Since back translation English->other language -> English seems like quite a useful data augmentation technique , I wanted to experiment with it. E.g. it occurred to me that languages from very different language families (but very well supported for economic reasons such as Chinese, Russian, Spanish, Korean, Arabic...) could make for a diverse set of effects occurring in the back translation.

Commercial translation APIs would be a straightforward way of doing this, but without free API key or budget from my organization (would not qualify as academic) that's quickly quite expensive for a private thing.

Pretrained translation models would seem like an obvious alternative (I have a GPU for inference, but clearly that's not enough to train all the models from scratch), but I could e.g. not find those for any OpenNMT variant. Are there any recommendations from others that have used this approach?


1 Answer 1


One option, which I have discovered, is back-translation via the Unsupervised Data Augmentation repository made public by Google Research. This is based on this paper.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.