TBSJ has seen dramatic change in the past few years with the development and launch of Leveraged AI (LAI), the next-generation translation solution that brings together the very best of man and machine.
As CTO and co-founder, Paul O’Hare has been leading the company’s efforts in technology, so we asked him to reflect on recent developments and consider the future of TBSJ.
How were the early days at TBSJ?
When we started TBSJ, everyone was wearing multiple hats. I was doing project management, translating, and editing. Over time, I moved slowly into IT. We had our own project management system and I began implementing translation technology to help our translators to work faster. I’d always been really interested in that kind of technology and I jumped at the opportunity to be involved. My work developed from selecting and implementing off-the-shelf systems to optimizing them. And for the past few years, I’ve been designing and building our own systems. Since co-founding TBSJ, I’ve moved completely out of project management and translation to working solely on translation technology.
What drives you to architect TBSJ’s translation tech?
I always knew we could make our own systems. Other companies were developing their own and I saw what they were doing right and doing wrong. Translators get paid per word or per character as it’s a more predictable alternative to payment by hour, but too many agencies use tech as a bludgeon, demanding a discount on those rates because of a tool or feature that they force on the translators. If the technology doesn’t actually make them faster, they shouldn’t be obliged to give a discount. That approach is not sustainable if the aim is to work with the best people.
At TBSJ, we do want to work with the best people. The backbone of our company is relationships and we want to use technology in a way that helps translators. Before I was involved in technology at TBSJ, there was an attempt to make a computer-aided translation tool as purely an IT project, but it didn’t work. A lot of money and time was thrown in and nothing came out of it. I wanted to make sure that translator input would help the next effort to be successful.
How and why did you develop LAI?
Some years ago at a translation conference, I spoke to an acquaintance who was an early mover into machine translation. He suggested that the translation industry would separate into two tiers: one for clients who need translation simply for understanding and one for clients who need translation to sell their products or services. When you’re buying translation just to understand something, such as research, you don’t need it to be perfect. Raw machine translation can be useful in this case. But I saw TBSJ in the second tier, with our clients mostly using translation to sell. Despite my natural resistance as a translator, I realized TBSJ couldn’t ignore machine translation, so I put some text into an engine and what came out showed that the technology had promise. But I didn’t want our clients’ data uploaded into any engine; I wanted to make our own, and put a lot of time into the idea.
About two years ago, we started building our own engines. I looked at the first equity research translations coming off production line and thought they were pretty good—and very useable. But it was pretty clear that the machine simply doesn't understand what it is translating, although it is pretty good at papering over the cracks. And even the best engines are unpredictable. So the input of highly experienced translators—or subject matter experts—would be crucial in reviewing the output and getting it right. I started thinking about how to employ this technology with all of our translators and keep them on board.
Then I went back to Ireland for my first machine translation conference, where I met Yury Sharshov who is now our chief scientist. The content of the summit, as well as my discussions with Yury, were a real eye-opener and convinced me that we were on the right track overall. I had enough information to build the first iteration of LAI but I didn’t have the technical background to do everything we wanted. I was dabbling and using a lot of off-the-shelf parts in the production line. What we needed was an expert understanding of how machines perceive human language, as well as skills in deep learning and AI. We were lucky to later hire Yury, who could take us to the next level.
What role do translators play in LAI?
As we’ve learned more about AI, it’s become apparent that machines alone cannot translate to the high levels of quality that our clients require, hence our development of LAI. A critical part of this process is people who can skillfully exert the required leverage; AI can make mistakes in bizarre and subtle ways so we need good translators with a deep understanding of the source material in order to reliably fix those errors. As we introduce our translators one by one to LAI, it’s a two-way flow of information. We’re very careful with onboarding.
What are you working on at the moment?
I don’t want to spoil the surprise, but we are working hard on some very exciting tools that we hope to integrate into LAI and also release as standalone products over the next 12 months. Our IT team has only five (soon to be six) members, so we’ve been developing these particular tools in partnership with Delta-N, an IT solutions provider based in the Netherlands. Internally, we’re revamping our existing infrastructure to incorporate the new components as they come online, and we’re also integrating the latest technologies into our data enrichment processes and engines.
At the end of the day, though, everything is focused on making the complete LAI experience as productive and convenient as possible. If translators and clients don’t like using our tools, we’ve failed.
How important has your background been in developing LAI?
I couldn’t be developing LAI without my decades of translating experience. But I don’t assume that I’m the only one with good ideas. Practically every translator has dabbled to make themselves more efficient, to various levels of success. Good translators are pretty smart people. A lot of them—including our translators—have programming skills or have built pieces of tech to help themselves work more efficiently. We need to canvas them, market research them and see if they like what we’re building. I also need to find time to translate again to get a feel for how our tools will work.
I graduated university as an automotive design engineer, and worked in that role for five years before becoming a technical translator. When I became a better translator I began translating patents and related reports. Over the years, I stepped up a gear each time an opportunity presented itself, and this also included legal and financial translation with TBSJ. All this experience helps me with LAI. There is a lot of linguistic knowledge involved in tweaking the engines and making them work better. An engine is like a black box—it’s almost impossible to see or understand how it actually produces specific translations. You have to carefully choose the data to build it on and then refine that data to influence its behavior. It’s like developing our secret sauce—the little things we do each time to produce a better quality machine translation, thereby saving the translators more time. It’s an accumulation of subtleties that delivers a meaningful result.
How has TBSJ evolved in the past decade?
There have been so many milestones, particularly in the past year with the development of LAI. We’ve designed and built something that our translators can use to work more efficiently. Working out how to do that and then seeing translations using our technology come off the production line was a huge step, as was deploying LAI as a service for our clients.
Over the years our company DNA and vision has become stronger. Our ideals when we set up TBSJ were to treat translators well and to be a company that people wanted to work for. Everyone buys into that when they join TBSJ so all our team are fully committed to it, perhaps even more so today than at the company’s outset.