A German AI translates better than Google & Co.

A German company makes the big ones look old. DeepL GmbH from Cologne has launched a free language translator online, which thanks to Artificial Intelligence delivers significantly more accurate translations than Google & Co. The architecture of neural networks makes the difference.
If you are not a firm in a language or at least want to understand the basic content of a text in a completely foreign language, you are welcome to use machine translators on the Internet. On the most famous, the Google Translator for example, or on the Microsoft Translator at Bing. Facebook has also integrated a translation function into its social network if you want to read a shared contribution from other countries in your native language.

However, the comprehensibility of the respective contents after machine translation is not always given. Many factors have a significant influence on the quality of translations. Correct spelling and punctuation of the original as well as the language level and, of course, the databases behind the translation tools. With complicated text structures and technical terms, Google & Co. to their limits. The result is also often curious sentence structures and meaningless word sequences.

More comprehensible translations in many samples

DeepL, the technology company from Cologne, Germany, is now attracting considerable attention with the launch of its translator on deepl. com. Within a very short time computer and technology magazines on the net jumped at it and praised DeepL's Translator in the highest tones. And indeed: even in samples from Springer Professional, DeepL delivers much more precise and understandable results than the "big" competitors, even when translating scientific texts from English into German.

Records in BLEU-Score broken

Such tests, including from other media, confirm the company's statements. DeepL reports that professional translators have rated the translator's results three times better than those of the well-known machine translators. DeepL also breaks records in the BLEU (Bilingual Evaluation Understudy) score, which is used to evaluate machine translators.

The company does not want to publish how exactly DeepL has expanded its artificial neural network. We have made some significant improvements to the architecture of the neural networks,"said Gereon Frahling, founder and CEO of the company. "By rearranging the neurons and their connections, we have enabled our networks to map natural language better than any previous neural translation network." According to the company, the DeepL translator is based on a single model, not an ensemble.

Supercomputer in Iceland with 5,1 Peta-Flops

The AI behind it runs on a supercomputer in Iceland that can perform 5.1 peta-flops (5,100,000,000,000,000,000,000 operations per second). That is enough power to translate a million words in less than one second. According to Jaroslaw Kutylowski, DeepL's CTO, Iceland was chosen as a location due to the oversupply of renewable energies. "This way, we can train our neural networks in Iceland at very low cost. We will continue to invest in high-performance hardware there,"explains Kutylowski.

The DeepL team uses the supercomputer to train neural networks with a huge collection of multilingual texts. The networks look at a large number of translations and learn on their own how to translate grammatically correct and choose good formulations. In doing so, the Cologne-based company uses its first product: Linguee, a search engine for translations. DeepL is a spin-off of Linguee. Over the past ten years, more than a billion high-quality translations have been collected on the platform, which are now being used to train the AI.

API as a business model in development

The DeepL translator currently supports 42 language combinations between German, English, French, Spanish, Italian, Polish and Dutch. According to DeepL, the neural networks have already been trained to master other languages such as Mandarin, Japanese and Russian. The Cologne-based company also intended to provide a programming interface (API) in the coming months. This means that the translation technology can also be used in other products such as digital assistants, dictionaries, language learning applications and professional translation programs.

This is also part of the new DeepL business model. "Our ambitions are not limited to translations,"says Gereon Frahling. "The neural networks have developed an incredible understanding of speech. This opens up many exciting opportunities for the future."

TechCrunch-Article

Summary

I have tested the translator and find him to be much more accurate than other translators of known sizes. For the first time I was able to translate longer or scientific texts with it. As the ultimate test I have written this whole article here in German and had it translated into English by DeepL. I have been using DeepL since August and I am completely satisfied. There are currently only 7 languages to choose from (English, German, Spanish, Italian, French, Dutch and Polish), but as announced all others will follow soon.

I hope you enjoyed this article and that you will use DeepL for your translations in the future.

Best regards to the SteemIt community
booster916

---

list of references:

https://www.deepl.com/press.html
https://www.springerprofessional.de/en/softwaretechnik/datenbanken-und-informationssysteme/eine-deutsche-ki-uebersetzt-besser-als-google-co-/14954442