Question About Decoder about xlm HOT 3 CLOSED

Yes, you are correct. The only difference between en->en and en->fr is the language embedding given as input to the decoder.

from xlm.

I'm not sure I understand the question. The setup here is the same as in UNMT, the language ID defines what language the model is encoding or decoding.
In UNMT it was a bit different, the language ID was only used to define the start symbol (i.e. <s_en>, <s_fr>, etc.) while here there is just one start symbol <s> and a language embedding is added at every timestep.

from xlm.

@glample Sorry for my poor expression in English. I want to ensure that the only difference between in DAE step and Mt step is the language embedding. For example ,to train en->en, decoder will add en language embedding. For en->fr, deocder will add fr language embedding. Is this the only difference and does language embedding play the crucial role in decode direction?

from xlm.