Closed LeeSureman closed 3 years ago
I almost figure it out. fairseq will set it as the bos of input in collator.
You are right, language token is added at the end of the target sequence. But then fairseq right shift the target sequence by 1 token which brings the language token at the beginning. So, the paper provides the correct information.
I find that in multilingual denoising task of fairseq , the target's language token is at the end, which is not like the table3 in your paper, where the language token is at the start of target. am I wrong?![image](https://user-images.githubusercontent.com/32452367/129170734-897369dc-96ab-4ba8-bc7f-c172640444fc.png)