Alignment not coming properly


(sriram chaudhury) #1


I have trained a transformer base model for English- Hindi pair. On 300K steps, I infer the translation with “with_alignmet = hard” flag on.

The alignments are not coming properly.

For e.g.
The translated o/p: उसने भी राक्षस देखे । ||| 4-0 4-1 4-2 4-3 5-4

source sent: He too saw the monsters .

In above, alignment is totally wrong, but the translation is correct though.

Can you please help me in getting the correct alignments?


(Guillaume Klein) #2


Did you train with guided alignment? If not, the Transformer model can’t produce usable alignment on its own.