Facebook has developed a new machine translation model open source, M2M-100, to avoid using English to switch from one language to another. This would be the first system capable of directly translating any pair from a set of 100 languages.
7.5 billion sentence pairs collected
Facebook must have gathered 7.5 billion phrase pairs collected from the web. It was then necessary to identify them and translate them by creating a mathematical representation of the texts to retain the sentences that say the same thing in two different languages.
The social network provides around 20 billion translations every day for its news feed. But even today, he goes through the traditional intermediate stage from English to Chinese to French, for example. The problem is that it affects the accuracy and fluidity of the translation. Facebook hopes that one day it will be able to use its M2M-100 system, which it believes performs better, and encourages other researchers to participate in its improvement.