Facebook, illustration -
Igor Golovniov / SOPA Images / Sip / SIPA
Facebook has developed a new artificial intelligence, called M2M-100, capable of translating from or into a hundred different languages without using English.
This language was indeed until now a mandatory intermediate step for most of the 20 billion translations made every day by the social network on the home page of each of its users.
An automatic translation from Chinese to French actually consisted of two successive operations, for example the company explained on its blog on Monday.
We're introducing M2M-100, the first multilingual machine translation model that translates between any pair of 100 languages without relying on English data.
We've open sourced the model, training, & evaluation set up.
Learn more https://t.co/9nszUF5nTj # t9n #machinetranslation pic.twitter.com/57kqbParp1
- Facebook AI (@facebookai) October 19, 2020
No need for English anymore
Previous automated tools worked by switching from Chinese to English and then from English to French.
These successive stages were due to the preponderant presence of English among the online content analyzed by the algorithms to enrich their knowledge of translation.
However, they caused a loss of quality and fluidity, while complicating the process.
The accuracy of results obtained by M2M-100 is now higher than that observed in machine translations performed by previous programs.
A base of 7.5 billion sentences
To achieve their ends, the engineers of the Californian giant have created a database containing 7.5 billion sentences in the 100 languages concerned.
The algorithm was then refined and configured using 15 billion parameters that "take into account a greater diversity of structure in terms of language and morphology," writes Angela Fan, AI researcher at Facebook.
Only a third of the billions of messages published daily by Facebook users are in English, specifies the specialist.
The data access and tools used by the company to create M2M-100 have been released as open source.
Google Translate: A new tool to transcribe and translate speeches live
An artificial intelligence capable of spotting errors on Wikipedia and correcting them