustering relationships. Specifically, transform previous process into two-stage decoupling labeling. And samples, relabeling performed calculating relationships Furthermore, are no longer restricted relationships, which allows us generate more reliable larger space. Extensive experiments CNN, LSTM, BERT five tasks show that outperforms baselines. Further also demonstrate advantage low-resource scenarios.