Thank you for the great work. May I ask what features are input into the back translation model?
As to the PHOENIX2014T dataset, I input your features directly (i.e. seqlen * 150) into the Sign Language Transformers for back translation, and I get the BLEU4 score of 12, which is not good enough.