Out-of-Vocabulary words in released models about muse HOT 4 CLOSED

Hey,

No, the projection is working with as many word embeddings as you want. But it is just learned with a limited subset of them. What you can do is:

• train fastText in English
• train fastText in French
• learn the mapping between the monolingual embeddings
• given some new unknown words, get their embeddings with fastText (which is possible if you saved the fastText model)
• project these new embeddings

So you can totally use fastText and the mapping to deal with OOV, and even get a translation for OOV.

from muse.

I am wondering how to use the previously trained mapping to project a new embedding

Thanks!

edit:

Nvm, got it!

mapping = torch.nn.Linear(300, 300, bias=False)
to_reload = torch.from_numpy(torch.load('best_mapping.pth'))
mapping.weight.data.copy_(to_reload.type_as(mapping.weight.data))
translated_vector = mapping(vector)

from muse.

Oh right, that makes sense, thanks. I will check it out.

from muse.

I am wondering how to use the previously trained mapping to project a new embedding

Thanks!

edit:

Nvm, got it!

mapping = torch.nn.Linear(300, 300, bias=False)
to_reload = torch.from_numpy(torch.load('best_mapping.pth'))
mapping.weight.data.copy_(to_reload.type_as(mapping.weight.data))
translated_vector = mapping(vector)

can you send you code to me

from muse.