How is language id embedding is learned?

[mani-rai]

I have been reading this paper to replicate the architecture for my unsupported language pretraining. I was able to grasp everything except learning of language id embedding. Can anyone tell me how embedding is learned? I am assuming it is learned during training as well. Is this true?

[hoschwenk]

Which language ID embedding you are referring to ?
The encoder does not use an LID, only the decoder, to know whether it should generate English or Spanish.
It is learned as a parameter as the rest of the network.

[loretoparisi]

@hoschwenk regarding LID / LASER I have a clue, and maybe you and the right answer
facebookresearch/fastText#1345

Thanks in advance!