English may be the world's most widely spoken language but the UK's monolingual population is stunting its growth. Government statistics estimate that the country currently loses around 3.5 percent of its GDP every year because of a lack of language skills.

Lisbon-based startup Unbabel thinks it has found a solution: machine translations that are refined by humans and embedded into every app, interface and business process.

© Techworld
© Techworld

"Unbabel's mission is to help companies communicate with their customers in any language," the company's cofounder and CEO Vasco Pedro tells Techworld in perfect English in London's Soho House.

It does this through a combination of natural language processing (NLP), neural machine translation, quality estimation algorithms and a global network of 55,000 translators.

How Unbabel works

The idea for Unbabel emerged from a conversation with an Airbnb host who was struggling to communicate with potential guests in multiple languages.

Pedro and his Unbabel cofounders realised that adding a translation layer to any interactions with customers would provide a solution.

When they founded Unbabel in 2013, machine translation services typically relied on phrase-based models. These translate individual words or phrases from one language into another and then reorder the words to make the sentence grammatically correct.

They often fail to recognise linguistic ambiguities, and end up making choppy and inaccurate translations.

"A lot of times the right translation was in there but because it's probabilistic in nature, if you didn't have the correct weights, you'd end up picking the wrong path," says Pedro as he sips on an espresso.

"We realised that if we had a human in the loop, the human could help the engine pick the right path in a super-easy way, mostly without even having to write. That was the original insight. That led us to think that the human in the loop component is a huge differentiation. 

"You had either machine translation engines that were all about just the machine, or you had translation services that were all about the humans. What you didn't have was a tightly coupled solution that would really create the benefits of both, where the AI is helping the humans do less work, and the humans are generating data to improve the AI."

Unbabel has since added a variety of features to improve the translation quality and efficiency. They include a neural network that decides in real time whether a machine translation needs human intervention and an AI assistant that augments human translations with automated tips such as a customer's desire for informal terms.

They've also added a smart-routing system that routs tasks to the most appropriate human translator.

"If you give one task to someone that is an expert in that topic, it fundamentally means that they're going to spend way less time doing the research," explains Pedro. "They're going to do a better job faster, so they're going to make more money and we're going to make more money."

How companies are using Unbabel

Unbabel helps companies including Skyscanner, Buzzfeed, and Under Armour conduct two-way multilingual written conversations in near real time, through integrations with marketing automation systems and CRMs such as Salesforce and Microsoft Dynamics.

It also helps publishers and broadcasters bring their social video into multiple territories through Unbabel for Video, a product that uses the platform's API to translate a video in seconds.

Machine translation is particularly effective for small interactions within a fairly narrow domain, such as customer service. Unbabel can do 80 percent of these entirely by machine, and only requires human assistance for the remaining 20 percent.

Other forms of content that people spend hours crafting, such as marketing materials, typically require two humans to generate an effective translation.

"Humans will be for the foreseeable future, involved in this process somehow," says Pedro.

"But what could happen is a shift from humans being responsible for delivering in-production the quality that we need, to taking the role more of training the machines."

What's the future of machine translations?

Neural networks have brought machine translation to new levels of accuracy that may now be starting to plateau due to limits in our understanding of the brain.

"We don't understand yet in humans how language actually works," says Pedro. "Because language is such a direct expression of our intelligence, to learn how to understand language we need to understand intelligence, and we don't really know how that works yet."

Exploring these connections is a full-time occupation for Unbabel's core product team, seven of whom have a PhD in AI or Natural Language Processing.

Pedro did his PhD in computational semantics. This largely focuses on computing the meaning in language, a big current challenge for NLP.

"I started looking at NLP because I felt language was probably the best way to understand intelligence," he says. "Language is the most obvious expression of our intelligence."

Nobody knows when human languages first developed, but much of our thinking today is conducted in natural language sentences.

One popular neurolinguistics theory is that the brain is constantly simultaneously processing a huge amount of disparate information, from sounds and images to feelings and preconceptions. It then blends these separate inputs into a single strand.

"That is typically when you put it into language," says Pedro. "So even when you think, you're thinking in a language.

"It's a way for your brain to create a funnel to serialise all of this parallels that are going on, and only at that moment is when you are conscious of your thought."

These thoughts manifest in an immense variety of languages.

Accents, dialects, slang and our individual characteristics mean that in some sense we all have our own personal languages.

Pedro thinks the translation layer will preserve rather than threaten smaller languages. The universal translator will help everyone speak any language they want, which will help tiny languages survive.

"I think that as machine learning becomes more adaptable, your universal communicator will adapt to your language. And your language evolved by itself, we'll have 7 billion languages instead of the 6,000 we have now. It's a little bit romantic."

Another neurolinguistics theory is that consciousness arises at the point at which we turn things into languages.

Applying this concept to natural language processing could trigger the next major breakthrough in machine translation.

"The neural stuff gives you a more sophisticated representation of language, but you still don't have the understanding of the meaning behind the language, which is something that really enables us to reason and to understand things in a more sophisticated way," Pedro says. "That's where it's headed."

Unbabel won Best Use of AI for Enterprises at the techies awards 2018.