Is machine translation terribly misunderstood?

Updated: Apr 15

It has been asked time and time again:

Can a machine translate as well (some say better!) than a human brain?
Is translator a doomed profession?
is a Deepl translation good enough?
Have machine translations become better?
Is AI translation the future?

In an attempt to persuade myself that I should really buy into this stubborn trend nagging me to accept MTPE (machine translation post editing) jobs or to give up the profession altogether, and since I had a free afternoon, I decided to give it a try.

Firstly, a bit of background...

What is a machine translation?

According to Wikipedia, the concept of machine translation started in the 9th century in the hands of an arabic transcriptionist and was further studied in the 17th century by the likes of René Descartes. But it's only since the 1950s that we really started to consider using computers to translate time-consuming texts. Fast-forward 70 years, and the internet is full of automated translation platforms such as Deepl or Google Translate, all claiming to use machine-learning translation to replace human translators and provide fast, accurate and cheap translations of pretty much anything.

Translators are worried (not really). Translation companies are all signing up in the hope of keeping up with this ever-accelerating market trend (not the good ones). The language industry is in uproar (at least, the cheap and nasty share) and multinationals (not the serious ones) are rubbing their hands at the idea of saving large budgets on their multilingual documents and online presence.

What if it was all a terrible mistake....?

Machine translation is not a free translation tool. It's a translation aid !

Just like translation memory software was designed to help translators with consistency and productivity, automated translations were originally meant as a translation aid. These API pull out existing translations from their large database to pre-translate documents, ready for translators to go through the chunks of text, edit the ones they deem useful and leave the rest. Without that native speaker's intervention, it's a very rough draft at best.

Using it as produced by the AI would be like asking your lawyers to meticulously review and extensively discuss a contractual text, only to then chose to publish a very early draft on your foreign pages...

Is Deepl an accurate translator? Is Google Translate accurate?

"Deepl itself is reluctant to commit to the accuracy of its translation."

For the purpose of this experiment, I trialled the market-leader (Deepl) but I think it's safe to say that the same conclusions can be drawned with any machine-translation algorythm or API (Google Translate, Lokalize, Wix Multilingual and all their friends).

I signed up for a free account. Let's note in passing that Deepl itself is reluctant to commit to quality, as shown by their disclaimer:

Is machine translation accurate ? — Deepl itself is reluctant to commit to the quality of it's machine-translations

I was thus given access to Deepl's translation API which i could then integrate it to my translation memory software (Trados SDL) so that I could pull upon its database whilst translating documents.

I'll spare you the long-winded fight with plug-ins and add-ons. I eventually got to something that looked ok at first glance. Then I started reading/editing.

Whilst some of the terminology saved me (little) time with some suggestions, there wasn't a single segment/sentence that didn't require extensive edits in order to make sense and meet my quality expectations. More concerningly, I also noticed that:

Similar/indentical sentences yelded completely different suggestions (i though consistency was the one thing machines were good at??)

The API doesn't know how to say "I don't know". So if there's nothing in its "brains" to translate the source text, it makes up rubbish.

Which is the best translation AI?

As a translation aid, it seems that Deepl, RWL and Language Weaver are well positionned to support translators productivity and consistency. If you're looking for an automated translation to translate foreign documents into your own language for internal purposes, Deepl and Google Translate offer simple, free interfaces that will probably do the job.

If you're looking for well written, accurate yet localised copy, you still need a human translator.

Why do machine translations seem to do better job in English than in any other language?

As a bilingual and bicultural french translator and content writer, I've often wondered why the English speaking world was so easily convinced that AI translations were good enough whilst the rest of the world, puzzled, would scratch their heads or laugh out laud when presented with these weird phrases.

It is the New Scientist who provided the best explaination. According to several studies, AI chatbot models "think" in English, which explains why so many english speakers and companies are under the false impression that it does an ok job.

Could you use Deepl to replace a real translator and save money?

Sometimes.

Automated translation is ok for...

Many of my clients have used Deepl or other AI translation software to translate french RFP documents into English, which is ok because they are aware of the potential interference and they're ready to question anything that doesn't make sense. They can then prepare a tender bid in English and ask a real translator to localise it into French.

Internal use
Short, basic correspondence

But a human translator is essential to...

At that point, I'll read the French tender documents, research the client's background, understand the products and finally inject all that technical terminology into a well-written bid, thus reassuring the bid reviewer that their requirements have been truly considered and that the bidder have real expertise. And that's something AI translation certainly can't do...

Customer-facing copy
Technical and regulatory documents

Automated translation is absolutely no good for...

The idea of trusting an AI translation software with a contractual, technical or any other kind of legally binding or customer facing text is ludicrous. And whilst your online content or marketing ressources may have slightly lesser consequences, these illegible, messy or sometimes plain hilarious machine translations still won't reflect very well on your brand. At best, it's a waste of your time, effort, and carefully crafted copy.