In a previous story, I half mockingly claimed not be impressed by the applications of AI and machine learning in everyday life. And then I even full mockingly made fun of the idea by claiming (in a post where I was attributing AI capabilities to a thermos) it might be time to change my view. But this time, it might really be time to revise my opinion.

About everyone will have used Google Translate (or the equivalent from Microsoft or Facebook) at some point by now. And about everyone will have concluded that is was kinda OK, especially if all you needed was to more or less understand what the text meant. But the text structure is almost always somewhere on the range from awkward to downright wrong. And translations of idioms remains troublesome. These problematic characteristics are linked  to the way computer translation has evolved. Originally, those translation services tried to analyze the syntax of the text in the source language (i.e. ‘understand’ the meaning) and then come up with the equivalent for each element in the target language. In the text parts where that doesn’t work, you end up with word by word translations, or even some words that remain untranslated. All in all a perfect illustration of the ideas uttered in my I’m-not-impressed post.

But recently I got to know DeepL Translator, a new translation service from a small German company. And I have to admit: I am really highly impressed by the quality and the natural feel of the translations that come out of it. The difference in quality results from a difference in approach. DeepL does not attempt to understand the syntax of the text, but completely depends on another method of machine learning (used by the other translation services too, but apparently in a much less efficient way): loads of translated texts have been fed into the system and the algorithms simply dig up the most appropriate equivalent in the target language — but with an amazingly accurate sense of context and overall meaning. And because the original input are actual translations predominantly made by professional translators, the output definitely feels more natural and ‘human’ than the output of Google Translate.

Obviously the resulting translations are not perfect (and in my view it will still take quite a while before they will come close to reaching the point where no human has to check the output), so by far not all translators should fear losing their job — perhaps only those that translate like a robot.