Hacker News Reader

Comment by distances | Hacker News Reader

> So far, I never felt that the translation is off, or terrible, and if not better in some cases, it's as good as Google Translator.

In my experience Deepl is consistently, without fail, considerably better than Google Translate. I basically use Google now only for more exotic language pairs, or full-page translations.

View on HN

> more exotic language pairs

I've found that sometimes what Google does for these is translate from Language A > English, and then English > Language B, which leads to bizarre results.

I have noticed that too. When translation libre from French to Azerbaijani, it gave the Azerbaijani word for free as in free of charge, rather than free as in freedom. That is an ambiguity that is mostly limited to English, maybe some other languages, but definitely neither French nor Azerbaijani.

From reports, Google Translate has effectively created its own internal (AI GD ML) metalanguage, which can interpolate between languages it's not been specifically trained on. E.g., with Japanese <-> English and Korean <-> English, Google Translate can manage Japanese <-> Korean, without being specifically trained to do so.

So yes, there's an intermediary language. But it's not English.

https://www.newscientist.com/article/2114748-google-translat...

The intermediary language is not English, but due to the way the training set is constructed (pairs of texts in various languages, with the vast majority of pairs having English as one of the languages, is my understanding) it can be very hard to tell apart from English sometimes.

For example, translating "рубанок" ("plane", in the sense of the carpenter's tool) from Russian to Polish used to produce "samolot" ("airplane") in Google translate up until sometime earlier this year, because in the intermediate representation "plane" was ambiguous just like it is in English. It looks like that particular bit is fixed now, which is at least progress! Maybe they've been adding more non-English text pairs...

That's almost certainly accurate. The metalanguage / interlingua isn't English, but being based on A <-> English and B <-> English training, is all but certainly influenced by English grammar, words, and idioms, in ways that direct A <->B training would not be.

It seems to be 'fixed' now, but once I was trying to translate from Hindi to Nepali (which are actually closely-related languages, think Italian and Spanish) with a simple sentence along the lines 'Ram came', where 'Ram' is a common Hindu name (effectively: 'John'), written in Devanagari with a long vowel: राम (rām).

And I gave the Hindi input in devanagari (राम आ गया), but still the Nepali translation ended up being the equivalent of 'the sheep came' (भेडा आयो), so somewhere along the line it seemed to be treating the name राम (rām) as equivalent to the English string 'ram' and translating accordingly.

So if the intermediate language isn't English, it certainly has some English-like properties....

This is interesting! So its not just about grammatical patterns, but other stuff that might be parsed as named entities.

But it's bizarre that, even if for some reason it doesn't recognise common Indian names, that it just treats 'unknown strings' as English words.

Supposedly, but it behaves suspiciously like English in practice, perhaps because of the input data (lots of texts originally in English then translated to many languages and fed in)

Since we're complaining about Google Translate, I'd like to mention how ridiculous their "verified translation" system is. It works by throwing automatic translations at people who, in their majority, have never studied English, and expecting them to tell whether it's right or wrong, but what happens is that most just confirm whatever they get as being correct. As a result, at least for Portuguese, many of them, if not most, are just plain wrong.

Considering Translate is such an important product, I can't fathom why they just don't hire a single linguist (or just anyone who isn't completely clueless, really) per language to register decent translations, or at least import them from a real dictionary...

Well, the more general problem about asking people visiting Google Translate to verify if a translation from A->B is correct is that generally people visiting Google Translate didn't know how to translate A->B or weren't very sure.

How many people on Google Translate actually are able to reasonably verify translations? Relatively few - and those qualified few who might poke at Translate out of curiosity are just as likely not to feel inclined to offer free labour to Google.

> Considering Translate is such an important product, I can't fathom why they just don't hire a single linguist (or just anyone who isn't completely clueless, really) per language to register decent translations, or at least import them from a real dictionary...

Perhaps professional translators rather than linguists. I imagine they have some linguists on the project, but they're likely to be more NLP-type linguists.

The difficulty is that they're interested not just in word-level meaning/translation-accuracy, but also phrase- and clause-level accuracy, and those are really large (i.e. theoretically infinite) spaces.

I've heard that the Spanish<->English translations aren't too bad.

Sometimes I have seen English words injected verbatim/untranslated in the middle of a phrase, when asked to translate between two non-English languages.

This is pretty standard, even in human (non-automated) translation.

One example is technical/service documentation for a heavy machinery company I worked for. The tech writers were based in Germany, but O&M Manuals were required in Japanese for sale in Japan. Those docs were translated using English as a "pivot language". Usually dictated by pricing (fewer German+Japanese translators = much higher cost).

But one advantage of automated translations is that there shouldn't have to be a pivot language.

And in any case, for German->Japanese, going through English probably has a lower cost. But for Hindi->Nepali, you'll lose a lot of information as Hindi and Nepali are closely related and similar not only in terms of correspondences between vocabulary items, but also grammatical structures, which is effectively 'thrown away' if there's an English, or close-enough-to-English-to-effectively-be-English intermediate translation language. (Not to mention the inefficiencies of the equivalent of sending a package from Delhi to Kathmandu via London.....)

I think in the case of automated translations it's a function of training data and confidence rather than cost. If you don't have a corpus of translation data for the source/target language combination to draw from, you're essentially forced into a pivot model.

To be frank, the Hindi<->English and Nepali<->English translation results are pretty poor to start with.

For Spanish, at least, it definitely isn't better than Google.

And recently everything I had to look up while reading Gabriel García Márquez, DeepL didn't know. After enough failures I gave up and returned to Google Translate for the remainder of the book.

Google may have a wider range of words, but DeepL is definitely much better in grammar and idioms, which Google tends to translate literally.

Moreover, Google often provides a single target word; whereas DeepL allows you to select from a range of synonyms clicking on a word, and will adjust the sentence accordingly to use the new word. When Google gets the context wrong and provides the wrong meaning for the translation, DeepL's capability to translate with a different meaning is invaluable.

same here. Admittedly it supports just a few language pairs but the translation quality is consistently and considerably better than Google Translate and other major offerings.

It's way off 'never off', I work with colleagues who use it en <> de.

For example it will always translate 'order' in a sentence as 'Befehl' (I order you to fix a steak) instead of 'Bestellung' (I order a steak).

Both are correct, but completely different, and in our context we never mean the former.

Reply on news.ycombinator.com