Unsupervised Neural Machine Translation with Generative Language Models Only

We show how to derive state-of-the-art unsupervised neural machine translation systems from generatively pre-trained language models. Our method consists of three steps: few-shot amplification, distillation, and backtranslation. We first use the zero-shot translation ability of large pre-trained language models to generate translations for a small set of unlabeled sentences. We then amplify these

[Ryobot]

“we achieve a new state-of-the-art in unsupervised translation on the WMT14 English-French benchmark, attaining a BLEU score of 42.1” 教師あり学習が得意とする分野ですら教師なし学習が競争力のある性能を出してきた