Is it possible to feed embeddings from XLM- RoBERTa to transformer seq2seq model? I'm working on NMT that translates verbal language sentences to sign language sentences (e.g Input: He sells food. Output (sign language sentence): Food he sells). But I have a very small dataset of sentence pairs - around 1000. And the language is a low resource language.

I am a new researcher on the field of deep learning. Please help with your valuable advice.


1 Answer 1


It is indeed possible, but the question is if it is a good idea.

FairSeq already contains a pre-trained XLM-R model, you can use by creating a new model: just copy the most suitable existing one and replace the encoder with XLM-R.

Another option would be using Huggingface's Transofrmers that also provides basic support for sequence-to-sequence models as they show in their blog post.

Now, why I don't think it is a good idea. Recent papers show that:

  1. Improving MT with pre-trained Transformers requires a lot of effort and it is still questionable whether it pays off, see Incorporating BERT into Neural Machine Translation .

  2. The encoder of MT capture different structures: see [What does BERT look at?](What does BERT look at?) and Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned.

  3. In an end-to-end trained translation system, a large part of the translation process is happening already in the encoder, see Analyzing Word Translation of Transformer Layers.

I assume you only have little training data and therefore you hope that pre-trained representation might help, but I believe you should put more effort in optimizing an end-to-end model to work in low-resource setup (have a look at Revisiting Low-Resource Neural Machine Translation: A Case Study) and data augmentation technique like iterative back-translation that it used in unsupervised machine translation.


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.