Facebook鈥檚 owner Meta has created an artificial intelligence model that can translate 204 written languages and has released it under an open source licence so that anyone can use or improve the software. The company claims that the AI supports more languages and provides higher-quality translations than world-leading software.
The model, called No Language Left Behind, supports dozens more text-based languages than , which currently works for 133, and , which caters for 110. The AI supports languages such as Acehnese, which is spoken by 4.7 million people, and Friulian, which is spoken by just 600,000 people.
Despite its name, No Language Left Behind covers only a small minority of the and doesn鈥檛 yet match Wikipedia, which has .
Advertisement
at Meta says the company will continue to add other languages. 鈥淎 lot of those languages are not spoken by a lot of people, and most of them don鈥檛 have written form,鈥� she says. 鈥淎nd so even though there are several thousand languages in the world, we estimate that only a few hundred really have standard writing systems. And so we focus on those first. This is just the starting point.鈥�
No Language Left Behind was developed using Meta鈥檚 new AI-specific supercomputer, called the AI Research SuperCluster (RSC). The machine is operational, but is still being added to, and when complete it will consist of 16,000 processors. Meta says at that point, it will be the fastest AI-optimised supercomputer in the world, performing at nearly 5 exaflops 鈥� meaning it can carry out 5 billion billion operations per second.
Fan says that although the AI model can run on less sophisticated hardware, the supercomputer鈥檚 power was vital for quickly training and testing iterations of the model. The completed model still requires more computer power than most people have on their desktops, so Meta has also released a smaller, slightly less capable model that is much less demanding.
The full No Language Left Behind model consists of 54 billion parameters, or individual stores of data used to calculate outputs, meaning it is much smaller than natural language AI models such as the Megatron-Turing Natural Language Generation (MT-NLG) model, which has almost 10 times as many parameters.
Meta will use No Language Left Behind in its own services, but because the code is open source, it can also be used by anyone else who chooses to implement it. Third parties can also submit improvements to the code, which can be integrated into the project.
It can be difficult to qualitatively compare the translated text from AI models, so Meta has also created an update to its existing translation benchmark, called FLORES-200, that evaluates the result of translating more than 40,000 standardised sections of text. The company claims that No Language Left Behind is 44 per cent better than Microsoft鈥檚 equivalent using this benchmark, and marginally better than Google Translate.听
