Automatic translation between nearly all common languages could soon be a possibility, thanks to a “superpower” new service in development at Facebook’s parent company, Meta.
A new paper from Meta suggests that scientists have put the groundwork together for a production-level, universal translation service. The 38 academics working on the paper say that not only could it translate more than 200 languages, but be open to everybody – fostering a new wave of collaborative work and research.
Using machine learning, the researchers say they’d be able to produce quality translations of 204 languages – double that of the next best options currently in use. These volumes can be reached thanks to the machine learning element, which doesn’t require languages to be so prevalent online for its source material – a factor that has hampered previous efforts. The result is a workable translation for even the languages of the Chokwe people of Central and Southern Africa and the Acehnese people of Indonesia, among many others.
Perhaps more important than the number of new languages open to translation is the quality of output. According to researchers, the new AI could be as much as 44% more accurate with its translations than current tools, with a more sensitive and sophisticated system mimicking particular patterns and phrases better than the more ‘by rote’ methods of old.
The development may be new and exciting, but the science behind it is somewhat less so, with the so-called Sparsely Gated Mixture of Experts method having been used in academic research for at least five years already. What impressed data scientists most, though, was the feat of engineering required to cleanse and process whole language datasets that don’t appear in large volumes online.
Facebook founder and chief Mark Zuckerberg called this new AI tool a “superpower”, whilst one of the researchers involved said it was a “pretty big deal”. However, the academics remained keen to set expectations, reaffirming that this was just the first step on a long and winding road. Their AI tools, the researchers said, providing the groundwork, but the process would be an ongoing one – and despite excitement, this remained a work in progress rather than a project with a firm (and imminent) end point.
