Artificial intelligence-based translation software DeepL is making its debut in the Forbes Cloud 100 list this year, thanks to a machine learning translation model that users say is more precise than Google’s.
With Rashi SrivastavaForbes Team
Jaroslaw Kutylowski speaks German, Polish and English (and can order a French Coke). With DeepL, the artificial intelligence supported translation tool of the company he founded, he can read and write in about 30 more languages.
Founded in 2017, DeepL has developed translation software that it says is far more accurate than competing products offered by Google and others, thanks to some powerful artificial intelligence working in concert with native speakers.
DeepL’s proprietary neural network architecture – a machine learning technique that helps a computer learn information the same way a human brain does – is trained on a large database of publicly available bilingual (translated) and monolingual (untranslated) text in 31 different languages, including Chinese. , Russian, Spanish and Italian.
That’s a tiny fraction of the 130 languages you can translate using Google. But DeepL’s translations are tuned for nuance by human editors and native speakers. The startup employs 20 in-house editors and more than a thousand contract human translators and native English speakers worldwide to evaluate the quality of translations produced by DeepL’s model and adjust for accuracy.
“We need a lot of high-quality human-translated data to learn about the quirks of a language to translate both everyday and formal text,” said Kutylowski.
DeepL has more than 10 million monthly active users, of which 500,000 pay between $9 and $59 per month. This includes grandparents who use DeepL to talk to their grandchildren in their own language, and romantic partners who struggle with a language barrier. But most of DeepL’s business comes from its 20,000 corporate clients – Mercedes Benz, Fujitsu and German rail company Deutsche Bahn – who use DeepL’s software to translate everything from websites, legal contracts and customer contracts to emails, marketing texts and PowerPoint slides. are a few. .
According to Pitchbook, in January 2023, the startup, headquartered in Cologne, Germany, has raised approximately $100 million in funding, valued at $1 billion, from global VC companies including Institutional Venture Partners (IVP), Atomico and Bessemer Venture Partners. CEO and founder Kutylowski has not confirmed the total amount of funds his company has raised to date.
DeepL has been downloaded on 25 million devices; that’s very little compared to Google Translate’s over 1 billion installs. But those who use it praise its accuracy. Nina Gafni, a Washington DC-based professional translator who previously worked as a linguist and translator at the Federal Bureau of Investigation, uses DeepL to translate French, German, and Italian into English. While it’s never exactly perfect at translating machine learning systems, it’s more culturally nuanced and precise than most DeepLs, she says.
“Sometimes machine translations can be very realistic, and that’s a big problem,” Gafni said. “If I’m in a difficult situation, I feel like I can trust DeepL much more than Google Translate.”
This may be because of the human calibration introduced to impress the pile of movie subtitles, book and patent translations, and forum conversations used to train DeepL.
Japanese native Akiko Taguchi, who works for DeepL, said: forbes spends most of his time making sure DeepL’s translations are contextually correct and humane. “I gave feedback as the machine mixed up the formalities used in my Japanese writing,” Taguchi said. forbes. “It has improved significantly since then.”
Originating from a search engine translation tool Linguee, DeepL allows users to translate all documents such as PDF, Word, and PowerPoint while keeping the formatting intact. It suggests alternative translations and tonal changes, and allows users to create a custom dictionary to specify how certain words should be translated.
In computer science, Kutylowski was developing the machine learning technology underlying DeepL in 2016, while at the same time Google researchers were working on Transformers, the translation-based technological breakthrough powering ChatGPT. But as of now, DeepL doesn’t use Transformers or major language models for translation, he said, and declined to specify DeepL’s exact architecture. “We knew that neural networks would dominate this field very soon,” Kutylowski said. “We were convinced that nothing previously used for translation, such as statistical methods, would have any application in the future.”
That turns out to be true, according to Karthik Ramakrishnan, co-founder of IVP, who led DeepL’s $100 million Series B investment round in January. There is a $27.9 billion global market for translation services and this market has not been well evaluated. “The vast majority of translation is still very manual, an outdated old workflow,” Ramakrishnan said. “People hire outsourced language service providers that take days or even weeks to deliver your translated content on time.”
With its applause and $1 billion valuation, DeepL is in a reasonable position to capitalize on the market that Ramakrishnan is watching. But so are the others. While the translation tool isn’t as accurate as DeepL, Google supports 130 languages and is backed by a giant that has invested decades in AI. Another major tech competitor is Microsoft Translator, which supports over 100 languages and dialects and has large enterprise users like the Volkswagen Group. Meanwhile, ChatGPT, possibly the bot that popularized artificial intelligence, can currently translate into 50 languages (although there are numerous translation issues, from hallucination to mistranslation of everyday language). This will only improve thanks to the continued development of mainstream OpenAI’s industry-leading major language models.
DeepL does not have its own LLM yet. Kutylowski said forbes plans to create one in-house and combine it with the existing smaller translation model for better results. To that end, DeepL uses Nvidia’s AI-driven data center in Sweden to build computational resources to train AI models. “Generative AI is a huge opportunity,” he said. “It will allow us to build functions on top of the translation where the translator becomes more interactive or more in dialogue with you.”
MORE THAN 100 FORBES THE CLOUD
#German #Unicorn #Google #Translate #ChatGPT