Image: © Haris/Stock.adobe.com
Meta’s AI model was designed to bring machine translation to low-resource languages and is now being used to help Wikipedia editors.
Facebook parent company Meta has develop an AI model that can translate 200 different languages.
Meta said its natural language processing (NLP) model is the first to be able to translate across so many languages, with a particular focus on African languages.
A handful of languages dominate the internet, which means only a fraction of the world can access content and contribute to the web in their own language. But machine translation models require lots of data for training, making it difficult to develop tools for languages that have fewer written examples online.
Meta CEO Mark Zuckerberg said the company’s new NLP model will allow “25bn translations every day” across the company’s apps.
“We call this project No Language Left Behind, and the AI modelling techniques we used are helping make high-quality translations for languages spoken by billions of people around the world,” he added.
Zuckerberg said the AI model, called NLLB-200, has more than 50bn parameters and was trained on the company’s AI supercomputer called Research SuperCluster.
Meta first shared details of the No Language Left Behind project in February, when the company showcased its AI research focus for the year ahead. This also included ambitious plans for a Universal Speech Translator to better support languages that lack a standardised writing system.
Meta said the new NLLB-200 model can translate 55 African languages, including many low-resource languages that have few written examples available online and are not typically supported by machine translation tools.
The tech giant said it has worked with professional translators for each language so it can develop a reliable benchmark and automatically assess the translation quality of low-resource languages.
Meta has released a research paper on the AI model and said it will provide tools for other researchers to extend the work to other languages.
It is open-sourcing its model and said lessons from the project are now being applied to translation systems used by Wikipedia editors.
‘Unwieldy’ models
Victor Botev, CTO of AI research start-up Iris.ai, said the engineering prowess needed to present enough data for these obscure datasets is a “marvel”.
However, he warned that these types of AI models are not necessarily the “cure-all” they may appear to be, as they can struggle with specific tasks due to their size.
“The models that Meta uses are massive, unwieldy beasts,” Botev said. “So, when you get into the minutiae of individualised use cases, they can easily find themselves out of their depth – overgeneralised and incapable of performing the specific tasks required of them.”
Botev added that the validity of Meta’s measurements have not yet been scientifically proven and its research has not been published for peer review.
“Doing a kind of peer review through Meta’s media publication creates bias for future reviews and puts public pressure on the reviewers. But despite all of this, I’m hoping that these points will be addressed and it will be a good foundation for some great work in the next few months in NLP.”
10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.
