Transfer learning works for NLP just as it does for computer vision — pre-train once, fine-tune cheaply for any text task. This paper helped catalyze the pre-training revolution.
Proposes ULMFiT, a transfer learning method for NLP that pre-trains a language model on a general corpus and fine-tunes it on target tasks using discriminative fine-tuning and gradual unfreezing. Achieves strong results with very few labeled examples.