Unlocking cross-cultural communication with Unbabel’s João Graça
For the next two weeks, we will continue sharing our series of conversations with founders and technologists at Web Summit in Lisbon, Portugal. To hear more, subscribe to all those conversations on Apple Podcasts, Google Play, Spotify, or via RSS on your podcast app of choice.
* * * * *
Over the last several decades, several companies have tried to better connect people and societies by using artificial intelligence (AI) to provide automated translation services. While faster than human translation, these AI-driven translations can be imperfect.
Today, one startup is trying to provide language translation with both speed and accuracy by leveraging both AI and human translators. The company is Unbabel, founded in 2013 by CEO Vasco Pedro and CTO João Graça, with offices in Lisbon, New York, San Francisco, and Pittsburgh.
Graça sat down with Fernanda Baker from Samsung NEXT during Web Summit in Lisbon to discuss the future of real-time translation and what it means for businesses and consumers around the world.
“Unbabel has completely solved the problem that people speak different languages,” João said in the interview. The secret, João says, is machine learning combined with humans in the loop to fix errors in communications such as chats, help desk articles, and support tickets.
The genesis of a solution
João has a Ph.D. in machine learning, with a focus on natural language processing. “This is always an area that I’ve been very passionate about,” he said.
The idea for Unbabel came to João and his co-founder when they learned from friends who were renting out rooms how difficult it was to communicate with German-speaking guests. “If it was from Spain or Italy, they kind of understood the message. If it was from Germany, they had no clue.”
Available translation services were no help, and the idea for Unbabel was born. With initial funding from Y Combinator, the company started in Portugal and has been growing ever since.
Today Unbabel helps corporate customers handle customer service in multiple languages, for example, through chat on a website. When the text comes in, João explained, it first goes to machine translation, which is completely adapted to the particular customer.
From there, another machine learning system judges the quality of the resulting translation. If it gets a pass, it goes straight to the customer. If not, it goes to human translators for correction before going on to the customer within about 20 minutes. And all the while, the machine learning models get better and better as they get more data to chew on.
Enabling a new business model
“The solution we’re offering didn’t exist in the market before,” João said. “What people normally did was just hire people that spoke the language. This made it much harder because it’s hard to hire people that speak German, for instance, in the Philippines.”
From the beginning, integration with leading communication tools such as Zendesk and Salesforce was a priority for the Unbabel team. The goal, João said, was to make the translation process transparent and seamless to both support staff and the customers they support. The company generates revenue from subscriptions to its products.
“So you have a customer service agent,” João said. “You pay X amount of money, and now he can speak 45 languages. It’s like a super-powered agent.”
The biggest challenge Unbabel faces, João said, is getting the word out about such a novel offering. “People don’t know the solution exists,” he added.
The technical challenge of translation
Asked why automated translation has been so difficult to crack, João said: “Language is a living thing, so it’s very ambiguous. Phrases have words that mean completely different things. Like, the word ‘bank’ can mean four or five different semantic things.”
The inherently ambiguous nature of language means there’s no one right translation for any given communication. “You can ask 10 different people to translate a sentence,” João added, “and they can come up with 10 different equally good translations, which makes it much harder for algorithms to learn.”
Unbabel’s technology is getting a boost from partnerships with universities. “From the beginning, we have had a strong focus on research,” João said. “We have PhDs besides us working on the fundamental problems.”
The Unbabel team includes some 20 researchers with advanced degrees in machine learning and linguistics. Most of these work out of the Lisbon office, although the Pittsburgh office supports a team dedicated to AI research and development.
Looking ahead, the Holy Grail is for AI-driven translation to achieve human-level accuracy. “That’s the $1 million question that we keep getting asked by our investors,” João said.
For relatively simple communications such as email, AI will get there, João predicted. For more complex writing, such as blog posts, however, he said it won’t be any time soon.
Innovation yet to come
Despite the challenges, Unbabel has grown rapidly. Today the company has some 250 employees in its offices around the U.S. and in Portugal.
“Our vision is to be a translation layer that eliminates language barriers,” João said. “If Unbabel succeeds, you’ll be on your phone, on your device, anything. Just thinking, talking, writing on your language, and you don’t really care if the person on the other end speaks Chinese. Everything will be translated seamlessly.”
Want to keep up-to-date on all our upcoming podcasts? Subscribe to What’s NEXT on Apple Podcasts, Google Play, Spotify, or via RSS on your podcast app of choice. Also, sign up for our newsletter to get weekly updates on upcoming events and podcasts!