CSLS calculation understanding about muse HOT 4 CLOSED

Because there is an option dico_threshold that looks at the difference between the score of the best and second best translation, and will only take translation pairs where this difference is above the threshold, i.e. there is no ambiguity and the best translation is clearly better than the second best. For this we need to select the best 2 scores, i.e. we concat all top2 words from each batch :)

But this is just an option which we do not use in practice.

from muse.

Hi,

The 128 here is just the batch size. For 128 source words we look at their best translations over all target words. We do that batch by batch because otherwise it involves to create a matrix of size (n_src, n_tgt) in memory (where n_src and n_tgt are the numbers of source and target words) and this will not fit for large vocabularies.

from muse.

I see. that makes sense. In that case, I wonder what is the reason to pickup the top 2 best scores, I mean to make candidate pairs, top 1 and its self should work already. Thank you!

from muse.

I see. Thank you for elaboration!

from muse.