Does it make sense to run the adversarial training on the already-aligned vector spaces?

[bdqnghi]

Supposed that I have already achieved the aligned vector spaces from another approach rather than MUSE, does it make sense to run those already-aligned vector spaces again on the adversarial training proposed in MUSE? Will it make the vector spaces to be more aligned? which can produce better results.

[glample]

Yes you can do this. But in your case you can directly start with the refinement procedure. It is much faster, and works better, as long as there is an initial embedding alignment that allows to get a reasonable accuracy. For instance, on English-French, starting the refinement from an adversarial alignment that gives 30% or 60% will end up giving the same final accuracy of around 80% P@1.

[bdqnghi]

i see, thanks