DeepL appears to be significantly better than Google Translate

[emk]

This has been getting some press lately, and some people are claiming it frequently outperforms Google Translate: DeepL.

I actually have a sample text that I use to test translation software. WARNING: Cruelty to seagulls ahead. This text from Topito grosses me out a bit, but I've found over the years that it tends to break translation software. This text is interesting for a couple of reasons:

• It deliberately mixes formal and informal language.
• It uses plenty of vocabulary you'd never find in a newspaper (and most translation software is trained on newspapers).
• It uses some nautical vocabulary, which is also outside of many training sets.

Basically, it's a text that would normally pose a lot of headaches for typical B2 students, though solid C1 students should be mostly OK if they've read a lot. Here's the original text in French:

Quand on croise un marin aviné au comptoir d'un bar côtier, celui-ci ne manque jamais l'occasion de narrer le dernier coup de pute qu'il a fait avec ses compères à une mouette chopée au large. Il s'agit en général de lui faire bouffer des roulements à billes et de regarder la pauvre bête clouée au sol sous les rires gras de ces rugueux marins. Ou de littéralement l’agrafer à une planche pour déconner. C'est certes cruel, mais ce sont bien ces gaillards au visage buriné par les embruns qui ont raison tant la mouette est un animal détestable. Les bonnes raisons de le penser ne manquent pas.

Here's the DeepL translation:

When you come across an avid sailor at a coastal bar counter, he never misses the opportunity to recount the last hooker's blow he made with his fellow sailors to a seagull caught in the open sea. It's usually a matter of making her eat ball bearings and looking at the poor beast nailed to the ground under the greasy laughter of these rough sailors. Or literally staple it to a board to screw around. It's cruel, but it's those guys with their faces chiselled by the spray that are right because the seagull is such a detestable animal. There is no shortage of good reason to think so.

There's a couple of weird translations here: aviné is a fairly uncommon French word, but the dictionary says it means "drunk". A couple of colorful words are translated as more generic ones, and so on. But basically this text is usable. I've never seen a machine translation handle this text well before.

I'm also fascinated by the translation of compères "buddies, partners in crime" as "fellow sailors". That's a clever bit of context dependence. And I'm also impressed by the "making her eat ball bearings", because lui in this particular context is gender-neutral, but the underlying referent is mouette "seagull", a feminine word. So it seems to be tracking long-range gender agreement.

Now let's compare what Google Translate gave me:

When you cross an avined sailor at the bar of a coastal bar, this one never misses the opportunity to narrate the last blow of a whore he made with his friends to a seagull choped off. It is generally a matter of eating ball bearings and watching the poor animal nailed to the ground under the fat laughter of these rough sailors. Or literally staple it to a plank to mess up. It is certainly cruel, but it is these fellows with a face engraved by the spray that are right so much the seagull is a detestable animal. There is no shortage of reasons for this.

Yeah, Google Translate is still a total disaster here. I have no idea how it gets "chopped off" from chopée au large, but DeepL chooses the much better "caught in the open sea".

So, my verdict: Go check out DeepL if you use machine translation to make sense of L2 texts. This one looks pretty nice after my single, unscientific test. And it seems to have a much better informal vocabulary than Google Translate. Let us know if it works well for you!

[tommus]

So, my verdict: Go check out DeepL

I just checked it out for Dutch to English, and I am very impressed. You are right. It is excellent. And if you find translation words that don't look quite right, you can click on them to get a range of alternatives. Thank you for this heads-up!

[aokoye]

It's from the same company that created and maintains Linguee which I am a big fan of. Given the corpora they appear use and web crawling they seem to do to generate those corpora I'm not surprised that DeepL appears to be as good as it is. Thanks for the find Emk!

[Serpent]

I wanted to check Finnish but of course it's not available.
Honestly English is a relatively easy language to translate into. When people are sceptical about automatic translation, it's mostly about translating into languages with a rich morphology and complicated word order.

I've tried this quote from LOTR:
Deserves it! I daresay he does. Many that live deserve death. And some that die deserve life. Can you give it to them? Then do not be too eager to deal out death in judgement. For even the very wise cannot see all ends.

And DeepL fails with it in Polish and Italian. Spanish and German seem better but none of the translations for "deal out death in judgement" made sense. In Italian it even turns "Deserves it!" into just "Deserves!" (without translating the word). A lot of the Polish output simply isn't grammatically correct.

This is Hemingway, Spanish to English.
His decision had been to stay in deep and dark waters, far from all traps and baits and betrayals. My decision was to go there to look for him, beyond all people. Beyond all the people in the world. Now we're alone for each other and it's been like this since noon. And no one to come to us, not him or me.

[Xenops]

When they have Japanese, that would be the real test for me.

[Stefan]

In their press release from today:

DeepL Translator currently supports 42 language combinations between English, German, French, Spanish, Italian, Polish, and Dutch. The neural networks are already training to master more languages like Mandarin, Japanese, and Russian. DeepL also intends to release an API in the coming months, allowing its superior translations to enhance other products such as digital assistants, dictionaries, language learning apps, and professional translation programs.

[emk]

I've tried this quote from LOTR:
Deserves it! I daresay he does. Many that live deserve death. And some that die deserve life. Can you give it to them? Then do not be too eager to deal out death in judgement. For even the very wise cannot see all ends.

And DeepL fails with it in Polish and Italian. Spanish and German seem better but none of the translations for "deal out death in judgement" made sense. In Italian it even turns "Deserves it!" into just "Deserves!" (without translating the word). A lot of the Polish output simply isn't grammatically correct.

Ah, nice test case! Going into French, we get:

Il le mérite! J'ose dire que oui. Beaucoup de ceux qui vivent méritent la mort. Et certains qui meurent méritent la vie. Pouvez-vous le leur donner? Alors, n'ayez pas trop hâte de juger la mort. Car même les plus sages ne peuvent pas voir toutes les fins.

This is actually a pretty usable translation for the passage. However, "deal out death in judgement" becomes juger la mort, which I'd have to check with a native speaker (or a large French corpus). And I can't judge how idiomatic or poetic some of those turns of phrase might be. On the other hand, I use machine translation only for L2 → L1 (or L3 → L2), but never for L1 → L2. So mostly I don't care about an odd word choice as long as the meaning is preserved.

Google Translate is once again a trainwreck:

Le mérite! Je dirais qu'il le fait. Beaucoup de ceux qui vivent méritent la mort. Et certains qui meurent méritent la vie. Peux-tu leur donner? Ensuite, ne soyez pas trop désireux de régler la mort dans le jugement. Car même les très sages ne peuvent pas voir toutes les extrémités.

DeepL is the clear winner here. In particular, Google's "voir toutes les extrémités", "Le mérite!" and "Peux-tu leur donner" are much weaker than the DeepL equivalents.

So at least for English/French pairs (which were always one of the strongest pairings in the Linguee database), it's remarkably good so far. Obviously, if you need Japanese, you're out of luck for the moment. But nonetheless, I'm excited: DeepL is more than good enough to automatically generate L2 → L1 translations for Anki cards.

[Elenia]

I wanted to verify the meaning of a Clozemaster sentence so I checked first google translate and then, remembering this thread, I checked DeepL.

Ich möchte ihn nicht.

I don't want it.

I do not want him

I don't want to kill him
I don't want to see him

Huh. Interestingly enough, the DeepL translation changes when I add the full stop to the end of the original translation, offering pretty much the same translation as Google. So clearly the translator anticipates the most likely ending to the sentence based on it's corpus and acts accordingly - fancy, but definitely not always useful.

[dampingwire]

When they have Japanese, that would be the real test for me.

Couldn't agree more. Especially if they can get Japanese done by the first Sunday of December and they add a direct neural interface

I tried a snippet in Italian (from Corriere della Sera) and it did reasonably well. When I threw the same passage at Google Translate it was neither significantly better nor significantly worse. A few areas where DeepL struggled, GT was fine, but the reverse was also true.

[Random Review]

I wanted to verify the meaning of a Clozemaster sentence so I checked first google translate and then, remembering this thread, I checked DeepL.

Ich möchte ihn nicht.

I don't want it.

I do not want him

I don't want to kill him
I don't want to see him

Huh. Interestingly enough, the DeepL translation changes when I add the full stop to the end of the original translation, offering pretty much the same translation as Google. So clearly the translator anticipates the most likely ending to the sentence based on it's corpus and acts accordingly - fancy, but definitely not always useful.

I actually think this is legitimate, because of the way German moves second verbs to the end of the clause, there is no way of knowing for sure what that sentence fragment means without punctuation or a context. A human translator might reason that people can be lazy and there is probably some missing punctuation (and translate accordingly); but expecting machine translation software to do that is not really realistic. I think the fair test is how it copes with all appropriate pronunciation.