Translate at your own risk
« previous post | next post »
Last month I posted a link to a Schott's Vocab Q&A with Claude Hagège on endangered languages. Some commenters immediately picked up on one of Hagège's statements about translation:
However, there exists an important activity which clearly shows that even though the ways languages grasp the world may vary widely from one language to another, they all build, in fact, the same contents, and equivalent conceptions of the world. This activity is translation. Any text in any language can be translated into a text in another language. These two texts express the same meaning. We can therefore conclude that despite the differences between the ways languages grasp the world, all languages are easily convertible into one another, because humans interpret the world along the same, or comparable, semantic lines.
Barbara Partee contributed this comment:
Emmon Bach has put it nicely: The best argument in favor of the universality of natural language expressive power is the possibility of translation. The best argument against universality is the impossibility of translation (i.e. that we often can't really translate exactly). [link added–EB]
Translation ain't easy, even for skilled humans — and (especially) for machines. Google Translate appears to be among the better tools out there, but as the comments section of what (I believe) was Language Log's first reference to Google's translation tool shows, you can have quite a bit of fun breaking it. Moreover, breaking it is easy and can happen completely inadvertently, a lesson that (from what I hear, anyway) is quite often learned too late by desperate students trying to take shortcuts while doing their homeworks for beginning language classes.
Almost exactly a year after that Language Log post, Google Mail added an automatic message translation tool as a Gmail Labs setting. I enabled the tool in my Gmail account and noticed that it easily recognizes every message written in Spanish that I receive from members of my family, suggesting that I translate the message from Spanish to English (and offering drop-down menus with other languages in case it had made the wrong guess, or in case I want to play around). So every once in a while I click on the "Translate message" link and casually examine the results.
Most of the time the translation is imperfect, but the gist of the original appears to be there. A couple of months ago, though, one of my aunts sent a message to several family members on the occasion of the 20th anniversary of my grandmother's (my aunt's mother) death. The most dramatic mistranslation was of the closing line, in which my aunt addresses her mother directly:
tu hija que te quiso tanto, tanto y no supo demostrarlo – PERDONAME.
Which I would translate as:
your daugher who loved you so, so much and didn't know how to show it — FORGIVE ME.
But which Google translated as:
you wanted your daughter so much and failed to prove it – Perdoname.
Some of this mistranslation is probably due to some systematic ambiguities. The verb "querer" in Spanish, represented in the original by the third person perfect form "quiso", does in fact mean both "to want" and "to love". Likewise, "demostrar" means both "to show" and "to prove", just as "to demonstrate" (or even "to show") does in English. Finally, the "tu" at the beginning is ambiguous between "you" and "your" (putting aside the negligible fact that "you" should be spelled "tú", with an acute accent). Throw all those together, shake it up a little, and produce an intelligible English sentence, and the Google translation is at the very least a possible outcome. But I got interested in the fact that the all-caps "PERDONAME", meaning "FORGIVE ME", was left untranslated (though it was changed to the initial-caps-only "Perdoname"). So I changed it to lower case and ran it through again.
tu hija que te quiso tanto, tanto y no supo demostrarlo – perdoname.
you wanted your daughter so much and failed to prove it – pardon me.
Now Google translated the word, but again an ambiguity interfered: "perdonar" does in fact mean "to forgive" as well as "to pardon" in the more mundane sense (much as both of these English verbs are technically also ambiguous in the same way). But then I wondered about how other manipulations of case might affect the translation. First, all-caps for everything:
TU HIJA QUE TE QUISO TANTO, TANTO Y NO SUPO DEMOSTRARLO – PERDONAME.
YOUR DAUGHTER THAT YOU WANTED SO MUCH AND KNEW NO SHOW – Perdoname.
Not sure why that would change the first clause into a noun phrase with a relative clause — nor how "KNEW NO SHOW" popped in there. Next, sentence-initial punctuation only:
Tu hija que te quiso tanto, tanto y no supo demostrarlo – perdoname.
Your daughter who loved you so much and she could not prove it – pardon me.
That did it! The right gist is there, even though the redundant "she" makes it somewhat less than perfect.
I was still interested, though, in how the manipulation of something as relatively meaningless as case could affect the translation so much. So I picked another relatively meaningless part of the original: because "tanto, tanto" was simply being translated as "so much" rather than "so, so much", I simplified it to just "tanto". Here are the results:
tu hija que te quiso tanto y no supo demostrarlo – PERDONAME.
your daughter that you loved so much and she could not prove it – Perdoname.
tu hija que te quiso tanto y no supo demostrarlo – perdoname.
your daughter that you loved so much and she could not prove it – pardon me.
TU HIJA QUE TE QUISO TANTO Y NO SUPO DEMOSTRARLO – PERDONAME.
YOUR DAUGHTER THAT YOU WANTED TO SHOW BOTH and did not know – Perdoname.
Tu hija que te quiso tanto y no supo demostrarlo – perdoname.
Your daughter who loved you so much and failed to prove it – pardon me.
In my estimation, that last one is probably the best of all of the Google translations: the redundant "she" is gone, which more than makes up for the harsher sense of "failed" instead of "could not".
I don't pretend to know anything about Google's translation algorithm(s), but I do find it interesting that what seem like very minor manipulations like those shown above can lead to both bizarrely different results as well as to subtle improvements.
Brian said,
January 7, 2010 @ 7:55 pm
I know nothing of Google's algorithms either. But it's very easy (even tempting) to speculate that they started with existing machine translation programs, and built on that by leveraging the sort of learning that their search algorithms use to gauge relevance. This might well get you the sort of algorithms for which careful punctuation gives you translations that were based on more formal source texts, and sloppy or absent punctuation (to say nothing of all caps) gets you translations based on the slapdash and fragmented texts of casual online communication.
Chaz said,
January 7, 2010 @ 7:57 pm
I've noticed a lot of the same sort of things when using the Google service to translate from (or, to a lesser extent, into) German => English; some of that makes more sense, though, since capitalization plays an important part in German grammar, since it denotes nouns. Google also seems to struggle with larger complex words, but can often handle them surprisingly well if they're split into smaller chunks.
As a general insight to what you posted, I imagine that in the first example(s), the fact that 'PARDONAME' is separated by a dash, and in all caps, is causing the algorithm to assume it as a signature / quote attribution attached to the preceding line. Changing the capitalization normalizes it within the sentence, and causes it to be parsed alongside it. I'd guess that this was implemented to stop the service from "translating" email signatures and the like in cases where a name (or surname) has a direct translation.
John Cowan said,
January 7, 2010 @ 8:01 pm
GT uses a statistical phrase translators, and upper/lower case information can't be just discarded as markup can be (consider English "polish" vs. "Polish"). Consequently, unexpected uppercase can confuse the heck out of them. Consider what happens when I feed variants of "confuse the heck out of them" to the en-fr translator:
"confuse the heck out of them" => "confondre le diable hors d'eux"
"confuse the Heck out of them" => "confondre les Heck hors d'eux"
"confuse the Heck out of it" => "confondre les Heck Out of It"
"You confuse the heck out of it." => "Vous confondez le diable hors de lui."
"You confuse the Heck out of it." => Vous confondez la Heck hors de lui.
(Disclaimer: I work for Google, but don't know any material non-public facts about GT. The above results may not be reproducible, because GT knows I'm a Google employee and gives me test rather than production versions of the engines, so the results may be either better or worse than what the public sees.)
Alexander said,
January 7, 2010 @ 8:02 pm
I had a similar experience with Mandarin — I was writing a note to someone and using Google to sanity-check what I was saying. I'm not sure it was wise.
Jim Regan said,
January 7, 2010 @ 8:09 pm
If you add the acute accent to 'perdóname', you get 'you wanted your daughter so much and failed to prove it – FORGIVE."
Google use statistical machine translation, so algorithms have little to do with it – the translation is created by matching all the translations available for the different parts of the sentence, and then ranked against an n-gram language model of the target language to see how likely it is that those particular phrase go together, to assemble the translation. As case can be significant – acronyms are usually all upper case, proper names use an initial capital, etc. – it makes sense that it affects the translation.
(I work on open source rule-based MT my system's output (with the accent added) is 'Your daughter that wanted you so much, so much and did not know to show it – FORGIVE ME.')
Al said,
January 7, 2010 @ 8:43 pm
@John Cowan:
"confuse the heck out of them" suggests «confondre le diable hors d'eux" – same as for you.
"confuse the Heck out of them" gives me "confondre la Heck d'eux" – subtle difference
"confuse the Heck out of it" => "confondre la Heck Out of It" – Sweet! We get the same result even though I don't get access to the newest tools
"You confuse the heck out of it." =>"Vous confondez diable hors de lui." – Hmmm… I'm missing the article
"You confuse the Heck out of it." => «Vous confondez la Heck hors de lui." Choette! Encore resultats identiques!
Given the tiny differences in the results John and I get, I doubt major differences exist in the algorithms we access. The second-phrase result intrigues me a little, as it implies a refining of preposition translation (what a gawdawful morass that can be).
And for the record? It confuses the heck out of me where the article went in the fourth one.
Nanani said,
January 7, 2010 @ 9:22 pm
>Any text in any language can be translated into a text in another language. These two texts express the same meaning. We can therefore conclude that despite the differences between the ways languages grasp the world, all languages are easily convertible into one another…
As a pro translator, allow me to laugh.
Even without that oh-so-hillarious word "easily", it's still not true.
This is a topic I've put some thought into, recently. Unlike computer languages, which are all fundamentally a transition from human language to machine-level, which is a perfectly understood stystem, neither human languages nor the harware they run on are perfectly understood. So, even theoretically, it cannot be said that "any text can be translated into another language".
Furthermore, only when the translation is done -well- can it be said that the two texts express the same idea. That result is the ideal goal, not the output of translation by virture of being a translation.
Kudos to Google Translate, my non-sentient and very entertaining collegue.
Bob Moore said,
January 7, 2010 @ 9:28 pm
There are different strategies for handling upper and lower case in statistical machine translation. Because of the vast quantities of data Google uses for training (billions of words of parallel data, trillions of words of target language monolingual data), they seem to treat different capitalization patterns of the same word differently. So I am not surprised that with Google, just changing the case of a word can change the translation. Most groups who don't have the vast data resources of Google do something like lower-casing everything for translation, and then trying to guess appropirate casing ("true-casing" the the term of art for this) at the end. Google evidently feels that they get better overall results the way they do it (e.g., there is less chance of translating George Bush as some sort of plant), but it leads to oddities like this in some cases.
George Amis said,
January 7, 2010 @ 9:42 pm
@Nanani
Well said.
I'm not a translator, but I've studied English translations of the _Iliad_ pretty extensively. Given that the original is poetry, written in a special dialect, and archaic in some complex ways. it's often not even clear what the idea being translated is. As a result, even very 'literal' translations are at best rather distant approximations.
Rick S said,
January 7, 2010 @ 9:51 pm
Chiming in to agree with Nani and George Amis. I often encounter impossible-to-resolve conflicts in translation goals: Do you change the wording to approximate the cultural context, or preserve it and hope the reader has the cultural background to appreciate it? It's easier when you know something of your audience, but if you're translating Spanish Wikipedia to English, it's a crap shoot.
I have to say, one thing about your experiments surprised me. I have a sometime hobby of using GT to translate stories from San Juan's daily El Nuevo Día, then fixing the translations and resubmitting via Google's "Contribute a better translation" link. (It gives me a chance to add to my vocabulary and knowledge of figures of speech, and hopefully I'm improving GT in the process.) One of the things I regularly have to fix is an incorrect gender assumption for the omitted pronouns in Spanish (such as "The victim's neighbor, Ms. María Delgado, told police that he heard loud noises coming from the apartment…"). So I'm surprised you got "she could not prove it".
Stephen Jones said,
January 7, 2010 @ 10:19 pm
What can't be translated is ambiguity.
JonJ said,
January 8, 2010 @ 12:16 am
I'm a translator, too, and have a couple of ideas about this discussion.
1) These "machine translations" are miscalled "translations." Computers can't translate, in the true sense, because they can't understand language (meanings), and a piece of language can't be translated into another language unless its meaning is understood—it's the meaning that is translated. "Machine translation" is a commonly used expression, but it's a misnomer.
2) The statement that any text in one language can be translated into another language is true in some ways and not in others. It is certainly the case that all human brains are basically built the same way (neglecting minor differences), and it is well known that infants can learn any language they hear spoken around them, no matter where they were born, and who their parents were. So at a deep enough level, it is the case that we are all human beings, and if one works hard enough at it, it is theoretically possible to convey meanings quite adequately from one language to another.
On the other hand, since languages are deeply related to their native cultures, there are also obviously overtones to most texts that may be very hard to translate within the space of a few words in the cases of many language pairs. But these are cultural differences, which can be conquered to a large extent by explaining the source language's culture to people in the target language's culture. But cultural differences can be very hard to deal with. Consider, for instance, the cultural gap we currently observe between atheists and religious folk, even when they all speak the same language. Often it amounts to total mutual incomprehension!
J. Goard said,
January 8, 2010 @ 12:45 am
I want to point out again something that I touched on in the previous post. Even if we were to grant that, for each sentence of each language, its meaning can be rendered by some translation in every other language, it would not by any means follow that languages "all build, in fact, the same contents, and equivalent conceptions of the world." That's because words, grammatical constructions, and discourse markers, all have very different frequencies across languages, and consequently very different degrees of felt "markedness". Furthermore, the myriad phonological and semantic associations with other words will necessarily differ (as they most likely do across native speakers of the same language, as well).
So this part of the argument is as absurd as saying that, because I can manage to eat with a bottle opener, and open a beer bottle with a spoon, that they are basically functionally equvalent utensils.
Matthew Stuckwisch said,
January 8, 2010 @ 1:41 am
In my experience translating, one can always translate the denotative qualities of a text. It's the connotative/phonological/metrical qualities that can difficult or impossible to translate. But that could even within a single language. Poetry that rhymes and has proper rhythm in my SAE might make no (metric/phonetic) sense at all in Scottish English where different distinctions are maintained and mergers realized.
Fernando Colina said,
January 8, 2010 @ 2:32 am
Jim Regan: If I add the accent in perdóname Google translates it for me as as "Sorry":
"Your daughter who loved you so much and failed to prove it – sorry"
Maybe the algorithm has changed since you tried? Sorry conveys the meaning here, but it seems too informal to me, given the heavy topic. Again, perdóname can mean sorry (as when you accidentally step on somebody's toes) and also forgive me.
The bottom line is that Google does not know if you are talking about playing soccer or mourning for your mother.
Michael Dunn said,
January 8, 2010 @ 2:46 am
The Italian phrase "Traduttore, traditore" (to translate is to betray, or something similar) seem relevant, but transliteration failures are an issue too. I'm reminded of when I was first starting out in Arabic nearly 40 years ago I was required to translate a news article and encountered the phrase هو شي من . This looks exactly like the Arabic words "huwa shay' min", which every first year student knows, but which means the puzzling "He (or a masculine it) is a thing from …"
The meaning defeated me.
This was during the Vietnam War. It was Ho Chi Minh.
Peter Taylor said,
January 8, 2010 @ 5:19 am
What on Earth do they have in their database which allows them to translate
as
? I can just about buy ignoring the double tuteo and treating "quiso" as second person formal, but even so it's dropping "que te" from the output altogether (and inserting a personal "a"). A word-for-word translation would be better, however unnatural it might sound.
Also absolutely jawdroppingly astonishing is the mysterious appearance of "both" later to translate whatever isn't dropped of "tanto y no (lo) supo".
Felix Ahlner said,
January 8, 2010 @ 6:35 am
The translation of proper names can be quite entertaining. Placing some Swedish artists in the sentence pattern "Jag tycker om att lyssna på X och Y" ("I like listening to X and Y") gives:
Timbuktu -> Justin Timberlake
Cornelis Vreeswijk -> Cornershop
Tomas Ledin -> the Tomahawk (but Thomas Ledin -> Prima Donna !)
I think I first saw this in the linguistics blog from Stockholm University, but I can’t find the original post.
However, writing "Jag besökte Timbuktu" gives the rather more fitting translation "I visited Timbuktu", whereas "Jag såg Timbuktu" again refers to the artist: "I saw Justin Timberlake".
Neal Goldfarb said,
January 8, 2010 @ 8:36 am
I just wanted to give a shout-out to Douglas Hofstadter's Le Ton Beau de Marot, which is a wonderful book about the difficulties of translation.
(For those who have already read the book, click on the link and scroll down to the first review.)
peter said,
January 8, 2010 @ 9:28 am
Nanani said (January 7, 2010 @ 9:22 pm):
"Unlike computer languages, which are all fundamentally a transition from human language to machine-level, which is a perfectly understood stystem, neither human languages nor the harware they run on are perfectly understood."
This may be true of low-level computer languages, or even simple high-level ones, but it is certainly false regarding some current high-level computer languages. Programs written using agent-oriented programming languages, for example, or machine-to-machine-communications languages, are by no means understood, not even imperfectly. Morever, features such as multiple threads of control or the problem of the objective verifiability of programmer & agent intentions may make understanding of (programs written using) these languages impossible not just in practice, but also in principle.
We can't ever ultimately verify what is in someone else's mind, whether that someone is a human person or a sophisticated computer program. This is why complex software can behave in inherently unpredictable ways, as with various telecommunications and electricity network outages which hit the US North East over the last decade.
Nick Lamb said,
January 8, 2010 @ 11:50 am
Regarding computers as "perfectly understood" systems I'll offer the standard Computer Science objection to this characterisation which is the Busy Beaver function. Computers are symbol manipulation engines, even a simple machine (many orders of magnitude simpler than the one you're using to read this) is capable of embodying mathematical problems that we believe are not only unsolved but maybe even unsolvable. For example, it might list all even numbers greater than 2 which aren't the sum of two primes. That's great – except we don't know if there even are any such numbers (see Goldbach's conjecture) so we have no idea whether this program will do anything, we can only try it and wait and see.
When arguing that language translation is in some sense "impossible" it's worth keeping in mind that we can have these same difficulties while communicating with someone directly in a shared native language. What did my neighbour mean by "I see you're back" ? I have no idea.
language hat said,
January 8, 2010 @ 12:42 pm
It astonishes me that anyone could say this with a straight face:
I would not pay attention to another word such a person said, because they clearly have not given a moment's serious thought to the problem. Sure, a translated text expresses the same meaning… except for the meaning lost in translation.
Peter Taylor said,
January 8, 2010 @ 1:56 pm
I understood Nanani to mean that if two languages are known to be Turing-complete then given a program in one it is provably possible to produce a program in the other which generates the same (lack of) output. Thus the halting problem, busy beavers, etc. are irrelevant to the question of whether translation is possible.
Non-determinism complicates things, but again it is possible to produce a program which will execute one possible run, or even (although with exponential blowup) all possible runs.
Stephen Jones said,
January 8, 2010 @ 5:30 pm
Arabic insists on using one of its three long vowels for every vowel in a foreign word. If you are not expecting the foreign name you're flummoxed.
They even transliterate silent vowels by long vowels, so Skype is duo-syllabic in its Arabic transliteration.
Alec said,
January 8, 2010 @ 5:48 pm
"Tanto" could be translated as "both" in a phrase like "tanto las vacas como las ovejas" – "both cows and sheep". But I agree it is nonsensical in the context of the post.
ellis said,
January 8, 2010 @ 8:45 pm
@Michael Dunn
I remember having similar problems over huwa gharatha in my oral final at Oxford. Hogarth, of course.
TB said,
January 9, 2010 @ 1:38 am
OK I'm confused about something. It's probably so basic as to be embarassing, but I really don't want you to think I'm being snarky or anything; I'm just actually confused.
In my experience, the LL comment community, such as it is, tends to come down hard on any sort of Sapir-Whorf suggestion, and on the idea that endangered languages should be preserved, and of course on the idea that there is or isn't "a word for that" in a given language says anything about the users of the language. But here, the idea that "though the ways languages grasp the world may vary widely from one language to another, they all build, in fact, the same contents, and equivalent conceptions of the world. This activity is translation. Any text in any language can be translated into a text in another language. These two texts express the same meaning" is the one getting hammered. But if translation is difficult or impossible, doesn't that mean that at least some of the ideas or feelings of a language community are also difficult or impossible to convey to another one? And if an idea is impossible to convey in the language that you speak, wouldn't it be very difficult to think it yourself? And if a language dies, and translation from it is impossible, then isn't it an irretrievable loss?
Essentially, I thought that what Hagège is saying here is what everyone says to dismiss Sapir-Whorf and concern about language death: that all languages are equivalent and express the same things, and any idea that can be expressed in one can be expressed by all, so language does not affect thought, and when a language dies it is no loss to the world. I think I must be making some extremely basic mistake in understanding what people are saying, and I feel really dumb about it. It might be too late for this thread, but if anyone felt like setting me straight, I'd love to hear it.
J. Goard said,
January 9, 2010 @ 5:02 am
I'm far from a long-term commenter, but I have lurked for years. Here's my 20 won.
Not necessarily in every case, because there may be other communicative channels (demonstration, iconic pantomime, video) by which a given aspect of meaning might may be easily conveyed. However, the very fact that one community regularly expresses some conceptualization linguistically, while the other could resort to other channels (but usually doesn't care to), is itself a crucial cognitive difference.
We all have many ideas that are immensely difficult to communicate through any language, like a person's face or how to ride a bike. We should not confuse "think" with "think in words". Moreover, every language can be augmented by creative use of words and constructions. Again, though, thinking something because you've done so 100 times a day since you were three, and thinking that thing for the first time by excercising creativity, clearly do not amount to having "buil[t], in fact, the same contents, and equivalent conceptions of the world."
Yes, but things come and things go in this world, and every such language will change dramatically anyway. (In a case where modernization is likely to produce language loss, I would certainly never favor keeping a group of people dirt poor.) The question of how much studying X more languages from X families is likely to bear on deep linguistic questions is something I don't feel qualified to answer. But obviously something is lost to linguistics, anthropology, etc.
Jim Regan said,
January 9, 2010 @ 10:19 am
Fernando Colina: mea culpa, given the topic at hand I should have been more explicit about what exact input I gave, which was:
tu hija que te quiso tanto, tanto y no supo demostrarlo – PERDÓNAME.
which yields:
you wanted your daughter so much and failed to prove it – FORGIVE.
As an aside:
tu hija que te quiso tanto, tanto y no supo demostrarlo – PERDÓNAME
that is, without the period, gives:
you wanted your daughter so much and failed to prove it – FORGIVE ME
But, if you understand the technology, that's quite uninteresting.
If I give it the input:
Tu hija que te quiso tanto, tanto y no supo demostrarlo – perdóname
I get:
Your daughter who loved you so much and she could not prove it – sorry
which is closer to your output, which leads me to believe that your input differed in other ways.
Östen Dahl said,
January 9, 2010 @ 4:54 pm
Here is the link to Lingvistbloggen that Felix Ahlner couldn't find. It's mainly about Google's inscrutable translations of Swedish place names, e.g. Örnsköldsvik>Visakhapatnam; Södertälje>Port Macquarie. Interestingly, at the time when I wrote this post, the Swedish West Coast town 'Varberg' came out as 'Sunshine Coast', when I tried now I got 'Dorset', for whatever reason.
.
Leonardo Boiko said,
January 10, 2010 @ 9:20 am
TB: that was my question too. No easy answers, eh?
Alf Mikula said,
January 10, 2010 @ 1:16 pm
I've noticed the issue with Swedish in the past…I was doing some work with Swedish translations and I know no Swedish at all, so I was using Google Translate to figure out what was going on. I noticed that in some cases if the English word "Sweden" was translated from Swedish to English, it was translating "Sweden" to "Canada". Today that no longer happens, but still today if you translate the following nonsense phrase from Swedish to English:
Sverige, "Svenska" till "Stockholm"
…"Stockholm" gets translated to "San Francisco":
Sweden, "Swedish" to "San Francisco"