Arabic and Urdu about multilingual_similarity_compare HOT 3 CLOSED

Hey @deepseek ,

For pure machine translation, those multilingual architectures are not really useful as in a machine translation scenario, we would need the decoder part as well. In our use case, we focus on the encoding part in a way that a sentence in arabic and its translation in urdu would have the same encoding in the semantic space. The quality of such encoding is what is evaluated in this repo.
Additional efforts have been done recently around languages such as Urdu with XLM-R.

What is your exact scenario? :)

N.B: Sorry I am currently having some time off so I would answer with some delay in the future.

from multilingual_similarity_compare.

Take this example, i have multiple books of the same author in Arabic and Urdu.
A new book is issued in Arabic, I want to translate it to Urdu.
i am sure that there must be a github repository that can translate such task.

from multilingual_similarity_compare.

Hey @deepseek

Yes there are but you should look at different architectures than BERT-like architectures which essentially just implements the encoding part.

Did you check with GPT-3 for example?

from multilingual_similarity_compare.