Unsupervised Data Augmentation (UDA) is a semi-supervised technique that\napplies a consistency loss to penalize differences between a model's\npredictions on (a) observed (unlabeled) examples; and (b) corresponding\n'noised' examples produced via data augmentation. While UDA has gained\npopularity for text classification, open questions linger over which design\ndecisions are necessary and over how to extend the method to sequence labeling\ntasks. This method has recently gained traction for text classification. In\nthis paper, we re-examine UDA and demonstrate its efficacy on several\nsequential tasks. Our main contribution is an empirical study of UDA to\nestablish which components of the algorithm confer benefits in NLP. Notably,\nalthough prior work has emphasized the use of clever augmentation techniques\nincluding back-translation, we find that enforcing consistency between\npredictions assigned to observed and randomly substituted words often yields\ncomparable (or greater) benefits compared to these complex perturbation models.\nFurthermore, we find that applying its consistency loss affords meaningful\ngains without any unlabeled data at all, i.e., in a standard supervised\nsetting. In short: UDA need not be unsupervised, and does not require complex\ndata augmentation to be effective.\n