Feature Internet

DeepL uses smart machine learning to teach other online translators.

  1. Armstrong Roberts/Getty Images/H. Armstrong Roberts/Getty Images/H. Armstrong Roberts/Getty Images/H. Armstrong

Google, Microsoft, and Facebook are all using machine learning to improve translation, but a small company called DeepL has surpassed them all and set the bar even higher. Its translation tool is not only faster than the competition, but it’s also more precise and subtle than any we’ve tried.

I barely know a smattering of French in addition to my rudimentary English, but fortunately, my colleague Frederic is a multilingual man. DeepL’s translations were generally superior to those of Google Translate and Bing, we both agreed.

Consider the following excerpt from a German news storey, as rendered by DeepL (top) and Google:

“Whereas Google Translate frequently goes for a highly literal translation that overlooks various nuances and idioms (or gets the translation of these idioms dead wrong), DeepL frequently delivers a more natural translation that comes closer to that of a trained translator,” Frederic explains.

The second sentence is more logical; the measure is “designed to” rather than “doing” something; the police are “on the road in armoured vehicles” rather than “on the road”; “martial appearance” may be imperfect (though inspired), but it’s better than the nonsensical “fighters’ turmoil…had come to the fore.”


DeepL consistently came out on top in a few tests I ran on some French literature I knew well enough to judge. A more legible translation has fewer tense, purpose, and agreement mistakes, as well as a better knowledge and application of idiom. We agreed, as did translators who participated in DeepL’s blind testing. But do not takes our word for it; try it out for yourself.

While it’s true that meaning can be delivered successfully despite such flaws, as proven by the utility we’ve all found in even the worst machine translations, it’s far from certain that anything more than the most basic facts will make it through.

Linguee has progressed.

DeepL arose from the similarly great Linguee, a translation tool that has been around for years but, despite its popularity, has never fully matched Google Translate — the latter, after all, has a big edge in terms of brand and position. Gereon Frahling, one of Linguee’s co-founders, used to work at Google Research before leaving in 2007 to pursue this new endeavour.

The team has been using machine learning for tasks other than core translation for years, but it wasn’t until last year that they started seriously working on a new system and company, both of which would be known as DeepL.

“We have constructed a neural translation network that encompasses most of the newest breakthroughs, to which we added our own ideas,” Frahling wrote in an email, indicating that the time was right.

An vast database of over a billion translations and queries, as well as a way of ground-truthing translations by scanning the web for similar snippets, provided a solid foundation for the new model’s training. They also constructed the world’s 23rd most powerful supercomputer, which is conveniently located in Iceland.


Convolutional neural networks, rather than the recurrent neural networks that Linguee had been utilising, were the way to go, according to research released by universities, research organisations, and even Linguee’s competitors. Because this isn’t the time to go over the differences between CNNs and RNNs, suffice it to say that the former is a better option for correct translation of long, complicated strings of connected words as long as you can account for its flaws.

A CNN, for example, might be thought of as tackling one word in a sentence at a time. This becomes a difficulty when, for example, a word at the conclusion of a sentence defines how a word at the beginning of the sentence should be produced, as is frequently the case. It’s an waste of time to go through the entire sentence only to discover that the first word the network chose is incorrect, so DeepL and others in the machine learning field use “attention mechanisms” that watch for potential stumbling blocks and correct them before the CNN moves on to the next word or phrase.

Of course, there are more hidden strategies at work, and the end result DeepL is a translation tool that I want to use as my new default. I’m looking forward to seeing how the others respond.

Related Articles

Wide Area Network Pros & Cons

Stefan Stefan

Leave a Comment