Facebook publishes a translation feature they refer to as the “Many-to-Many multilingual machine translation” (MMT) model. While most translators use English data as a stopover, M2M – 100 does not take this step. So it is translated directly from Chinese into French.

According to Facebook, most of the training data is available in English, which is why previous models are about Chinese translate into English and from there into another language. Which accordingly creates an additional source of error. The new model can handle 100 languages ​​- in all directions. This is particularly important for the social network, since the news feeds are automatically converted into the language that the user has set. Two-thirds of the account holders are not English-speaking.

Billions of data for 100 languages ​​ On the rating scale for machine-translated texts, BLEU, M2M – 100 scored just as well as simple bilingual models and even, according to a Facebook blog post better than the English-centered models. 7.5 billion sentences in the 100 languages ​​have been used for this purpose, 15 billion parameters are used. The mass of data that has to flow in to enable the direct translation channels was one of the major difficulties. The training data required grew to the square: “If we need ten million sentence pairs in each direction, we need a billion sentence pairs for 10 languages ​​and 100 billions for 100 languages. “The data comes from the existing collections ccAligned, ccMatrix and Laser, among other things, whereby Facebook built Laser 2.0 from them as part of its work.

The code is available on Github. Fac