Translating company websites and apps is tedious, in some ways more so than a straight-forward white paper or article. With language spread throughout a platform, developers and translators alike have been irritated by the process of picking out text and then putting it back into beta versions of new sites in a new language. That frustration got under the skin of one pair of Austrian entrepreneurs as they launched their own translation business a few years ago. Now, LingoHub is pushing a technology that offers continuous translation, i.e. automatic language updates for beta sites using automatic machine translation (MT).
“We wanted to create a solution focused on software. So, how I came up with this was I [had] worked at several startups before. The translation process was always a hassle. If I can create the software so that all the tests run correctly, so why not create the same concept for human language?” LingoHub Co-founder and CEO Helmut Juskewycz told Geektime. He wanted to streamline the process of localization, the term developers use for adapting sites to other languages. “I was facing the challenges of localization everyday: localization was time-consuming, complicated and costly. I knew that the entire process can and must be easier.”
He and Senior Developer Markus Merzinger pivoted the company in 2013, which had originally launched in 2010 as a conventional contractor sort of company with a focus on translation. Juskewcyz says the goal of automating translation services, or perhaps better-put streamlining them, was there long before the shift. Using what they dub Translation Memory Autofill, they can automatically populate beta sites for alternative language versions of clients’ current sites and apps via integrations across GitHub, Bitbucket or REST API.
“It popped up in my head while working as a developer. I was facing the challenges of localization everyday: localization was time consuming, complicated and costly. I knew that the entire process can and must be easier.”
Juskewycz says with his computer science background, it took some time but he’s got a full handle now on the translation market. They’ve bootstrapped their whole operation up until this point and are close to breaking profitability. They would take files, send them to translators, then send the finished text to developers who would have to input it. But that process, Juskewycz thought, should have been more seamless.
“I always wanted to free up the developers. They should program, they shouldn’t translate. The best answer you’d get on translation is ‘who cares? I care about the technology, not the translation. It should work, but I have my own focus.'”
The system that eventually evolved into their product uses machine translation to provide a rough translation of a given text before it is passed onto a professional human translator for review. This allows them to use technology by saving time on much of the grunt work, while the humans check for accuracy and smooth out the edges.
As they built up, they recruited some 3,500 vetted translators who are fluent in 35 languages. LingoHub now offers more than 200 language pairs for translation (English-German, German-Spanish, etc.). The bulk of their customers are in mobile right now, so while they have the ability to auto-fill test sites for desktops, their business is extremely focused on apps at the moment. While they can claim clients around the world, the bulk is in the German-speaking area in Austria, Switzerland and of course Germany.
Juskewycz explains to Geektime the entire process is actually cheaper for the customer because it cuts down on human errors but still uses human translators for the most essential parts of the process with their personal expertise. The company also allows access to their software if clients have their own in-house translators they would want working on the project.
“For me the role of every Machine Translation tool is to provide translators with an initial translation phrase. When it comes to high-quality translation there’s only one option go with: professional human translators. Machine Translation is not able to understand cultural intricacies and often struggles with grammar and syntax. Just a certified translator can translate your text in a way it reflects your local target audience.”
By the sound of it, from this writer’s perspective, that option will likely always exist, or continue to exist for some time. Companies can more easily rely on in-house people familiar with certain terminology and concepts to get something more quickly than even the most sophisticated translation memory (TM), a sort of machine learning program that can record how certain words are translated over time relying on a number of samples. However, if the technology gets good enough, could their translators be in fact working themselves out of a job in the future?
The rise of machine translation startups
A very, very similar company that Geektime has profiled is Portuguese-based Unbabel, which also boasts a global network of translators and allows both low-level and more experienced translators to work off of a translation produced by a machine learning program that will pick up on edits over time. In LingoHub’s case, they are more focused on integrating their software with their customers’ digital media and translating app features into new languages to allow those customers to more quickly release their products across borders. Each client has its own translation memory, so an engineering company can expect machine interpretations that fall back to engineering-related translations.
“Unbabel provides translation with a lot of post-editing of machine translations, providing translators with a combination of MT and suggestions. With us, once we have established [a] connection to the beta [site], we check everything happening in the beta with an automatic process that we call ‘continuous translation.’ It’s like a living thing that always changes.”
The MT industry is growing fast, expected to grow at 23.53% annually through 2019 according to Research and Markets, perhaps to as much as $983.3 million by 2022 according to Grand View Research (expect numbers will vary widely on this industry, as very few companies are in a position to provide translations that would need very little human correction and it is still unclear whose machine learning algorithms will be the most useful).
“What’s unique is obviously is the process integration of the USP. We do a lot of stuff that is really important for the developers so our platform is specialized for software developers.”
Google Translate and the future of translation
That’s the big key for the vertical’s future Juskewycz suggests, as you have to allow developers to focus on their own language — i.e., coding — and not worry about the human translation work that should be left to the people accessing the system.
“The trend is the hardware system, between machine translation and translation memory.” That allows companies like LingoHub to constantly refer back to the linguistic style and vocabulary needs of individual customers, relevant to their particular industries. “Translation memory is something you store from what’s already translated. What you always have is some sort of text style in the TM really focused on the company. If you want to support the translator you need to have some sort of TM. A hardware system that supports a human.”
The translators are paid well and the rate varies depending on the language pair. The usual rate is 0.15€ to 0.19€ per word, with an extra 0.05€ if the customer wants a second translator to review the work afterwards.
“It depends a bit on the language and the language pair, but we do we pay the translators well because we think it’s real work. It’s tough work. We have 25% commission but the rest goes to the translators and it is always by the word. In some rare cases it is per hour.”
When asked what he thought about Google Translate’s free service compared to his own, he said that he will never be unimpressed with what Google has achieved, but for the time being it has its limits.
“In my opinion [Google Translate] is one of the best, if not the best, MT services out there, [but] MT isn’t going to replace human translation any time soon.
The primary reason he says is the issue of context, which machine translation simply hasn’t been able to get just yet. Issues like nuance and tone are still not fully integrated into natural language processing software, much less software that can reconfigure processed language and then convert it into another tongue. That probably goes double for written language, which loses auditory and visual cues that are involved in regular face-to-face convos. One example Juskewycz gives is marketing language, which is meant to convey sophistication on the one hand and capability on the other. Usually written with a positive spin, it would take time to line up the proper equivalents of marketing language in two languages.
“There is always some kind of text style in every language you have to consider. But marketing is so much more, it’s about how you position itself. If you want to translate for example, a tweet, it’s not really sophisticated text.”