[INTERVIEW] This German start-up is taking translation technology to the next level

While tech giants are betting big in an intensifying race for global artificial intelligence (AI) dominance, a German start-up is taking on the Goliaths in the highly-competitive machine translation market.

Jaroslaw ″Jarek″ Kutylowski, DeepL CEO and founder, speaks during an interview at Josun Palace Hotel in southern Seoul, on May 9. [DEEPL]

While tech giants are betting big in an intensifying race for global artificial intelligence (AI) dominance, a German start-up is taking on the Goliaths in the highly-competitive machine translation market.

DeepL, based in Cologne, Germany, is a machine translation company founded in 2017.

DeepL Translator is currently available in 31 languages including Korean, which the company first started offering in January. The service is known for its strong performance in understanding the context of the inputted language and therefore comparatively more natural-sounding results in the target language of translation.

DeepL Translator is powered by the artificial neural network, which can teach and improve itself in any language — possibly even those from outside the planet, if there are any — as long as enough training dataset of sufficient quality is provided.

“The nice thing is that neural networks are actually capable of learning languages on a certain abstraction level,” says Jaroslaw “Jarek” Kutylowski, DeepL CEO and founder.

“You don't have to teach the neural networks explicitly what the grammar of a language is; it's going to be able to learn that on its own, based on the examples you've shown it.”

“If there is an extraterrestrial language that we would like to train AI on, as long as we have enough examples of translation from English to the language that's coming from the alien civilization, that would actually work,” Kutylowski added.

DeepL describes its service to be “the world’s most accurate and nuanced machine translation,” citing improved network architecture and the high-quality training dataset assessed by special web crawlers.

In blind tests, where professional translators evaluated the quality of translations, DeepL’s translation of European languages to English, and vice versa, was rated about three times higher than the scores of its bigger competitors, according to the company.

While DeepL’s primary focus was on the European markets, it is now striving to further expand its foothold in Asian countries.

DeepL began its Korean service earlier this year, and plans to release a paid service, DeepL Pro, starting in August.

Kutylowski believes Korea is one of the largest markets for the company in the future, with demand growing in the country at a steep pace.

The Korea JoongAng Daily sat down with Kutylowski for an interview on May 9, the first day of his two-day visit to Korea, at Josun Palace Hotel in southern Seoul.

The following are edited excerpts from the interview.

Q. DeepL Translator is said to be “three times more accurate” compared to its competitors. How did your company achieve that?

A. We started our journey when neural networks appeared as a way to translate, which was around 2017. We adopted that technology very quickly and went deep into research on how we could improve it.

As with a lot of achievements of humans, it is step-by-step work.

Fluency is something pretty interesting for translations — you can either translate very literally, word by word, which makes the translation exact but may not really work in the target language.

On the other hand, you can translate freely. The danger here is that you may not be hitting the message that was originally there.

So we've been putting in quite a lot of work to make sure that we’re right there where we need to be, as a translator both producing a nuanced and fluent translation, while also making sure that the real message is being conveyed.

I think that our users and customers like DeepL because of its ease of use. It's not only about the quality of the translation itself but also about how the whole product works.

What are the challenges specific to the Asian languages and market? And do cultural differences or linguistic nuances in Asian languages affect translation accuracy? If so, how does DeepL address such challenges? In general, Asia, from a Western perspective, has a lot of languages that are really different. It is not only about Asia — we are also working on Arabic, which also has a totally different grammar and structure. So we had to rethink how a language is encoded into a form that is digestible by computers.

One of the beauties of AI and neural networks is that they can actually learn a lot of that from the examples that we are providing. So it's more of a question of what kind of data are we providing to the AI.

It needs to see examples of translations that contain cultural differences. It can then abstract those, and in some way understand and translate things like idioms that are not really encoded in books normally, but come with a deeper understanding of the knowledge.

The materials that you find on the internet, whether it's a translation or not, they have varying levels of quality. We’re pretty cautious about which kind of data we're giving the neural networks to train, and how we are ranking, preparing and filtering that data. This is a large part of the process we are working on.

Since DeepL began its Korean service in January, have you seen meaningful growth in the number of users or the market share?

Yes, the number of users is growing very strongly. We’ve been comparing the figure to those of other languages when they were initially launched over the last years, and I would say Korean definitely shows one of the strongest growths we’ve ever seen.

This is why we also decided to move faster in introducing DeepL Pro. We usually follow this path of first introducing the language, then moving on to DeepL Pro, and maybe at some point in time also creating a team within the country. We have accelerated this path here because the number of users is growing so quickly.

DeepL completed a successful funding round earlier this year, despite the ongoing market uncertainties worldwide. What do you think is the main driver for the company in attracting investments?

The company is growing very fast, and we've been successful in growing steadily with around twice the figure every year in all of the relevant metrics of the company, such as revenue, the number of employees and more.

At the same time, we have always been profitable. That makes us different from some other start-ups, which can only rely on external funding, and therefore the risk is high. We’ve always been able to maintain profitability, which is why we are very stable financially.

What’s your ultimate goal?

We see ourselves as an AI communications company, so we want to help people communicate in any way that technologies enable us to do so.

Translation of written texts is one important part of it, but there are also other problems we see in communications. Anything we can do to overcome not only the language barriers but also the general communication barriers that are encoded in our vision — that is what we are working on, and will continue to work on.

BY SHIN HA-NEE [shin.hanee@joongang.co.kr]