In our quest for insight into the creation of unicorns in the language services sector, we spoke to Vasco Pedro, the CEO of Unbabel, a technology-centric translation company founded in 2013. We wanted to know which ingredients would be crucial to building out current companies in this space, given the standard inhibitors of fragmentation, product delivery complexity, and hesitations about the role of AI.For Vasco Pedro, the implicit destiny of translation for buyers is for it “to become invisible.” You can also hear this message from large corporates with their own translation departments. For a supplier, though, this is a hard lesson to learn in a world of brand image projection and the attention economy.
“Sometimes it doesn't feel like a translation industry,” says Vasco. “We are focused on delivering very high-level customer service as a use case. What we are experiencing with our clients is much less conversation about translation and much more about augmenting a company’s customer support agents. This changes the whole value proposition.”
For Unbabel, this can mean augmenting a client’s customer support team and handling everything from low-volume languages to very high-level customer support. In particular, because there may be lots of different use cases in the service mix, it no longer makes sense to pay per word.
“What we are focusing on,” says Vasco, “is a single use case because it can unlock real value. So as we are increasingly applying our services to tickets, live chats and phone conversations in customer services, we are using these and not word rates as the unit of billing. When we talk about reaching the 1 bn mark, this represents translating just 25% of all digital interactions. This is a huge number, yet customer service content is itself only a small use case!”
So where will the growth come from? As a technology-centric company – i.e. with a technology-led workflow and tool-base driving production - Unbabel is naturally working on various different components of its model. These range from their NMT solution, based on modified Open Source engines (in-house rather than third-party), their AI-assisted CAT tools, and their automatic translation quality assessment. But Vasco Pedro is also thinking about the problems looming ahead.
“No one,” he says, “has figured out how to deliver seamless website localization, which is still a huge problem because it involves more hard work plugging things, making them work together and then managing the content.
A second problem that will have to be solved involves the speech domain - including speech to speech translation, voice transfer, and using speech synthesis that imitates real individuals: “That would be very original, and enable a voice layer to be added on top of all translation. We need a business model to handle these needs.”
A final prediction sounds like science fiction but “will be relevant in ten years’ time” and that is the advent of brain-to-computer interfaces, operating in symbiosis to increase human processing and output capacity.
Back on the ground, Unbabel has to deal with the current reality of delivering 10-minute translations to clients: “There are large language barriers,” says Vasco, “and the very way you speak creates trust, so we have to be very careful. At the same time, there will be so much content that we are obliged to use an AI-human hybrid approach so that the translation is almost immediate. And suddenly brands are dependent on this solution. But there will always be humans in the loop, as I see nothing that says ‘AI only’. We can’t solve document-level semantics yet and we cannot handle the translation of real-time conversation.”
What kind of impact has this model had on corporate growth, from output to staffing? In the first three years, Unbabel built out its infrastructure. In 2015 it decided to focus on customer service and notched up 300% growth each year. “We are now translating over one million words a day with human intervention, giving us more than 600% growth over the year.”
For staffing, the company has developed a sophisticated QA process so that anyone who is bilingual (not necessarily a professional translator) can apply to join, take the automatic quality tests, and then train up for tasks. Vasco: “We get linguists to annotate 5 to 10% of content to generate scores every week, giving us continuous feedback from editors. This has led to seeing our people as a genuine community.” It also means using AI tools to help the staff master such unfamiliar knowledge as linguistic rules. Yet this approach has produced a very powerful resource and it would cost a lot to shift to another type of staff profile.
As part of the community building, Vasco Pedro says that translator schools are now using Unbabel tools, and the company’s own support plan uses everything to make bilinguals into teachers of translators ranging across 46 countries. The idea is to start with a local hub and then build to critical mass by 2020.
So far Unbabel has pushed in depth into the one, albeit critical use case of customer support. It will be interesting to see if and how they apply their model to other use cases where traditionally translation has been as visible as the company now is!