US immigration uses Google Translate to scan people's social media for bad posts – Er, don't do that, says everyone else Google recommends that anyone using its translation technology add a disclaimer that translated text may not be accurate. The US government's Citizenship and Immigration Services (USCIS) nonetheless has been relying on online translation services offered by Google, Microsoft, and Yahoo to read refugees' non-English social …
I had to quickly translate an user documentation from English to German. I told my boss it would take 3 days, I was give 4 hours and told to put it through GT. I tried bunging it through Google Translate.
After I stopped rolling on the floor laughing, I went back to my boss and told him 2 - 3 days, unless he wanted it wildly inaccurate.
It converted phrases like "do not open the case, high voltage inside" to the German equivalent of "open the case, high voltage inside". I thought Google was trying to kill me... The best was "do not open the case, no user serviceable parts inside", which came out as "open the case, nothing inside", not something want to read, when handing out over $3,500 for an industrial terminal!
I've also had friends send me dissertations, auto-translated from German to English, because they have to post an English version or CVs in English. The translations are generally laughably inaccurate and don't leave a very good impression. I would say, in general, about 10-20% of what Google Translate chucks out is usable - depending on context and subject. That said, it is probably still one of the better online translation services.
My translations are generally okay, but I did an internship at a translation bureau and they are at a whole other level. Translation is a very difficult thing to do, somebody who is fluent in both languages can run rings around Google Translate, and that is still baby-steps compared to what a professional translator does.
My wife's CV was once translated using GT by a colleague. She apparently became an E.U. "deep cohesion expert", whatever that might be (instead of "cohesion fund program" specialist).
Although I must concede that GT has been greatly improved over the years and does a lot less stupid mistakes like that one (and it also seams to learn take user corrections into account, although I have no guess about it beeing automated - i.e., machine learning - or done manually)
GT is generally good for single words or short phrases. As long as you have some knowledge of the languages in question and the context, you can fill in the blanks yourself. If you have absolutely zero knowledge of the language you're translating to, don't trust the output blindly!
In the last year or so I have been using deepL translator, it's much better than GT, being especially good at grabbing some context and adapting the translation. Of course, still far from perfect, but for people at beginner's level in one of the languages, it's a great tool.
Covering what another poster mentioned above with German technical translation, German constructions around negatives can be very particularly nuanced and it's easy for a beginner (or a machine translation) to mix up what is essentially a *do* with a *don't*
Google does indeed have a habit of turning negatives into positives in languages like German or Czech, but that's a limitation of the machine.
On the other hand, in French sometimes the only was to know if a sentence has a positive or a negative meaning is to ask the author to pronounce it.
It can usually, but not always, be inferred from the context but for instance « il y en a plus » means either “there are more” or “there aren't any more [left]”, depending on whether the ‘s‘ is pronounced or silent.
And in the latter case it could mean either “there aren't any more [left]” or “it rained some of that”.
[...] and it's easy for a beginner (or a machine translation) to mix up what is essentially a *do* with a *don't*"
I use Google translate to handle web page paragraphs in various languages - mostly European. It quite often drops a negative - even when it is absolutely clear in a free-standing word. It also silently drops other words in a sentence - and sometimes puts them in twice. Occasionally the translation is just a jumble of English words making no sense at all.
It is not a problem for me in the Germanic/Romance languages with which I have some familiarity - otherwise it can get tricky.
My first job in IT was translating manuals and software, so I would read an English paragraph for understanding, then write the same content in German, often badgering colleagues who knew the topic for details so I wouldn't introduce errors.
I learned quite a bit about image processing, typography, and DTP in general during those times...
So far as I can tell GT relies on people submitting better translations than the ones it has come up with, so that GT can “learn”. I wonder if it’s just so bad that people who can spot the errors are just too overwhelmed by the sheer scale of GT’s uselessness to bother sending in corrections.
Plus there’s plenty of room for japes and giggles. If enough people put in a duff and funny translation, GT will start believing it...
essentially, it's one of the tools (crowdfunding), and no serious translator would do that, because they have bills to pay and to do that they have to work. For money, not for the sense of achievement (never mind the feeling you're undermining your own future). So yes, 10 million monkeys will produce some results. It won't be of high class, but how many people want to struggle with Chaucer, or even contemporary literature? BO-RING! Simplify, dumb down, optimize, outsource your brain! Playing 1st person shooter is so much more fun than climbing pesky language constructions.
My favourite, from a translated list of countries
Bosnia
Croatia
Slovenia
Serbia
Fruit Salad
If you go into a cafe in Italy and ask for "macedonia", they will bring a fruit salad to the table. My list of countries was in an Italian document. Google couldn't tell that it was refering to the country rather than the food product, but it would be instantly obvious to any human translator.
my experience is different (and I hate to say it, because I'm a translator, though the language pair is different). Recently (over last 2 - 3 years) I have noticed that google translations have become much more than just barely tolerable, i.e. I could finally understand the translation without struggling too much. And, more to the point, I have recently tried a different auto-translator, which turned out even better than google. To test it, I used it for a shortish, easy translation job, and, in general, there were very few howlers to require my intervention. I mean, sure, it is not exactly machine translation, it's just a vast, and growing fast, database of human translations with more rather than less perfect matches. I remember that test a few years ago, when they combined a large number of pics of female faces. The result wasn't a Queen of the Universe, but it wasn't fugly either. It was bland, but fairly likeable. And this is the direction of "machine" translation, I think. Bland, but good enough.
p.s. one of the more popular translation courses at one UK university recently has been something along the lines of Machine Translation post-edition. So, over 30 years or so, the roles of central/peripheral elements have shifted already. from a word editor spell-checker, not long ago (with its own, human-introduced follies) to the human verifier gaining traction. Time to die. Or become a truck driver...
I'll point out there is not much that Umberto wrote that one should *not* read. Fiction/essays/notes on the back of matchbooks. I rather like most of what he's written, and that which I don't like at least makes me think about the world.
old story, possibly apocryphal
translated a sentence from English into Japanese, then taking the translation to another Japanese speaker to ask them to translate it back into English
SENT ---- Out of sight, out of mind.
RETURNED ---- Invisible idiot
and that wasn't even using AI, just good old people power :o)
anything going back to pre-internet times is apocryphal (or so my kids say). And, every story that comes from internet times may be fake. It doesn't look good for the post-internet times...
p.s. it wasn't an "invisible idiot" but a "blind idiot". Also, EN-RU (or the other way round) translation of "the spirit is willing but the flesh is weak" was, allegedly "vodka is strong, but meat is rotten".
ah, there it is:
www.hutchinsweb.me.uk/MTNI-11-1995.pdf
1. strangely enough, the most often quoted variant in that article is "invisible", rather than "blind" (and I would swear I read it as "blind", in a book from early 1990s
2. the article, which comes from those almost-pre-internet times (1995) comes with a quote which everybody should find familiar:
If nothing else, this should serve as a warning not to believe everything read in newspapers
(A. Hitler)
That's actually not too bad a translation, certainly compared to the other one
Out of sight - I can't see
Definitely says the same thing
Out of mind - I don't care
Describes the same state of being, but different intent.
But overall, the meaning is much the same.
The statement "USCIS follows up with human translators as needed", if true - and no reason has been presented to doubt it - makes it clear what's going on here. Use Google Translate, in conjunction with Search tools, to scan for words/phrases that might suggest an issue exists, then use that information as a filter to determine which posts/feeds get attention from a human translator.
Seems eminently sensible to me.
Now, granted, I don't know if that's what the story means. But, unlike other interpretations, it's 100% consistent with all the stated facts.
What you are saying also crossed my mind, and I hope that is true.
However, there is a downside to this. Only posts who's machine-translation is flagged as concerning would get referred.
But this means they will miss actual terrorists who's translation was incorrect, but in the 'this isn't a problem' direction.
That is, this system means a translation something like the following would be referred:
native text: I want to go and sledge the All Blacks at tonights game.
machine translation: I want to beat up all the blacks at tonights game.
Hopefully the referral would give an accurate translation of a common, everyday social activity - going to a football game and shouting abuse at the opposing teams players.
native text: I want to bomb the blacks at tonights game in Durbin.
machine translation: I want to abuse the All Blacks in Durbin.
The translation seems to imply that the post is talking about the afprementioned acceptable social sporting activity, this will not get referred for an accurate translation of it potentially being an actual terrorist activity.
Therefore I conclude that the process is flawed, and they are just as likely to let in 'bad actors' as they are to incorrectly detain 'good actors' while awaiting an accurate translation.
Yes, Google Translate has real problems translating from formal English. For the longest time it didn't know what to do with the word "not" and just dropped it out of the translation completely!
Do not enter the cage, dangerous animals -> Enter the cage, dangerous animals.
The story cited of Ismail Ajjawi might suggest there's a higher and more problematic hurdle somewhere in the system. Although there's no indication (other than the fact of it featuring in this article) that translation issues played any role in that story.
All your security are belong to us.
Although speaking and comprehending the languages, for written work I regularly use Google translate between English, Catalan, French and German. The first three are no problem: its the German with its weird grammatical structure which doesn't work well. Putting the verb at the end was not a good idead.
" Dear friends and relatives I must today announce to everyone to our great sorrow that after a long illness and despite much suffering, our grandfather has himself shot.....
well, I say it IS very near good translating (and I'm serious). But, like your claim, mine is based on a very limited number of translations, although I've already switched to their engine for when I want to read what I can't understand.
"Putting the verb at the end was not a good idead"
A supposedly true story from the EEC (EU) about a real time human translator.
A German speaker was being very voluble - then the English translation went silent. Suddenly there was an exasperated - "the verb, damn you man, the verb!"
I think the problem is threefold.
1) People, when they are young, can do things they (and other people) end up regretting. We've seen it with James Gunn. We've seen it with a dozen other famous people when they have to apologise because 15 years ago, they made a joke others now find massively offensive.
2) Machine translation is rather basic, and often wrong. I'm no translator, but I was listening to someone talk about the problems of machine translation while talking about a kickstarter for a little bluetooth headset that connects to an app on your phone that translates spoken words. He said the problem is that while machines can translate the actual words easily, different languages have different syntaxes, so a sentence might have a totally different meaning from that intended when it is translated (sometimes with potentially dangerous consequences, as noted above) even if all the words are correct. Certain words in certain languages also change meaning depending on context. He was talking about translating the spoken word, but the comments apply to the written word as well. A human translator will often pick this sort of thing up, and will adjust their translation to deal with it.
3) People are just going to learn to avoid posting radical things on social media, or create false accounts with generic, non offensive postings to get past any checks.
4 anyone who speaks any giddam foreign took language is probably. A goddam foreign gook who can't be trusted. Or if they are an american who learned foreign they are obviously a traitor.
This was actually the reason for the founding of a lot of AI research in the 59s. They needed to translate a lot of Russian, but anyone who knew Russian must be a communist.
What's scary is that there're actually quite a few people who actually do think this way.
OTOH, there're also quite a few people who think that learning (or failing to learn) a language is something that randomly happens to people. I see this often in 'heritage language' situations. Those who failed to learn it more often than not refuse to acknowledge that family members who did learn it actually put any effort into it. Yes, I'm aware that makes no sense, but there you go.
"Certain words in certain languages also change meaning depending on context"
One of my favourite examples...
"bum a fag"
UK English - fairly common and completely innocuous expression meaning "borrow a cigarette"
US English - highly offensive for both sexually graphic content and a homophobic slur
The US army had a similar problem in Iraq.
Written arabic is one language (=the Quran) but spoken Arabic is as different as Spanish/Italian/Portuguese.
Obviously you can't hire Iraqi translators (security), so you hire from friendly Arab countries = Egypt/Jordan/Lebanon
A journalist was asking one Egyptian interpreter how he was doing interrogating Al-Qaeda prisoners - I can't understand a word, but you learn what they - the US army guys - want to hear.
Ironically the US army has some of the best language schools in the world, specialising in places that the US might have as enemies years in advance. But they are so prestigious that no graduate of them is going to be anywhere near the front line.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
