New Thinking

Machine Translation in the Age of Displacement

Share this post:

Just a few years before the launch of Unbabel Inc., a global machine translation agency based in Portugal, the world was a very different place. In 2009, Greece conceded its €300 billion debt, sounding the alarm bell on a crisis in the Eurozone. Within two years, social media communication assisted in fomenting the wide-scale protests that gave rise to the Arab Spring. Those demonstrations sparked a political firestorm that began to displace Syrians in even greater numbers than those driven out by war in Afghanistan and Iraq. And life online looked different, too: It wasn’t so long ago that a person in need might turn to Babel Fish, that Yahoo!-owned translation engine of yore.

“We always had the vision that we wanted to use Unbabel for something more than just selling translation,” says João Graça, CTO and co-founder of Unbabel. The company, which came of age in the midst of the most critical years of the European migrant crisis (it was founded in 2013), and which now provides AI-powered translations edited by a network of 50,000 individuals around the world, has had plenty of opportunity to apply their insights to people in need.

Language may be one of AI’s toughest nuts to crack: It changes rapidly, its meaning is transformed based on culture and context, its rules are fluid, and equivalencies between languages can be evasive. But advancements in machine translation have made headlines in recent years thanks to tech behemoths like Facebook and Google, who have helped to advance the capabilities of the technology in building their own platforms. Google unveiled a major upgrade to its automated translation system last fall with Google Neural Machine Translation, or GNMT. And in August, Facebook announced that it, too, had completed a transition to neural machine translation, a system the company developed internally. (Translations on the social-media platform were powered by Microsoft search engine Bing until 2015.) Neural machine translation engines may have the potential to churn out much more reliable translations than rule-based or statistical machine translation systems because they use neural networks to decode whole texts rather than disparate fragments. These companies’ products also employ Deep Learning techniques to provide better, more accurate translations over time.

Several researchers dedicated to machine translation development also sit behind the wheel at Unbabel, including Graça, who holds a PhD in Natural Language Processing and Machine Learning from Portugal’s Instituto Superior Técnico and the University of Pennsylvania. Unbabel’s team of some 70 employees orchestrates its massive network of independent contractors to sift through translation projects from around the world, for clients among the likes of Pinterest, Skyscanner, and The company’s business proposition rests on the vast potential for globalization to boost e-commerce: “If you want to go global, you’ve got to get local,” Unbabel’s homepage stresses. “Reach your customers in their native language today.” Unbabel translators are vetted with a language skill test and regularly re-evaluated and trained under the auspices of the company, and they receive weekly payouts to PayPal and Skrill. Today, clients can request translation of English texts into 28 different languages.

Unbabel, which adopted neural machine translation early this year, is unabashed about the fact that its services are a product of human–AI collaboration. The input of its Unbabelers—as the company calls its editors and translators—is used to continually improve and expedite its automated translation. “I think we’re going to be getting much better and keep [improving],” Graça says, “but there’s always going to be need for humans, in different ways.” Human Unbabelers will gauge a client’s writing style and tone and adapt a text to their particular terminology, which can help to train the AI for future work, and also examine the context in which the translation will be used. “If you train a system for customer service, or a particular domain, that system will not be a good for another domain, and if you just have an open-channel domain, it will never be good enough by the simple fact that some terms are different for different domains,” Graça says. He stresses the importance of effectively and efficiently sorting tasks for machine versus human translation: “Instead of just going with pure [machine translation], you need to have a combination that knows when machine translation is enough and when it’s not enough,” he says.

Unbabel’s humanitarian interests date back to its earliest days. In response to the Eurozone crisis, Graça and his colleagues launched Unbabel News, which sought to provide translated news articles to the public for free, with the aim of illuminating the ways in which people in different countries were interpreting the crisis in different ways. At the time, however, Unbabel “didn’t have the manpower” to meet the company’s own quality standards, Graça says. When the Ebola epidemic hit West Africa, Unbabel began to translate information about the virus that had only been available in English, until the crisis subsided and that project, too, fell by the wayside. “But we kept thinking… ‘How can we use our community to serve other purposes?’” Graça recalls.

Most recently, Graça and his team launched, the company’s nonprofit arm, which provides free translation services to NGOs supporting refugees. The new initiative works the same way as its commercial service, but it allows the 50,000 Unbabelers to select and work for pro-bono for organizations that are important to them. Unbabel also supports INTERACT, the International Network on Crisis Translation, a research project funded with a grant from the European Union. INTERACT seeks to provide language support for people in crisis situations, with a focus on four main topics: policies, training, simplification, and machine translation. Translators Without Borders, whose “Words of Relief” project was deployed to assist victims of Ebola and of the Nepal earthquake in 2015, also supports the project. “A lot of times [a crisis] happens in parts of the word where you don’t actually have good machine translation engines, because most machine translation engines that work well are made for languages where there’s actually…transactions involved, and there’s a commercial part,” Graça says, citing the need to develop support in the languages native to these regions.

Efforts in the global language-tech community to provide financial assistance and services for refugees extend beyond those companies working with machine translation. A spate of relocation apps connects migrants with information and services to help them get proper documentation, find healthcare and employment, and learn local languages or connect with interpreters. Tarjimly, an app founded by three Muslim Americans that pairs refugees with volunteer translators through Facebook Messenger, launched this February. Babbel, a paid language-learning app founded in 2007 and headquartered in Berlin, has supported migrants in Germany by partnering with local NGOs to provide refugees with the Babbel app for free and training volunteer German instructors to teach students without a common language. Christian Hillemeyer, Director of Communications at Babbel, credits the ubiquity of one key piece of hardware with getting real help into the hands of migrants: “Basically every refugee has a smartphone,” he says. “You have to see: The smartphone might be the most important invention since the wheel.”

From his office in Lisbon, Graça described his own sprawling ambitions for the future of Unbabel and for the technology the company has developed and continues to refine. He still nurses a desire to revive Unbabel News, he says, and as a contributor to INTERACT, to developing global translation support that people can put to use the moment a crisis hits. But each of Graça’s goals for Unbabel come down to one main objective—to increase translation capacity by reducing human effort. And the continued improvement of the company’s technology rests on its continued use, be it for commercial translation or in service of NGOs. “We want to be the translation leader of the world, so everything that has to be translated will come by us and have professional quality,” Graça says. “So there’s still a lot of things to do to get there.”

More stories

Prediction Machines: To Fight Misconceptions about Artificial Intelligence, Clarity is Key

Cheap, ubiquitous, and incredibly valuable—that’s how the economists behind a new book about AI see the future of prediction, thanks to the technology that helps fill in the blanks. However, misperceptions about what AI is and how it works continue to prevail, even among tech-savvy marketers. In order for AI technology to achieve its full […]

Continue reading

Changing the World, One Website at a Time

Corporate social responsibility is important. Very important. With the ability for widespread internal communications and the advantage of robust organizational structures, corporations are poised to help in a big way. And it seems the timing is right to double down on CSR for two reasons: Consumers want to support brands with charitable missions (see: Toms […]

Continue reading

Thomas Marckx, Co-founder & Solutions Architect, TheLedger

Fireside chat at THINK 2018 with Thomas Marckx, Co-founder & Solutions Architect, TheLedger — moderated by Robert Schwartz, Global Leader of Agency Services, IBM iX. Subscribe to IBM thinkLeaders! Learn more about IBM iX

Continue reading