Meta has released a massive new artificial intelligence (AI) model capable of translating between over 200 different languages.
Trained on one of the world's most powerful AI supercomputers, the No Language Left Behind (NLLB) model already supports advanced translation features across Meta's suite of applications and services.
"Advances here will enable more than 25 billion translations every day in our apps," said Mark Zuckerberg, CEO of Meta. “The AI modeling techniques we have used enable high-quality translations for languages spoken by billions of people around the world.
“Cross-language communication is a superpower that AI provides, but as we advance our AI work, it improves everything we do, from posting the most interesting content on Facebook and Instagram to recommending more relevant ads and making our service be safe. for everyone."
Cross the language barrier
In addition to supporting the Meta family of apps, releasing the template under an open source license, the company also paved the way for a new wave of third-party services designed to break the language barrier.
In a promotional video, Meta raised the possibility of augmented reality tools that translate cookbooks into foreign languages and apps capable of unlocking literature from "low-resource languages" such as Icelandic or Occitan.
“Language is the key to inclusion; if you don't understand what people are saying or writing, you can be left behind,” Jean Maillard, a research engineer at Meta, said in a post on camera.
Naturally, possible metaverse applications for the technology were also mentioned; Meta envisions a future where people of all backgrounds can converse freely in virtual reality, supported by live translation.
"The idea is to completely get rid of the language barrier, so that everyone understands your experience without changing the way you communicate," added Necip Fazil Ayan, research director at Meta. "I think the metaverse is going to be where all these things come together."
The key to achieving these goals is the ability of a single AI model to cover a large number of languages. Comprised of 50 billion parameters (the more the better), NLLB can translate almost twice as many languages as the next best performing models, but Meta aims for even greater heights.
The techniques employed to achieve this level of performance are detailed in a research paper (opens in a new tab) published by Meta, which describes "multiple architectural and training enhancements" that it expects "will be important to the realization of a universal translation". .