Abstract
Semi-Supervised Learning (SSL) has seen success in many application domains, but this success often hinges on the availability of task-specific unlabeled data. Knowledge distillation (KD) has enabled effective optimization of compact neural nets, achieving the best results when the knowledge of an expensive network is distilled via fresh task-specific unlabeled data. However, task-specific unlabeled data can be challenging to find, especially for NLP. We investigate the use of generative models in synthesizing unlabeled data and present a simple and general framework called "generate, annotate, and learn (GAL)". A language model (LM) is used to synthesize in-domain unlabeled data. Then, a classifier is used to annotate such data. Finally, synthetically generated and annotated data is used to advance SSL, KD, and few-shot learning on NLP and tabular tasks. To obtain a strong task-specific LM, we either fine-tune a large LM on inputs from a specific task, or prompt a large LM with a few input examples and conditionally generate more unlabeled examples. It also yields a new state-of-the-art for 6-layer transformers on the GLUE leaderboard. Finally, self-training with GAL offers large gains on four tabular tasks from the UCI repository.