DeepL CEO envisions an interactive AI assistant that can help with your language

Jarek Kutylowski, founder and CEO of DeepL, speaks during an interview with the Korea JoongAng Daily in Seoul. [DEEPL]

[INTERVIEW]

DeepL eyes an AI-powered communication tool that is not only precise in its own right, but is able to ask questions to the user for higher quality output.

The German startup, best-known for its translation service, recently launched a sentence-editing service in Korea powered by its own large language model (LLM), currently available in English and German.

"[What I ultimately aim for is] a personal language assistant," said Jarek Kutylowski, founder and CEO of DeepL in a recent interview with the Korea JoongAng Daily.

"It will serve as both a translator and editor that is super proficient with language. It will ask questions when things aren't clear, like an editor [at a news publication] who corrects us and gives feedback so that we can learn."

DeepL's translation service stands out among similar services provided by the likes of goliaths such as Google and Microsoft by scoring higher points in multiple tests by various agencies in the translation or publication sectors.

In a test conducted by DeepL with a number of external professional translators, the company's tech was preferred over Google's 3.9 times for English to Korean translations, the company said.

The startup, founded in 2017, launched the translation service in Korea last year.

A year into doing business here, Kutylowski sees more potential in the country.

"Korean language is very complicated, which is very far away from Western languages. So the gap between English and Korean is obviously large and therefore there's more potential for us to close that gap," he said.

"What we are seeing specifically to Korea, also, is this really strong outward-facing economy that relies heavily on external markets. It also has strong research and development, particularly in electronics and therefore has a strong need on the academic side to be very connected to the world."

Koreans' tendency to speak English at a high level compared to others doesn't come as a risk to the CEO.

"Regardless of how people already speak certain language very well, there is just a need for quicker way to write something," he said.

"And for business purposes you need particular set of skills and vocabularies."

DeepL's dive into sentence editing derived from the pattern of how its customers used its existing translation service.

"We saw a lot of people writing in English, translate for example into Korea and then translating back into English because it helped them improve the text," he said.

"We thought, 'O.K. then maybe we can do a better tool for that.'"

Another reason for expanding into editing was due to its enterprise clients, DeepL's primary consumer base.

"Enterprise clients said they have a lot of communication problems within the company which doesn't have to do with foreign languages and asked us to help with our technology."

In order to offer an editing tool that can change the style and tone according to user preference, DeepL developed its own LLM using Nvidia's H100 processors.

Right now, the LLM model doesn't interfere in the translation service, which relies on a smaller algorithm, according to the company.

Ultimately, however, the CEO believes its array of services will pick and choose which algorithm or language model it will utilize.

"When the AI for translation notices that it lacks some data to create a good translation that something is ambiguous, it will ask a user for example, 'What do you mean by that?'," he said.

"It can have this kind of interactive dialogue with the user and with the user input, it will be improving the translation even further."

DeepL plans to expand the translation service to spoken language in the near future as well. Whether it will be with an external partner or not cannot be confirmed as of now, according to Kutylowski.

BY JIN EUN-SOO [jin.eunsoo@joongang.co.kr]