What if you could train task-specific NLP models without a single human-labeled example? "Towards Zero-Label Language Learning" leans hard into that question. The trick is Unsupervised Data Generation (UDG), which uses few-shot prompts on a large pretrained language model like GPT-3 to synthesize its own training set, then fine-tunes a downstream model purely on that synthetic data.

This long review walks through the UDG pipeline in detail and its surprising results: models trained only on synthetic data match or beat baselines trained on real human annotations, and when you mix synthetic data with labeled data as augmentation, the approach sets new state-of-the-art numbers on SuperGLUE. For speech and voice AI teams, the implications are direct, this is the playbook for bootstrapping labels for intent classification, semantic parsing, and any downstream NLU task where human annotation is the bottleneck. The deep-dive walks through every piece if you're planning your next synthetic data pipeline.