Perhaps the AIpocalypse isn't imminent – if Google Translate is anything to go by, that is There's been quite the little chortle in this part of Central Europe this week regarding the actions of a tourist board in Moravia. To set the scene, Moravians are thought of as the slightly slower country brothers of the Bohemians, (or “true Czechs” ... one local bar has “We have Moravian and Czech wine” in the window) in …
The irish tell jokes about people from Kerry. People in Kerry tell jokes about one particular village. The people of that village tell jokes about one family. The people in that family tell jokes about their dad. Who actually does all these things. He gets home from work and says 'I saw a guy on the bus today. He had custard in one ear and jelly in the other. I said "Are you a trifle deaf?" and he said, "No, I'm a mental patient."' And the family stand around saying 'I don't get it'
-Paul Merton
Unlike the US , where Racial Theory and it's subsequent Eugenetics "science" was actually conceived and heavily promoted, even up through the 60's, amongst scientists and government, and still holds firm ground in quite a few of the more ....conservative... areas ( be it physical or political) ?
The only difference between Germany and the rest of the "Civilised World", is that they incorporated and formalised the concept with the usual Gründlichkeit into Policy. In the US ( and the British Commonwealth/Colonies) it was already so ingrained into culture it didn't need formalising. It was already in effect in full force, and it got worse in the US after WW II.
Oh , and the basic concept of the concentration camps weren't new either.. The Nazis were quite taken with the efficiency of Dealing With Things in the US, and simply copied the concept of native-american "reservations", and put them through the Gründlichkeit mill.
While we're tangentially talking about the Irish and mistranslations, let's not forget the Gardaí scooping the Ig Nobel prize in Literature back in 2009 for something similar.
The economics of why they couldn't hire a professional translator despite having a huge budget. Or cultural working practices which allow bodge jobs like this.
Not AI. Who said AI has to understand the world using human language? The original Terminator used 6502 machine code and still managed to wipe out a load of humans.
The usual 'fiddle' for euro-dosh, in fact it hard to really call it a fiddle. EU offers money for regional development/ tourism etc. Enterprising mayor or other local official plays the game for all its worth. Local economy benefits. Keeping the benefits local means employing who is available locally rather that bringing in qualified outsiders.
I know a village in Italy which has an extraordinary number of brown tourist signs and maps of (imagined) footpaths and cycle ways. There are few tourists but it keeps the local signmaker in business
ref. economics of big budget v. AI translation? Simples. They apply for an EU grant because they're starved of local (national) funding. Or they believe they're starved. Or believe they should get more funding. Or somebody showed them how it can be easily got by applying for a grant. So they get back to this grant application support service (you can't apply for grants directly, and where you can, rules are fiendishly complicated), which take a fee. Plus a cut from a grant once receive. Actually, 25K is peanuts, I once saw a grant worth a couple of milion EUR, to restore and scan several dozens old maps at some Lithuanian library. That was a good few years back, though the scans haven't surfaced yet, one must be patient with such maps). So, they get a grant, buy a couple of pretty toys (at hugely inflated prices), also find a couple of "roles" for their existing staff to maintain the project, then obviously comes a web designer biz (completely unrelated to the grant applicant and middle man) with a turnkey solution, including the abovementioned cost-optimised translation. End result - everybody happy.
Why, are you not, the dear end user and potential income source? Well, fuck you, we did our best to make you come!
On a relevant and related note; IIRC Arnie is always dubbed into German by someone else, despite speaking the language natively. Apparently he hails from a rather bucolic corner of Austria and nobody is scared of a killing machine that says:
<Wurzel>
Oi'll be baaaach!
</Wurzel>
Who said AI has to understand the world using human language?
The people who want to use it for Natural Language Processing.
I suspect someone will say Turing did, thereby demonstrating a complete failure to understand Turing's point - which was essentially a rejection of metaphysics over pragmatism in developing an epistemology of mind.
Saved my lard, in French speaking bits of Africa. Anyway, hints and tips:
- Replace proper nouns with something known, before you translate (and put them back in, after translating).
- Same with any industry specific technical terms, as they can cause sentence structure to break down badly.
And remember, it's a courtesy machine translation, not an attempt to trick people into thinking you know the language (Google Translate will gladly prove otherwise). For web sites. brochures and so on, consider employing an actual human translator. They'll appreciate the business.
Now, did you hear the one about the Irish.....
No its utter crap. I regualrly use it to 'translate' German and Italian. It puts words in the wrong order, often says the opposite of what the original lanuage says and for some stupid reason translates people's names. Sometimes you need a degree in translating the output of Google Translate to make any sense of it.
This post has been deleted by its author
@Irongut agreed.
I had to translate a user manual into German from English. I thought I could save a bit of time by putting it throught Google Translate and just tidying up. After my sides stopped hurting and I could climb back on my stool again from laughing so hard, I started from scratch.
Translate seems to have real problems with formal English as well. If you use "do not" it translates it as "do", if you use "don't" it translates it as do not... Not very helpful if you are writing safety instructions. It translated "do not open the case, high voltage electricity inside" into the German equivalent of "open the case, high voltage electricity inside." Google had obviously noticed that I had swapped my Android phone out for a Lumia.
Or, how about, "do not open the case, no user serviceable parts inside"? That got translated into "do open the case, no parts inside." I was robbed during translation!
I wouldn't like to use it to translate into a language I didn't know - that's what professional translators are for - but for translating some unknown foreign text into 'usable' English it is pretty good. And free.
I wouldn't like to use it to translate into a language I didn't know - that's what professional translators are for - but for translating some unknown foreign text into 'usable' English it is pretty good. And free.
I found, for my industry (software for the food industry, especially meat processing), Google is fairly useless at going back and forth between German and English.
Translating documentation and press released that my company produces (I am theoretically a project manager, but have to rush out translations all the time), I find that Google is good for about 20% of the translation, I tend to use Leo and Linguee much more often. I have noticed that Bing Translate has improved recently, I did a head to head between Goolge and Bing on a press release, around 20% in Google was usable, in Bing probably closer to 40%.
Using your noggin and Leo / Linguee is still a much faster, safer and more accurate translation, if you understand both languages.
No its utter crap.
At least it knows the difference between "its" and "it's".
It's very successful at doing what it was designed to do. You're asking it to do what it's advertised as doing, which is quite a different thing.
I find Google Translates passable if a) you're translating into your language (i.e. for informational purposes) and b) you have at least some knowledge of the from language to spot the obvious errors. Even then, I've had it completely reverse the meaning of a sentence - I've no idea how.
My favourite Google Translate error, though, was merely a matter of it trying to be too clever. We do various Oracle DB software, and one of our Swiss customers raised an issue talking of "der DB-Fehler", which Google Uebersetzer decided to render as "the German Railways error".
"My favourite Google Translate error, though, was merely a matter of it trying to be too clever. We do various Oracle DB software, and one of our Swiss customers raised an issue talking of "der DB-Fehler", which Google Uebersetzer decided to render as "the German Railways error"."
Dealing with that is easy. Replace DB with Pig, before translating from English to German. Then replace Schwein with DB, in the German translation. With this approach, you stand a chance of the grammar butchered but still good enough for a native German speaker to get it.
I've had it completely reverse the meaning of a sentence - I've no idea how.
Not at all an unexpected result if you do a bit of research into linguistics and machine translation, particularly the huge-corpus approach used by Google.
You take some corpora in various languages that have been annotated by human judges - parsed, basically, into part-of-speech and phrasal grammatical structure. You train some ML model (Google likely use NNs, because they have an unnatural affection for the things, but you could use HMMs or MEMMs or various others) based on those inputs. Presto, you have a probabilistic model for parsing the language.
Then you take your huge corpora of texts that you have versions of in multiple languages. Remember when Google indexed the web and scanned all those books? Yeah, that. You use that to build probabilistic maps from language A to language B, for various pairs {A,B}. You can use longer chains (A->B->C) to add more information to the model, but you have to weight it lighter because successive translation introduces more noise, so there's a point of diminishing returns.
When you get fresh input to translate from A to B, you first check to see if you already have a translation of the whole thing, or of big chunks (sentences, say). If not, you parse the input into phrase-level chunks, using a model from the first set, and then translate those chunks, using a model from the second.
This sort of approach, with a bit of tweaking and a really big set of corpora to train from (which is what Google have) has a pretty good success rate - somewhere in the 90%-95% range on typical inputs.
What about the sentence-meaning-inversion thing? Natural languages have many, many ways to invert meaning at the phrase and sentence level, and language users keep coming up with new ones. Often this can hinge on the presence or position of a single word, or punctuation, or context from other sentences in the text - consider sarcasm, for example.
And then there are sentences which are simply ambiguous in language A but can only be translated into one of a set of non-ambiguous sentences in language B, for example because of grammatical inflection in B. (All natural languages admit all sorts of ambiguity, of course, but they have different constraints on its particular forms.)
Meaning inversion is a very easy trap for blind-translation models like Google's to fall into, because a given phrase often has inverted-meaning local maxima in its probability distribution.
I've found out the same thing.. it's mostly usable if.. I do a lot of communicating with several Frenchmen on a hobby interest (18th Century Warships) so what we do is type emails in our native language and paste in the Google Translate. Give many terms are archaic, we manage pretty well all things considered.
Google Translate often offers alternatives. Presumably all the "corrections" that users are invited to submit are given a weighting depending on the number of times they have been submitted. It is tiresome having to keep correcting its translation of the German "messe". It uses "fair" rather than "mass" - although possibly that is down to it not understanding the context.
I use it every week when trawling web pages in English, German, Austrian, French, Dutch, Italian, Spanish, Catalan, Norwegian, Danish, Swedish, Icelandic, Russian, Polish, Latvian, Estonian, Lithuanian, Bulgarian, Czech, Slovak, Chinese, Japanese - plus a few other less common ones.
A slight problem is that it often completely omits words in the translation - usually adjectives. Yet elsewhere in the same piece it shows it can translate that word.
Fair is the correct translation for Messe, though not fair as in fair play, but as in a fair being held in a fairground. The translation of "mass" would be "Masse", not "Messe".
As for the corrections: that options is gone from the mobile app and I don't have the idea that Google does anything with it. It still translates Dutch "zwaluw" to the French "avaler" instead of "hirondelle".
Why? Because a zwaluw is a swallow and the verb to swallow is avaler in French.
So another tip for using Google translate: translate from English, even if English is not your first language. (Although in this case you still don't get the bird, just the verb)
This is quite interesting. Out of curiosity, I just typed "Is that a swallow" (no question mark) into Google Translate and I got the expected "est-ce une hirondelle". The funny thing is, it changed into that at the last moment - while I was typing, one letter short it was still showing "est-ce une déglutition" which seem to be related to the _other_ noted meaning...
Maybe the addition of "a" or "een" signals that the verb should not be used.
Think of the input as an N-dimensional vector pointing into an N-space of possible translations. Each complete word adjusts that vector. Google shows you the data point in the translation space that's closest to the end of the vector.
<<It [Google] still translates Dutch "zwaluw" to the French "avaler" instead of "hirondelle".
Why? Because a zwaluw is a swallow and the verb to swallow is avaler in French.>>
"One swallow does not make an orgy" - the late, great Willie Rushton on I'm Sorry I Haven't a Clue's complete the proverb round.
"The translation of "mass" would be "Masse", not "Messe"."
I have seen the German "Masse" translated as "mass" when talking about the mass property of an object.
German music sites use "Messe" for the religious service and the English translation they give is the expected "Mass". Google Translate usually gives "fair" or "trade fair".
Googled using the keywords : facebook knabenchor messe
Facebook postings for several German choirs use "Messe" in the context of a religious service.
Here is an interesting example where in one paragraph Google gets it right and wrong. Correlated with other examples - Google Translate appears to apply the correct context if it recognises the format of a church name in the same sentence. Bold indicates the translated words.
"Mit einem Foto von der Abendmesse in San Bonifacius, Verona, und einem anschließenden kleinen Konzert gemeinsam mit dem Chor Hartolan Viihdekuoro aus Finnland wünschen wir eine gute Nacht. Am Sonntag singen wir noch einmal in einer Messe in Verona"
"With a photo of the evening Mass in San Boniface , Verona , and a subsequent small concert together with the choir Hartolan Viihdekuoro from Finland , we wish you a good night . On Sunday we sing once more in a trade fair in Verona"
The robots, the algos, unrestrained aren't about to take all our jobs. Simply because they're not yet very good at doing things which we humans do without much effort, which is to distinguish between different potential meanings of words and put them into context on the fly.
This project strives to solve exactly this problem. Not yet there, but...
You have to be so careful with automatic translation... Machines have no idea about idioms and colloquialisms, so avoid them like the plague.
Sometimes people just don't realise that their local name for something isn't universal... This is especially true of food. I've seen English menus which are a literal translation of the local language, which are no more help to me than the original.... "Princess steak - Steak cooked in the traditional princess style" (I think it was princess)... Perfectly good English, just completely useless!
If you want to see truly bad machine translate, try bing. I've yet to use it to translate any of my foreign Farcebook friends into anything more than gibberish.
Ah yes, reminds me of my favourite restaurant in Brussels. Now long since ruined. Although, I've not been in a while, so I can hope for redemption.
They had menus in 6 languages. But the English translation of tete de veau was "tete of veal". Which is all very well if you remember that tete means head - but if you just think it's a french cut of meat, then you're in for a surprise when you open the stock pot and there's a baby cow's head staring up at you.
There weren't many screams while I was eating there, so I can assume that not too many people made that mistake.
Or perhaps the waiters were aware, and warned people. My brother bought some horse from the butchers, when on holiday in Sardinia. It was deliberate. But the butcher knowing that the English are traditionally squeamish about eating horses tried to make sure, to his credit. He didn't speak english though, so this involved making ear signs with his fingers while blowing out his lips and making brrrr and whinnying noises.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
