Abstract: Data Augmentation (DA), i.e., synthesizing faithful and diverse samples to expand the original training set, is an effective strategy to improve the performance of various data-scarce tasks.