2
$\begingroup$

Can it be possible to train a neural network with no parallel bilingual data?

$\endgroup$
2
$\begingroup$

In this paper: Unsupervised Machine Translation Using Monolingual Corpora Only the authors proposed a novel method.

enter image description here

Intuitively it is an autoencoder, but the Start Of Sentence token is set to be the language type.

One other advanced method is to use the pre-training model. In this paper: Cross-lingual Language Model Pretraining researchers proposed an algorithm that utilized the pre-trained multi-lingual trained BERT(with labeled data but we don't need to have a labeled dataset for our task) and the autoencoder mentioned previously.

| improve this answer | |
$\endgroup$
  • $\begingroup$ That's crazy! I wonder if it could be a way of addressing the grounding problem too, if media experiences could somehow be mapped to a shared embedding space . . . $\endgroup$ – Neil Slater Feb 29 at 10:17
  • $\begingroup$ @NeilSlater I know only a little about the "grounding problem" but I thought this paper would be helpful to you: arxiv.org/pdf/1609.05518.pdf $\endgroup$ – Lerner Zhang Feb 29 at 10:27
  • $\begingroup$ Pre-training sounds like cheating. $\endgroup$ – nbro Feb 29 at 12:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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