Meta is building an AI-powered universal language translator

Meta has said it’s working on a big new AI research project. The company wants to make translation software that works for “everyone in the world.” The project was announced at an event highlighting the benefits AI may bring to Meta’s metaverse goals. 

“While languages like English, Mandarin, and Spanish are well-served by current translation tools, about 20% of the world’s population doesn’t speak any of these languages”, the company claims. As a result, many of these under-served languages lack easily accessible corpuses of written material that are essential to train AI systems. 

The company will therefore address this issue by implementing new machine learning algorithms in two distinct areas. Building AI models that can learn to translate languages using a smaller number of training samples is the initial goal of the No Language Left Behind project. Another project dubbed Universal Speech Translator (UST) aims to create real-time translation systems that do not require textual translations between languages (a common technique for many translation apps).

Meta’s goal for No Language Left Behind is an ambitious one. Working with a smaller training sample means that the training samples used need to be of the highest quality possible. It is often difficult to source such data, especially for under-served languages with relatively few resources available online. While there are many different methods of obtaining linguistic data through open source means, it may be necessary to reach out to other industry sectors to get the data needed.

“The language industry works with this kind of data every day,” says Ofer Tirosh, CEO of global translation agency Tomedes. “Translation agencies also have access to linguists in resource-poor languages, who are necessary for creating the kind of high-quality data required for machine learning.”

There are a number of language service providers that specialize in creating and curating datasets for different languages, though at present the problem of scarcity remains for under-served languages. Tirosh says that translators in these languages would be crucial partners in developing the kind of data the No Language Left Behind project requires.

Importantly, Meta believes that such technology would greatly enhance this global solution, expanding its reach and transforming it into an indispensable communication tool for millions of people. As noted in a blog post, universal translation software would be a killer app for future wearable devices such as augmented reality glasses and would also help to break down borders in “immersive” virtual and augmented reality environments. In other words, while the development of global translation tools may have humanitarian benefits, it also makes good commercial sense for a company like Meta to pursue such a venture.

It is true that recent breakthroughs in machine learning have greatly increased machine translation speed and accuracy. From Google to Apple, prominent tech companies now give free AI translation tools for work and tourism, bringing immeasurable global benefits. In addition, critics claim that machine translation ignores important nuances, introduces gender bias, and makes strange, unanticipated errors only a computer can make. Some speakers of rare languages fear losing control of their language and culture if big tech controls the power to translate their words.

In the case of social media platforms like Facebook and Instagram, the importance of avoiding such mistakes cannot be overstated. Think of the 2017 arrest of a Palestinian man in Israel after a Facebook post he made was mistranslated by the company’s machine translation algorithms. “Good morning” in Arabic was translated by Facebook as “hurt them” in English and “attack them” in Hebrew, despite the fact that the man had written it in Arabic.

If Meta wants to achieve its goal of creating a worldwide translation system, it will have to show that its technology is up to the task and that it is able to apply its findings properly.