DeepL schools other online translators with clever machine learning
Tech giants Google, Microsoft and Facebook are all applying the lessons of machine learning to translation, but a small company called DeepL has outdone them all and raised the bar for the field. Its translation tool is just as quick as the outsized competition, but more accurate and nuanced than any we’ve tried.
I only speak a smattering of French in addition to my passable English, but luckily my colleague Frederic is a man of many tongues. We both agreed that DeepL’s translations were generally superior to those from Google Translate and Bing.
Take, for example, the following passage from a German news article, as rendered by DeepL (top) and Google:
As Frederic puts it: “Whereas Google Translate often goes for a very literal translation that misses some nuances and idioms (or gets the translation of these idioms dead wrong), DeepL often provides a more natural translation that comes closer to that of a trained translator.”
The second sentence is parsed more naturally; the measure is “designed to” accomplish something rather than just doing that thing; the police are “on the road in armoured vehicles” as opposed to merely on them; “martial appearance” may be imperfect (though inspired) but it’s far better than the nonsensical “fighters’ turmoil…had come to the fore.”
A few tests of my own on some French literature I know well enough to judge had DeepL coming out on top regularly, as well. Fewer errors of tense, intent and agreement, plus a better understanding and deployment of idiom make for a much more readable translation. We thought so, and so did translators in DeepL’s own blind testing. But don’t take anyone else’s word for it — test it out yourself.
While it’s true that meaning can be conveyed successfully despite errors of that class, as evidenced by the utility we’ve all found in even the poorest machine translations, it’s far from guaranteed that anything but the barest facts of will make it through.
DeepL was born from the similarly excellent Linguee, a translation tool that has existed for years and, while popular, never quite reached the level of Google Translate — the latter has a huge advantage in brand and position, after all. Linguee’s co-founder, Gereon Frahling, used to work for Google Research but left in 2007 to pursue this new venture.
The team has been working with machine learning for years, for tasks adjacent to the core translation, but it was only last year that they began working in earnest on a whole new system and company, both of which would bear the name DeepL.
In an email, Frahling told me that the time was ripe: “We have built a neural translation network that incorporates most of the latest developments, to which we added our own ideas.”
An enormous database of over a billion translations and queries, plus a method of ground-truthing translations by searching for similar snippets on the web, made for a strong base in the training of the new model. They also put together what they claim is the 23rd most powerful supercomputer in the world, conveniently located in Iceland.
Developments published by universities, research agencies and indeed Linguee’s competitors showed that convolutional neural networks were the way to go, rather than the recurrent neural networks the company had been using previously. Now isn’t really the place to go into the differences between CNNs and RNNs, so it must suffice to say that for accurate translation of long, complex strings of related words, the former is a better bet as long as you can control for its weaknesses.
For example, a CNN could roughly be able to be said to tackle one word of the sentence at a time. This becomes a problem when, for instance, as commonly happens, a word at the end of the sentence determines how a word at the beginning of the sentence should be formed. It’s wasteful to go through the whole sentence only to find that the first word the network picked is wrong, and then start over with that knowledge, so DeepL and others in the machine learning field apply “attention mechanisms” that monitor for such potential trip-ups and resolve them before the CNN moves on to the next word or phrase.