In automated text translation, Facebook — now Meta — leads the way. On July 8, 2022, the social media giant (which also owns Instagram and WhatsApp) announced that it had developed an artificial intelligence program capable of supporting the translation of 200 languages. This is double the previous record, held by a Microsoft algorithm. Meta’s numerical model, called NLLB-200 (like “No Language Left Behind”), takes into account rare languages such as Lao (spoken in Laos), Kamba, Fulani or Lingala (spoken in Africa). Little or not integrated with previous translation software, they are nevertheless used by millions of speakers. No less than 45 million people speak Lingala in the DRC (Democratic Republic of Congo), Republic of Congo, Central African Republic and South Sudan. However, there are – notes Meta – only 3260 Wikipedia articles in Lingala; to be compared, for example, with the 2.5 million articles in Swedish, a language used by “only” 10 million people in Sweden and Finland! This is what NLLB-200 is all about: offering better access to Web content to billions of people who until now had no access to it because of the language barrier. If Meta offers its program in free access, which allows any actor to seize it, the company created by Mark Zuckerberg has an immediate interest in developing this type of service, Facebook making nearly 20 billion daily translations on his newsfeed… Antoine Bordes, director of the Paris center in artificial intelligence of Meta (FAIR, Facebook Artificial Intelligence Research), laboratory whose researchers were in the front line to develop NLLB-200, answers questions from Science and Future on the development of this model, the culmination of six years of research on translation by artificial intelligence. And evokes the way in which, in the future, it could fit into the scenario of the metaverse.
“A thousand billion parameters!”
Sciences et Avenir: the NLLB-200 program is backed by a supercomputer. Why ?
Antoine Bordes: To ensure the automated translation of 200 languages, a considerable amount of data must be handled. On the order of a trillion parameters! It’s huge, even if this data can be categorized in three different ways.
What are these three types of data?
First, there is — and this is the minority part — translation data. Basically, all existing translations in the public domain of the languages we are dealing with have been put into our model. These are texts translated by humans.
Then there is the “monolingual” text, untranslated, which only exists in its original version. It can be texts in English, French, Italian… but also in Zulu, Assamese or Bengali (respectively languages of southern Africa, Indo-European and spoken in Bangladesh, editor’s note). We will look for them on the Internet to feed the model; moreover, there is then the issue of knowing which language is in question!
Finally, the third type of data corresponds to the process created by Meta and whose code is in open source : Laser 3 will scour the web for phrases in different languages that mean the same thing. We are talking here about translated texts in which matches have been found automatically. I insist: these are not translations written by translators but “approximate parallel texts” that our system evaluates by saying: “Yes, it’s the same thing”. Think of press dispatches: when an event of international scope occurs, it is covered all over the world. Also the texts of the journalists will integrate sentences or expressions which one will find in all the languages, and between which Laser 3 can establish correspondences.
“The scale of the data is that of the multilingual Web”
So the NLLB-200 model is powered by real translations made by humans, monolingual texts and approximate translations?
Yes, the scale of the data is that of the multilingual Web. Hence the need to use a huge calculator to have a model that is capable of ingesting all this, milling, learning, and ultimately translating more languages first, but also in a more qualitative. There is indeed an average 44% improvement on the languages already covered by automated translation.
This type of estimate, to sanction the quality of the translation, is made by an evaluation tool called FLORES-200. However, it was also developed by Facebook’s AI lab. Aren’t you judge and party?
First, it is not a desire on our part to design the assessment tool ourselves, but a necessity. There was simply no such device covering rare languages. This is complex to implement: you have to find speakers capable of translating. We had to hire translators who worked in two separate teams: one produces the translation of a text, the other judges the translation thus produced. It’s a way to remove bias. But, indeed, Meta designs the validation tool and also the model: aren’t we congratulating ourselves? The question is legitimate. We respond to this by making everything “open source”. Everything is published, it is our justice of the peace. Our scientific paper is huge, it’s full of details, but it’s available. Meta is really open to improvement and criticism in a purely scientific way, in terms of reproducibility and peer review.
“The Web will evolve from 2D to 3D”
How does automated translation fit into the scenario of the metaverse, this evolution of the Web driven by Facebook and Meta?
Very centrally. We want an inclusive metaverse that is a source of opportunity for as many people as possible. One of the levers is to break the language barrier. The metaverse must be polyglot in essence. Imagine participating in a virtual meeting in sub-Saharan Africa. With automated translation, it is possible to speak to everyone even without practicing the language mainly used by the participants. The metaverse must create this environment where we can talk, exchange, dialogue — including giving a chance to people who don’t speak English well; after all, 80% of the information available on the Web is in this language!
NLLB-200 is oriented towards written text. But Méta has another, related project: the Universal speech translator, the translation of “voice to voice” in real time. By keeping the accents, the silences, the prosody… everything that makes the conversation. The two projects are entirely complementary — their teams work closely together — so on the one hand we are extending the number of languages translated in an automated way, on the other we are looking at the subject of “speech to speech”. In the end, in the longer term no doubt, we could be able to do 200 languages in “voice to voice” in real time. This will be the metaverse.
Meta is very voluntary on the subject… but aren’t you the only ones to believe it?
I don’t think so at all. Take a look at the recent Vivatech show: it wasn’t just the Meta stand that was talking about metaverses, far from it! If the challenge is to say that we risk finding ourselves alone there, I am not worried at all. Many players consider that the Web will evolve from 2D to 3D, with an immersive side that is decisive for issues of office automation, entertainment, work or gaming. We are convinced that this is the future of the Internet, and in particular the future of communication, social projection and online relationships. We believe in it, and we go there.
–
