The Perils of Machine Translation.

Arthur Goldhammer, “a writer, translator, scholar and blogger on French politics” who “has translated more than 120 books from the French,” writes about translation for Aeon. He begins with an anecdote about “a voluble young Dutchman” who asks a couple of nuns where they’re from; “Alas, Framingham, Massachusetts was not on his itinerary, but, he noted, he had ‘shitloads of time and would be visiting shitloads of other places’.”

The jovial young Dutchman had apparently gathered that ‘shitloads’ was a colourful synonym for the bland ‘lots’. He had mastered the syntax of English and a rather extensive vocabulary but lacked experience of the appropriateness of words to social contexts.

This memory sprang to mind with the recent news that the Google Translate engine would move from a phrase-based system to a neural network.

Go to the link for his thoughts about Google Translate and pattern matching; I want to quote this passage:

Google’s translation engine is ‘trained’ on corpora ranging from news sources to Wikipedia. The bare description of each corpus is the only indication of the context from which it arises. From such scanty information it would be difficult to infer the appropriateness or inappropriateness of a word such as ‘shitloads’. If translating into French, the machine might predict a good match to beaucoup or plusieurs. This would render the meaning of the utterance but not the comedy, which depends on the socially marked ‘shitloads’ in contrast to the neutral plusieurs. No matter how sophisticated the algorithm, it must rely on the information provided, and clues as to context, in particular social context, are devilishly hard to convey in code.

Take the French petite phrase. Phrase can mean ‘sentence’ or ‘phrase’ in English. When Marcel Proust uses it in a musical context in his novel À la recherche du temps perdu (1913-27), in the line ‘la petite phrase de Vinteuil’, it has to be ‘phrase’, because ‘sentence’ makes no sense. Google Translate (the old phrase-based system; the new neutral network is as yet available only for Mandarin Chinese) does remarkably well with this. If you put in petite phrase alone, it gives ‘short sentence’. If you put in la petite phrase de Vinteuil (Vinteuil being the name of a character who happens to be a composer), it gives ‘Vinteuil’s little phrase’, echoing published Proust translations. The rarity of the name ‘Vinteuil’ provides the necessary context, which the statistical algorithm picks up. But if you put in la petite phrase de Sarkozy, it spits out ‘little phrase Sarkozy’ instead of the correct ‘Sarkozy’s zinger’ – because in the political context indicated by the name of the former president, une petite phrase is a barbed remark aimed at a political rival – a zinger rather than a musical phrase. But the name Sarkozy appears in such a variety of sentences that the statistical engine fails to register it properly – and then compounds the error with an unfortunate solecism.

Also, as I wrote to Paul, who sent me the link: “125 books!! When did he have time to eat? I presume he never slept.” (Past tense because he’s apparently given up translation and now writes code instead.)

Comments

  1. “125 books!! When did he have time to eat?”

    Or rather, “When did he have time to text?” I suppose that over time, many avid texters type in the word equivalent of many novels, without noticing it.

  2. The article is presentable enough, but there’s a bit of petticoat peeping out.

    Goldhammer writes that he is “full of admiration for the technical complexity and virtuosity of Google’s work”. Many people either spread their legs for technical complexity, or shut them tight. Christian scientists (small “s”), for example, are likely to admire the watchmaker complexity of the world. I myself find mere complexity of a mathematical kind to be off-putting, along with those who wallow in it.

    It’s a matter of taste, I suppose. My goal is to fight complexity in every form by reworking it into simpler terms and techniques. Keeping my legs in any one fixed position would hold me back. My working assumption is that there must be ways to understand and formulate whatever it is with fewer frills and fritillation. I think pretty much everybody proceeds along those lines.

    There are lots of different kinds of code – the kind I write and review doesn’t require any math knowledge, but only an ability to think straight and organize those thoughts in a particular fashion. That’s pretty hard for some, it turns out. Goldhammer says he has a PhD in math and writes code – algorithm code, I suppose. Well, so what ? There are people who can read three languages, and work in comparative literature. Others raise six kids, and work at raising them.

    Getting back to the second item: virtuosity is an outsized knack. Trump has a knack for throwing well-spoken people into confusion by mere rambling. I am torn between admiring this technique and scoffing at the complex-headed people who can’t rethink in order to handle it. I cite the transcript of the meeting between him and the NY Times people.

  3. 125 books!!

    Not to mention that one of them was Picketty’s weighty tome Capital in the 21st Century.

  4. It was weighed by many, and found vaunting.

  5. Trond Engen says:

    How does Google Translate avoid a feedback loop when it builds “corpora ranging from news sources to Wikipedia”? Does it (the new or the old algorithm) have a way to recognize its own output?

  6. That question about GT has been raised here in the past. As far as I remember, nobody could point to an answer.

    Goldhammer links to an article he says explains the “technical differences” between old and new Google technology: Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation. I merely skimmed it, looking for some indication that self-recognition was addressed in the models described by the authors.

    All I found was the usual tendency to treat all inputs as unequal, but some outputs as less unequal than others. If I understand the authors’ conclusions, they are: dinner in, garbage out, but recycling is still possible.

    The authors don’t explicitly address the question of whether they are garbling corpora over time. Such an outcome would not necessarily be a bad thing. It would make it easier to identify human authors that make sense.

  7. I have trouble often enough understanding the English generated by non-native speakers in IT. I couldn’t say how much of that output is the result of having sucked in too much machine-translated English.

  8. 125 books!! When did he have time to eat?
    Let’s assume that Piketty with its 700 pages is the upper end of the scale and let’s assume that the lower end is closer to 120 pages, that will give us an average of 410 pages and that times 125 equals 51250 pages total. Assuming the usual average number of words per page as 250, that translates to 12 812 500 words total. Assuming some 10% leverage through the use of translation memories and intra-text repetition, we’ll end up with 11 531 250 translated words total. And finally, assuming the usual throughput of 2000 words per standard 8-hour work day and working 365 days a year, we’ll end up with 5765 days total of working on translation which amounts to 15.8 years. Now bear in mind this is a conserv… lib… I’m probably overestimating stuff, like the average book length or the number of words per page (some of it could have been children’s books like the one I’m holding right now which only has 150 words per page), but it sounds about right. I remember when I stopped working as a full-time translator in 2007, I reviewed all my projects between 2004 and 2007 and ended up with some 2 million words translated during that period (this includes translation memory matches). So plus or minus yellow school bus, he could have pulled that off starting in his early twenties and ending in his late fourties, even with (… what do you call that thing when you don’t work for a few days … oh yeah) vacation. And since Goldhammer was born in 1946, this checks out even though looking at the list, I don’t see any children’s books.

    Past tense because he’s apparently given up translation and now writes code instead.
    Good for him.

    I couldn’t say how much of that output is the result of having sucked in too much machine-translated English.
    Occam’s razor and Churchill’s rule lead me to believe it’s none. Rather, it’s simply bad English combined with years of reading bad English. There’s enough bad written English to go around.

  9. Stu,

    I myself find mere complexity of a mathematical kind to be off-putting, along with those who wallow in it. It’s a matter of taste, I suppose.
    I sense some deeper issues, do you want to talk about it?
    And while I see what you’re saying and agree with it to some extent, I also know what Goldhammer means, having spent a few days playing around with TensorFlow (Google’s Neural Network framework-thingy) and SyntaxNet and I know it’s not mere complexity of a mathematical kind he’s talking about. I don’t know what your programming job entails, but trust me, the stuff Google’s gals and guys are up to is pretty far removed from your usual database + web interface or fuel control software (both of which I’ve done).

  10. Bulbul, the deeper issues are in the shallower fields: the math of economics and sociology. Or this, which would be nice if it were a joke: A Cohomological Viewpoint on Elementary School Arithmetic

    Notice that I said “mere complexity” and “wallow”. My beef is with the beef-eaters, not with the beef. In particular, I see no good reason to “admire” complexity or virtuosity, any more than to scorn simplicity and best effort. I just deal with them.

  11. Stu,

    I see no good reason to “admire” complexity or virtuosity
    I understand that, you appear to feel about it the same way I feel about elegance of code or algorithm, an empty and meaningless attribute.
    Except the actual quote is “I am aware of the remarkable feats of which machines are capable, and full of admiration for the technical complexity and virtuosity of Google’s work.” And work is something that can be admired and in case of Google’s work on MT and AI, it should be.

  12. Google Translate doesn’t seem to be improving and may be getting worse. I haven’t been using Chinese much recently but I don’t notice any massive pick up. I also note that my suggestions for improving the translation of individual words seem to have been ignored.

  13. Bulbul: oh well, I guess I’m just not an admirer of admiration. It never bought me a meal. On the other hand, mutual admiration is a free lunch, if you don’t mind the company.

  14. The money thrown at machine translation would be more productive if, instead, it were invested in training and hiring human translators and their subs. Somebody should put the idea to Trump.

  15. I skimmed through the paper that Mr. Clayton referenced and don’t see any problem with it. Nobody’s suggesting to teach kids cohomology before they know how to carry digits forward, but it is a reasonable enough illustrative example of a concept that might be useful for people who are learning cohomology.

  16. Mr. Clayton, I once heard with my own ears a politician suggesting that joggers should be given shovels and made to dig tranches to channel their otherwise useless efforts. And a few decades before that there was even a better suggestion to hook up an electric generator to a ballerina. I hope Mr. Trump will give you an office next to Mr. Bannon’s.

  17. That is exactly why I found it interesting, not being a cohomology buff ! A better title would have been “Basic cohomology ideas explained in terms of school arithmetic”.

  18. I could never propose such an idea, due to conflict of interest – having worked a few years as a translator, many decades ago.

    I doubt that Bannon could retain his sanity (what there is of it) with me in the next office.

  19. Bathrobe says: “Google Translate doesn’t seem to be improving and may be getting worse”. I agree about the lack of improvement. I judge GT by the results. Oft-repeated, starry-eyed promises based on current fancy research don’t float my boat any more.

    From time to time I have tried to use GT in my internet ramblings. I have fed into it snippets of Danish,Thai and Russian, and contemplated the results in German, English and French. The results almost always made hardly any sense at all.

  20. I find GT very useful for producing an English version of a Danish text, or vice versa; much of the rote work is done for me, and in some cases it does pick the right term for the context, or allow me to chose a better one directly in the output window. But of course this only works because I know both languages will enough to weed out all the errors when I then transfer it to an editor and revise it.

    However — I imagine that if I did translations for say an hour per day I’d quickly get to the point where GT was a hindrance more than a help, and typing the first draft myself would be faster.

  21. This machine-translation business has a number of suspicious formal similarities with economics. On the one hand, there are big theories knocking around on which the theorists themselves disagree. On the other, you have non-theorists exchanging anecdotes and disagreeing about what they mean.

    Nothing changes apart from the number of theories and anecdotes to choose from. It may be that that is the important function (not purpose !) these subjects serve – everybody can get in on the act and interact. It all helps you to meet new friends and even your future spouse.

    Meanwhile, the Ding an Sich has a good chuckle.

  22. Bathrobe says: “Google Translate doesn’t seem to be improving and may be getting worse”. I agree about the lack of improvement. I judge GT by the results.
    And you are basing this on careful research, right?

  23. Lars,

    However — I imagine that if I did translations for say an hour per day I’d quickly get to the point where GT was a hindrance more than a help, and typing the first draft myself would be faster.
    No. See for example the experiment / study done at Autodesk where they found that the use of machine translation resulted in significant improvement in translator productivity. In my experiments conducted at my (rejoice, oh multitudes) former job, we observed improvements in terms of translator throughput of 10-25%, depending on language.

  24. The paper in question has “Group cohomology is a vast field of mathematics, and only the most elementary parts of the subject are presented here” on the first page. This is a trope in mathematical works, and a very useful one to me, as it provides fair warning that most of the rest of the paper will be unintelligible. I spoke of this trope to a mathematician friend, who said that a book called Basic Set Theory was the most difficult work on set theory that he had ever attempted. Between the two of us, we came up with the title Prolegomena Toward an Introduction to the Elementary Theory of X-functions, which would be so difficult that no one could understand so much as the first sentence, not even the author.

  25. Nothing changes apart from the number of theories and anecdotes to choose from.
    Again, no. If you bothered to actually read the paper, you would see there has been a whole lot of improvement, especially since Google started using neural networks instead of their simple phrase model, the concept of word-piece is also new and seems to have helped a lot.
    Now it is true that the promise of machine translation as panacea has not materialized, nor has the fear of 2007-2009 that MT will replace human translators. But guess what? Those promises were never made by people who know shit about shit, they have always been PR bullshit combined with the self-delusion and hubris exhibited even by people who do know shit about shit. Which is why you should always read the paper. It may talk about bridgeing automated and human translation, but it qualifies that sentiment all over the place, by pointing out that, for example:
    – “Our best model, WPM-32K, achieves a BLEU score of 38.95”. Translation: our machine translation is about 40% as good as human one.
    – “WMT En-De is considered a more difficult task than WMT En-Fr as it has much less training data, and German, as a more morphologically rich language, needs a huge vocabulary for word models”. Translation: this only works for languages with relatively simple morphology and big bilingual corpora.
    So no, even Google MT is not perfect, not even close. It may not even be good at times. But there is improvement and it is usable. And that’s all you were ever promised by people who work on it.
    Also, let’s not forget that most Google’s products are permanent beta and so being constantly tinkered with, so what works one day, may not work the next. But hey, you get what you paid for.

  26. And you are basing this on careful research, right?

    No. I said I judge GT by the results – what comes out when I put a nickle in the public vending machines. To judge these results no research at all is needed. GT is offered for ad hoc use in the internet, and the results are ad hoc useless to me in the internet.

    As you say, I get what I paid for, namely nothing. It might be better PR to take GT out of the public eye until the promises can be kept.

    Of course GT is useful in other contexts – to the people who wrote that paper, say. I’m not knocking it for them !

    I still feel that my comparison with economic theory and practice is accurate. A particular example of that was real existierender Sozialismus in the DDR. There you had dogmatically reliable caders unable to buy a banana.

  27. No. I said I judge GT by the results – what comes out when I put a nickle in the public vending machines.
    Except you don’t put any nickel into GT and a MT system is infinitely more complex than a public vending machine. So the comparison is useless.
    But my larger point is this: you are basing your evaluation on a tiny and biased sample which makes your evaluation completely worthless. I do hope you realize that.

    It might be better PR to take GT out of the public eye until the promises can be kept.
    Like I said, Google is not making promises. All they’re saying is “here’s a toy we’ve been playing with, we think it’s pretty cool, feel free to use it.”
    There are companies who make money producing machine translation who do make promises regarding MT, but we are not talking about them.

  28. you are basing your evaluation on a tiny and biased sample which makes your evaluation completely worthless.

    Worthless to you, but not to me ! I clearly belong to the anecdote-trading group. You seem to place yourself in the theorists’ camp. Just as I described it – a standoff. Each side accusing the other of being worthless. The pot calling the kettle Mr. President.

  29. Just as I described it – a standoff. Each side accusing the other of being worthless.
    Um, no. I described your opinion of GT as worthless because it is, as it is based on biased and skewed data sample. It’s like describing all German cars as bad because of that one time when you rented one it broke down on you. We’re talking about cognitive biases and logical fallacies.
    Also, there are no actual two sides here, just some people with data and some people who go “Me no likey!”

  30. That’s another way to put it ! We seem to agree after all.

  31. If we are trading personal experiences, I find GT useful not for my work (which doesn’t include translation), but for leisure activities. For example, my capacity in French is almost non-existent and German is even worse. If I need to make sense of something written in French or German, GT saves me hours of time of looking up words and grammatical constructions, including most trivial and mundane words from the first 2000 which any semi-demi-intermediate speaker knows. Or rather, if forced to do it all by myself, I would simply pass. With GT I can then read for meaning and occasionally look up a few words or expressions where translation looks suspect or doesn’t make sense. In other words, it is much better than a vending machine, which eats my money with some regularity without any output and the rest of the time provides me with junk food (or drink) that in all probability I would be better off not ingesting.

  32. In my job I sometimes need to translate documents I wrote in German into English or vice versa. Here I use GT for getting quick and dirty translations which I then correct. I also often need to look for information in documents written in languages which I don’t know or know badly; here GT is helpful in understanding the gist of a document and sorting documents according to whether they contain what I need (and hence I’ll have them translated by a human) or not (so I can ignore them).

  33. Stu Clayton says:

    I buy German, Spanish, French and English products fresh from the farm. My leisure recipes sometimes call for exotic bits of Danish and Thai, say, but GT is almost never any help there. What comes out in these cases is not a bad translation – I couldn’t judge of that anyway, which is why I reach for GT in the first place – but just plain gobbledygook. German is always my GT target language, and it shows the limitations of GT clearly.

    For amusement I have occasionally tried GT from English to German, and usually get a semi-mess that would have to be considerably reworded and rearranged to become understandable. It might save a bit of time over working from scratch when you know both languages and are just casting about for more accurate words, as Hans implies. But I don’t work like that myself.

    I don’t know what use a semi-mess might be for people don’t know both languages already. What is for me a semi-mess might be a complete mystery to them.

  34. Aus meinen großen Schmerzen
    Mach ich die kleinen Lieder;
    Die heben ihr klingend Gefieder
    Und flattern nach ihrem Herzen.

    Sie fanden den Weg zur Trauten,
    Doch kommen sie wieder und klagen,
    Und klagen, und wollen nicht sagen,
    Was sie im Herzen schauten.

    Translates to

    From my great pain
    I make the little songs;
    They lift their sound feathers
    And flutter to her heart.
     
    They found the way to the Trauten,
    But they come again and lament,
    And lament, and will not say,
    What they looked in their hearts.

    It’s not Heine, but you can work from there.

  35. Stu Clayton says:

    For Chrissake ! I can just see the burrowed frowse of someone reading that who has no German.

    Who is this Trauten filly ? Why did the songs go all that way to look in their own hearts ? Why did they flutter if their feathers are hale and sound ?

  36. Stu Clayton says:

    I myself wouldn’t judge GT on the basis of pomes. A lot of people even find them too hot to handle.

  37. David Marjanović says:

    Trump has a knack for throwing well-spoken people into confusion by mere rambling.

    It’s called a Gish gallop, as explained in this rationalwiki article… oh, I see it’s been updated. 🙂

    And yes, there is enough badly written English around for everyone’s needs, and then some. I get to review it sometimes…

  38. And I get to edit it!

  39. @bulbul the use of machine translation resulted in significant improvement in translator productivity. In my experiments conducted at my (rejoice, oh multitudes) former job, we observed improvements in terms of translator throughput of 10-25%, depending on language.

    That has been my experience translating English into Chinese (which I no longer do). Starting from scratch is hard work; it’s easier to cut and paste from a halfway job. Sometimes you don’t have to do much at all; at other times it is hard slog and there is a fair amount of virtually unusable detritus. But if you don’t know both your languages reasonably well you would be hard put producing a reliable translation.

    As I’ve mentioned at other times (many years ago), one of my greatest peeves was that Google Translate continued to mangle numbers and figures for year after year. There was a certain consistency at times, like generally translating 50 million as 5 million (IIRR), but when you had multiple figures distributed among different parameters it was pure carnage.

    Just for fun, I’ve tried an old document on the new Google Translate. While it seems better than before (at least everything is in its place, which is a quantum leap), it’s still pretty unreliable. For instance:

    First paragraph:

    “303.1 mln t of cargo”

    translated as

    货物303.1百万吨.

    Correct, although it needs to be converted to “3.031亿吨” to match the Chinese number system.

    But by the fourth paragraph, our neural learning system seems to have forgotten how to do what it did just a few paragraphs ago:

    “17.1 mln t of cargo including 12.3 mln t of dry cargo and 4.8 mln t of liquid bulk cargo.”

    translated as:

    17.1万吨货物,包括12.3万吨干货和4.8亿吨液体散货。

    That is: “171,000 tonnes of cargo, including 123,000 tonnes of dry cargo and 480 million tonnes of liquid bulk cargo”.

    A translation, no matter how beautiful, that gets figures wrong is almost useless, and will be so judged by the reader.

    I’ve tried giving input to Google Translate in the past. One I remember recently is the term for “Ambassador’s residence” within an embassy. This is a fixed term in many languages but Google Translate gave me this for “Ambassador’s residence”: Chinese 大使的住所 literally ‘Ambassador’s residence’, Japanese 大使の住居 literally ‘Ambassador’s residence’, Mongolian Элчин сайдын оршин суух literally ‘residence of the Ambassador’. And this for “Ambassador’s Residence”: Chinese 大使馆 literally ’embassy’, Japanese 大使の住居 literally ‘Ambassador’s residence’, Mongolian Элчин сайдын оршин суух literally ‘residence of the Ambassador’. These are all literal translations and not the correct term. (Although the neural learning circuits at Chinese have got the capitalised version completely wrong, unlike the Japanese and Mongolian phrase-based translations.)

    After quite a bit of hunting around I found the correct Chinese term, 大使官邸, and suggested it to Google Translate. It’s now there for both “Ambassador’s residence” and “Ambassasor’s Residence” as an alternative to their first offering. The correct Japanese term, 大使公邸, which I similarly suggested to Google Translate, is also there as an alternative to their first offering. The correct Mongolian term, Элчин сайдын өргөө, doesn’t appear at all.

    It’s a trivial example but it shows the difficulty for Google Translate to get it right.

  40. I’ve just noticed that “ambassador’s official residence” yields the correct term for both Chinese and Japanese (although not Mongolian). I guess a lot depends on what you input.

  41. I also notice that Google’s neurons still can’t manage to come up with a translation for ‘vapid’ in Chinese. Although I guess that’s an improvement on the translation that they give for Japanese, which is 激しい (normally meaning ‘fierce’, ‘intense’, ‘violent’, ‘furious’, ‘vehement’). Maybe I had it in antonym mode…

    There seems to be something broken at Google Translate. After all these years, it should surely be able to do better than that.

  42. “The money thrown at machine translation would be more productive if, instead, it were invested in training and hiring human translators and their subs.”

    Why? So much of a translator’s day-by-day work is simply mentally draining and soul-sucking. It is in no way ennobling and the only reason real humans do it is because they otherwise couldn’t eat. Let translation of sales catalogues, boilerplate contracts, and vacuous press releases all be taken over by machines. I’d be a lot happier in a futuristic heavily automated world where I could live off basic income and have more time for what I really want to do with languages, namely linguistics research and translating literature.

  43. Of course, that requires there to be basic income to live off of.

  44. So much of a translator’s day-by-day work is simply mentally draining and soul-sucking.
    This. And that’s not counting all the mentally draining and soul-sucking translation project management work.

    Let translation of sales catalogues, boilerplate contracts, and vacuous press releases all be taken over by machines
    Amen, I say, amen.

  45. Bathrobe,

    I also notice that Google’s neurons still can’t manage to come up with a translation for ‘vapid’ in Chinese.
    Have you actually read the article? If you have, you may be able to guess why. Hint: GT does not translate words. Not even human translators do.

  46. GT does not translate words. Not even human translators do.

    Oh but it does. GT just told me that “disingenuous” is “unaufrichtig” in German.

    It also told me that “vapid” is “schwach”. This translation itself is schwach, which is pretty cute but doesn’t help. “Banal” is more accurate. Thanx to Pons for that one.

  47. I do so love a round of trash-talk ! That’s all we’re doing here. We might as well be dissing each other over Keynesian economics. Nobody is the wiser for it, but maybe Mr. and Mrs. Right will find each other beneath the salvos.

    Also, everybody needs a way to earn money. Google keeps the economy rolling by hiring people to work on GT and other things. I see nothing wrong with that. Unfortunately, performance reviews come in all shapes and sizes. Google has not yet bought up that area of activity.

  48. Oh but it does. GT just told me that “disingenuous” is “unaufrichtig” in German.
    No, it does not translate words. It may appear to do so, but it does not. It is not a dictionary and besides, there is no such thing as “translating works”.

    I do so love a round of trash-talk ! That’s all we’re doing here.
    Again, no, we’re not. Some of us are making reasoned arguments. Others are basically going “durpa durpa durpa duuuuur.”

  49. “Others are basically going ‘durpa durpa durpa duuuuur’ ” is not a reasoned argument. You vindicate my claim that this is trash-talk.

    El sueño de la razón produce razones.

  50. Hint: GT does not translate words. Not even human translators do.

    It does actually translate words, and always has. Indeed, it has a small dictionary at bottom left that gives you definitions, synonyms, and examples of the word ‘vapid’ as used in sentences, and at bottom right a list of equivalents to the word ‘vapid’ in the foreign language and their meanings in English (in the case of Japanese only one, 陳腐 stale, old, moldy, vapid, flat, putrid). That’s an awful lot of stuff for an automatic translation service that only translates sentences.

  51. Above, bulbul tried to beg off by distinguishing between real and apparent GT. Such detritus from the history of philosophy is not helpful. It seems he doesn’t like GT out there in the internet appearing as pragmatic, instead of all-potent in a special restricted area (“sentences”).

  52. A full translation of the sentence “Somewhere through the course of their vapid conversation, she caught my eye and smirked knowingly” also uses 激しい hageshii ‘fierce’, so the observation that Google Translate doesn’t translate words is also beside the point.

    彼らの激しい会話のどこかで、彼女は私の目をつかんで、故意に笑った。

    And this is how Google Translate puts the above Japanese sentence back into English:

    “Somewhere in their intense conversation, she grabbed my eyes and deliberately laughed.”

    This is actually what the Japanese means. The point is that the translation is wrong, and it’s wrong because Google Translate gets one word exactly wrong (plus a few dodgy renditions as well) — although I suspect sabotage. Some joker may have suggested the wrong equivalent just for fun.

  53. That’s odd: “smirked knowingly” has semantic components similar to those of “deliberately laughed”. A smirk is a subdued laugh, knowing and deliberation are cognitive thingies.

    I sense here an algorithm confronted with two roads diverging in a yellow wood.

  54. I think it’s a bit less subtle than that. “Knowingly” can have (at least) two meanings:

    “She looked at me knowingly” is one.

    “He knowingly transgressed the law” is another.

    笑う in Japanese can also mean either ‘smile’ or ‘laugh’.

    GT picked the wrong equivalents, that’s all.

  55. Ah, so that’s why you suspect a joker may have rigged it up. A rigger in the woodpile of words, as it were.

  56. “Others are basically going ‘durpa durpa durpa duuuuur’ ” is not a reasoned argument.
    True, it’s not, it is a description of the lack of the existence of such.

    It does actually translate words, and always has.
    It uses words – or rather tokens – but it does not translate them; for the translation process, it works with groups of words (see also the paper where they spell this out (“Practical implementations of SMT are generally phrase-based systems (PBMT) which translate sequences of words or phrases where the lengths may differ”).

    it has a small dictionary at bottom left that gives you definitions, synonyms, and examples of the word ‘vapid’ as used in sentences,
    This is not a dictionary, but a (bilingual) concordance.

    Above, bulbul tried to beg off by distinguishing between real and apparent GT. Such detritus from the history of philosophy is not helpful.
    No, I didn’t. The distinction I made is between what you see and what is actually happening. Think physics, e.g. sunrise and sunset.

  57. Actually, 笑った waratta is not wrong. It’s the translation back into English that forces a choice, and the choice happens to diverge from the original sentence as translated into Japanese. The cumulation of word choices (wrong or subtly wrong) is what gives rise to the rather large divergence in meanings.

  58. The distinction I made is between what you see and what is actually happening. Think physics, e.g. sunrise and sunset

    “Sunrise” is what I see, “sunset” is what is actually happening ? “What is actually happening” does not imply “real” ? That’s not yet ‘durpa durpa durpa duuuuur’, but it’s close.

  59. This is not a dictionary, but a (bilingual) concordance.

    It is still a concordance of words. Extremely useful for the translator given that words are at the bedrock level of translation, but words none the less.

  60. The cumulation of word choices (wrong or subtly wrong) is what gives rise to the rather large divergence in meanings

    I tried several A->B->A conversions this morning, for languages A and B. The original input was never the same as the final output. I wonder whether it could be helpful to let the A->B conversion be guided by whether A->B->A restores the original input. I have doubts about that, but can’t yet figure out exactly why.

  61. A long time ago LH had an entry about a fun site on the Internet that had GT translate the same passage back and forth between languages until a state of equilibrium was reached. The final version was mostly wildly different from the original. I don’t know whether that site is still around, and whether it would yield the same results today.

  62. Not guided only by such fixed-point considerations, of course, especially if A->B and B->A rules are inverses of each other. In that case any old set of A->B rules would give you a fixed-point.

  63. until a state of equilibrium was reached

    I remember that now ! It gave equilibrium a bad reputation.

  64. I did notice that while GT translates “vapid” into Chinese as “vapid”, and “vapid conversation” as “vapid交谈”, it translates the sentence “Somewhere through the course of their vapid conversation, she caught my eye and smirked knowingly” as 在某个地方通过他们的虚假对话,她抓住了我的眼睛,并故意嘲笑. “Vapid” is 虚假. So it does appear that the new neural networks have (as bulbul pointed out) decisively moved away from the word-based level of translation at Chinese. Japanese translation still uses the old phrase-based system.

  65. Trond Engen says:

    Bathrobe: I’ve just noticed that “ambassador’s official residence” yields the correct term for both Chinese and Japanese (although not Mongolian). I guess a lot depends on what you input.

    Chinese: 大使官邸
    Japanese: 大使公邸

    Playing with GT letter by letter and looking through the alternatives I get something like “Major envoy official mansion”*. That suggests to me that GT doesn’t really know the terms. It just happens upon them when you give input that corresponds to the output you want.

    *) In Chinese, that is. The Japanese obviously means the same, even if only “big” and “use” are suggested for the first two, and even with a different word (or letter) for “official”.

  66. “Sunrise” is what I see, “sunset” is what is actually happening ? “What is actually happening” does not imply “real” ?

    That is the trope of ‘The sun isn’t rising, it’s really the Earth turning towards it’. But there really is no ‘really’ about it, if you want to be annoying.

    It is true that Galilean frames centered on the Sun and with stationary directions to the far galaxies work better for orbital mechanics than the ones where an given observer on the earth is stationary — but physics doesn’t have an opinion on which is ‘real’. They are both approximations.

    And your local Galilean frame where the stars are rotating is perfectly fine for all the physics you do on a daily basis; even meteorologers use such a frame, and add the ‘Coriolis force’ where needed. Of course it doesn’t explain why the Sun goes around the Earth instead of flying off somewhere, but that’s why we don’t try to use it for that.

    (Also we operate with a ‘gravitational force’ pulling us towards the center of the Earth instead of doing our calculations with a non-diagonal local curvature tensor. Because the results are the same to the precision we need, except when running a GPS satellite system).

  67. But there really is no ‘really’ about it, if you want to be annoying.

    That’s annoying only to people who cling to realism like a cat to a storm-tossed tree. Of course you also can’t really insist that there is no ‘really’ about it. That would be self-contradictory.

    Frames of reference (not just those of physics) are ways of looking at things that can be entertained from time to time, alternating with realism. They contradict realism in the sense that A and not-A can’t occupy the same mental location at the same time, to put it crudely. But this difficulty can be “resolved in the time dimension”, as Luhmann would say, by first doing one, then the other, then perhaps going out for a beer. Bounded oscillation, you might call it. Or cognitive flexibility.

    I was startled to learn recently that having the speed of light as an upper bound to the speed of anything really changes the ballpark. Newtonian physics has no such general speed limit. The multi-body problem there has weird solutions. In 1992 a guy named Xia showed mathematically that you can set up five point particles moving in such a way that one of them “shoots off to infinity in a finite amount of time”. His solution involves unbounded, accelerated oscillation.

    This is the article where I read about it.

  68. I suspect that “no fixed point” approach is here because of the probabilistic nature of translation. It very well may be that the most probable meaning of expression A in source language corresponds to expression B in the target language, but the most probable meaning of expression B is something different. Quick example: divan is a somewhat unusual word for a certain piece of furniture in English, but диван is the first thing you will call the same piece of furniture in Russian. It would be pretty pretentious to translate диван as English divan ahead of sofa and couch. And I don’t think there is enough possible context a translator (either human or machine) can have to reach that point.

  69. David Marjanović says:

    And I get to edit it!

    Me too, pretty often. I mean, I’m not expected or paid to do it, but I tend to do it anyway because I know nobody else will do it.

    Thanx to Pons for that one.

    Pons, I should perhaps explain, is a dead-tree dictionary publisher that tried to establish the verb ponsen in analogy to “to google”. That didn’t work.

  70. I’m glad I missed that campaign. All the big dictionaries are now on-line 4 free, btw. No trees were harmed.

  71. Except the OED, alas.

  72. Amidst the shouting match between Stu and bulbul and the excursion into Ambassadors, my point that GT still can’t get numbers right seems to have been lost. Why do numbers denominated in millions in the same sentence lurch from the hundreds of thousands to the hundreds of millions?

    As for Ambassador’s residence, I tried inputting the Wikipedia article for Chancery into GT.

    A chancery is the type of building that houses a diplomatic mission or an embassy. The building can house one or several different nations’ missions. The term derives from chancery or chancellery, the office of a Chancellor. Some nations title the head of foreign affairs a Chancellor, and chancery eventually became a common referent to the main building of an embassy. The ambassador’s quarters are generally referred to as the Residence.

    This is the result for Japanese:

    大統領選は、外交使節や大使館がある建物のタイプです。 建物には、1つまたは複数の異なる国のミッションを収容することができます。 この用語は、大統領または大統領、首相の事務所から派生しています。 いくつかの国が首相の外交の頭に就任し、大統領選は最終的に大使館の本館に共通の指示対象となった。 大使の四半期は、一般にレジデンスと呼ばれます。

    Back-translation:

    The presidential election is the type of building with a diplomatic mission and an embassy. The building can accommodate missions of one or more different countries. This term is derived from the President or the President, Prime Minister ‘s office. Some countries took over as Prime Minister ‘s diplomacy head, and the presidential election finally became a common direction for the main building of the embassy. The ambassador’s annual quarter is generally called a residence.

    一个庄园是建立外交使团或大使馆的那种建筑。 该建筑可以容纳一个或几个不同国家的任务。 这个词来源于大法官办公室的大法官或大法官。 一些国家将外交部长称为大臣,而大法官最终成为大使馆主楼的共同指示。 大使的宿舍一般被称为住所。

    Back-translation:

    A manor is the kind of building in which a diplomatic mission or embassy is established. The building can accommodate one or several different national tasks. The word comes from the Lord Chancellor’s office or Lord Chancellor. In some countries, the Minister for Foreign Affairs is called the Minister, and the Chief Justice becomes the co-directive of the main building of the Embassy. Ambassadors’ dormitories are generally called dwellings.

    Is the neural network version really that much of an improvement on the phrase-based version?

  73. Marja Erwin says:

    I think the first problem is that if we’re using it, we may not have the energy to correct all its mistakes at once…

    I suspect that similar problems contribute to the inaccuracy of speech-to-text software, although with the lack of any working interface rather than a lack of energy.

  74. Ambassadors’ dormitories are generally called dwellings.

    This is so good, it should be posted into THE HAZY YON thread.

  75. Surprised that no one has yet brought up “No no, you’re confused, the horizone is moving up”, but I guess us old-timers are getting thin on the ground.

    I’m with D.O. up there. Years ago if I wanted to find out what the German or Italian newspapers were saying about some issue I was out of luck, because I don’t speak those languages. But now with very little effort I can get a half-assed translation which in most cases is sufficient for my needs. I call that an improvement.

    The other cool thing about GT is that you can enter Chinese characters by drawing them in. This is one feature that has definitely got better over the last few years. When it was first introduced it often avoided the correct answer when presenting you with alternatives, but more recently it has become much more useful. This is a time-consuming way of going about things, but for my limited needs it works pretty well.

  76. I call that an improvement.

    Me too. There seems to be a divide between people (like thee and me) who are interested in having an idea of what’s being said in languages we don’t read and are excited about having the means to do so, even if occasionally through a glass darkly, and people who are interested only in perfection, and if GT can’t produce it go “Pfft, sucks.”

  77. @maidhc, I did.

  78. There seems to be a divide between people (like thee and me) who are interested in having an idea of what’s being said in languages we don’t read and are excited about having the means to do so, even if occasionally through a glass darkly, and people who are interested only in perfection, and if GT can’t produce it go “Pfft, sucks.”

    It’s not a divide, but merely a path from one little node to two other littel nodes in a web of skills and interests.

    I myself don’t want to have an obscure idea of what’s being said in a book. What actual people in actual life are saying, is another pot of pisces. There I would even use a 19C phrase book.

    But I also don’t expect perfection, otherwise I would never read comment threads.

    I’m on a thrid node, and there are others. I read and/or speak more than enough languages already. GT is no help there, and I prefer not to dilittate when GT does.

    I gather from the most recent comments that it helps greatly with Chinese, in some respects.

  79. dilettieren => dabble

  80. “No no, you’re confused, the horizone is moving up”, but I guess us old-timers are getting thin on the ground.

    The same goes for Google itself. If you cast your mind back to the end of the 20th century, you will remember what a revolution Google was in web search. Before that, you took it for granted that you probably wouldn’t find what you were looking for. Remember Alta Vista? I myself remember enthusiastically recommending Google at my own website around 2000. But a lot has happened since then, and not all of it is Google’s fault. Google’s very existence has provoked changes in the way it operates. There were the tireless efforts of the web site optimisers and link farms that wrought havoc with its algorithm. There was also the problem that once a site was at the top of Google’s rankings, inertia tended to keep it there at the expense of many newer and better sites. There has also been the explosive growth of the Internet itself. Google has worked hard to fix these problems, but despite everything there are certain ways in which the initial Google was possibly better than what we have now. It may have been easier to find what you wanted on Google than it is now. To point this out is not to be someone who forgets what it was like in the bad old days.

    For Google Translate, of course we are grateful that we can get an idea what is written in a whole swathe of major languages. But Google Translate has its own problems. One is the initial problem that it was always better (significantly so) at some languages than others. There is also the fact that, as Google Translate becomes ubiquitous, the output of Google Translate inevitably runs the risk of becoming input for the Google Translate machine, potentially creating a self-reinforcing loop. And despite bulbul’s protestation that it’s not meant to be used that way, originally it was quite possible to use Google as a dictionary. Now we have a concrete example, one that I pointed out (“vapid”), where Google either fails to come up with a translation equivalent, or comes up with the wrong equivalent. It’s not possible to go back in time to test this, but I’m pretty sure that you could look up “vapid” and get a decent Chinese or Japanese equivalent not so long ago.

    So it’s fine to be grateful, but that does not mean that those pointing out the problems and frustrations — and backsliding — of Google Translate should be regarded as ingrates.

  81. I agree with pretty much everything you say (and I too remember the joy of the first discovery of Google — so much better than Altavista!), but:

    originally it was quite possible to use Google as a dictionary.

    That’s as may be, but it certainly wasn’t intended as one, and surely the plethora of online dictionaries now available makes that a pretty faint regret.

  82. David Marjanović says:

    Not that many Internet ages ago, a recommendation for what to do with unintelligible Internet crackpottery was to transkoreanize it: GT it to Korean and back in the hope that the retranslation will make more sense (and/or be funnier) than the original.

  83. it certainly wasn’t intended as one, and surely the plethora of online dictionaries now available makes that a pretty faint regret

    No, that’s not true. When you are trying to triangulate among languages, or are simply working with several different languages, there is nothing better than having them all together in one place. It is a pain (and sometimes slow or impossible) to open different websites in different countries trying to find a translation equivalent. Sure, for Chinese I can use iciba, for Japanese goo, and for Mongolian Bolor, and there are probably more out there now. Both goo and Bolor feature other languages in addition to English. But if you want “one-stop shopping”, nothing beats Google Translate, supplemented by the other dictionaries if you strike a snag.

    For Mongolian you will sometimes need Russian, which is right there at Google Translate. I once had a colleague who was trying to translate a document from Mongolian into Chinese and was stumped by a strange-looking technical word (this was before GT got Mongolian). I immediately identified it as Russian, found the English equivalent to the Russian, and then found the Chinese equivalent to the English. Voilà, that was the word he wanted. All this thanks to Google Translate, which saved a lot of time.

    Really, it’s a bit silly to say that “Google Translate isn’t meant for looking up words, so don’t do it”.

  84. originally it was quite possible to use Google as a dictionary
    That’s as may be, but it certainly wasn’t intended as one

    “Translate” is not a word Google owns. When people enter single words in GT to get these translated, it’s perfectly reasonable for them to think of GT as a dictionary. By the duck principle, it is a dictionary.

    When people enter sentences in GT to get these translated, it’s perfectly reasonable for them to think of GT as “not [just] a dictionary, but a translator”. Also by the duck principle, it is not [just] a dictionary, but a translator.

    I would guess – but it’s only a guess – that the Google people who decide what interface to present to the public as “Google Translate” are not concerned that GT can be used as a dictionary, by entering single words. Of course it’s also possible that nobody has told them how important it is to stop people from using GT as a dictionary, so that they don’t get any funny ideas about its being a dictionary in addition to being a super-duper translation thingy.

    bulbul insists that GT is only “apparently” a dictionary. He seems to want even Joe Blow users of GT as a dictionary to acknowledge in speech that they are bigly wrong to be satisfied with using it as a dictionary, and calling it that.

    There may be worries among some back-end GT workers as to their honor and reputation. These worries could be dispelled by a simple change to the GT public interface: when someone enters a single word there, a pop-up appears with the message: “Invalid input. You are attempting to use GT as a dictionary. Our code of honor does not allow this.”

  85. I’ve already debunked the claim that GT is designed only for use in translating passages, but in case you missed it:

    A. Input the sentence “I like to play the flute” into Google Translate for translation into Chinese and below the translation box you will get only the pinyin for the Chinese translation: “Wǒ xǐhuān wán chángdí”.

    Input the word “flute” and you will get what bulbul calls a “(bilingual) concordance”: 1. definitions of “flute”, synonyms, examples, and see also, and 2. translations of the word “flute” into Chinese. Why would Google Translate provide a feature that is visible only if you look up individual words? Answer: It is obvious that Google Translate intends for people to look up individual words.

    B. Translations of sentences clearly reflect the properties of individual words. An example is the sentence about “vapid conversations”, where “vapid” comes out wrong in Japanese whether you look up the individual word or translate the entire sentence. The idea that Google Translate is valid only for entire sentences and not for individual words is nonsense. Google Translate (at least in its phrase-based incarnation) operates equally well on the word and sentence levels.

    The only “proof” we have that GT should be used only for the translation of passages is bulbul’s say-so, and that only seems to have been thrown out there in annoyance at Stu’s comments. I don’t think that it can be taken seriously as a statement of fact. Not much different from prescriptivist claims about language, which often have no basis in reality at all.

    (It struck me that bulbul’s response might be: “Yes, but you’re using it as a translation aid, not a dictionary”. To which I can only reply, what difference does it make?)

  86. Yes. Still, modifying the GT public interface might make some GT people *feel* better about the whole business, and that is no mean thing.

    Dragging facts into the discussion hasn’t helped so far. It has been centered around the presentation of self in everyday life, viz. at the GT public interface.

    Bathrobe, the addition to your comment makes me realize that the warning message should avoid the term “dictionary” altogether, even though a dictionary is a translation aid: “Invalid input. You are attempting to use GT as a translation aid. Our code of honor does not allow this.”

  87. In fact when inputting a phrase you can double click to select a word on the left hand side and you’ll get dictionary lookup on both sides — English senses at the left and glossed Chinese translations on the right. Not so much for other input languages, though.

    Also sometimes you can select a part of the translated text and get a list of other possibilities, but for ‘I like to play the flute’ it only gives two whole sentences. I thought that might be a property of the neural network system, but other inputs do allow it — or maybe it falls back to the phrase based system if you are sufficiently weird. (My input was ‘hans støvlehæle gik han skæve’).

  88. When you are trying to triangulate among languages, or are simply working with several different languages, there is nothing better than having them all together in one place.

    An excellent point, and one I hadn’t thought of. But:

    Really, it’s a bit silly to say that “Google Translate isn’t meant for looking up words, so don’t do it”.

    Come, come, nobody’s saying that, that’s silly.

    Dragging facts into the discussion hasn’t helped so far.

    Come, come, I say again. You are in the business of providing attitude, not facts. Which is delightful, of course, but distinguamus.

  89. Really, it’s a bit silly to say that “Google Translate isn’t meant for looking up words, so don’t do it”.

    Come, come, nobody’s saying that, that’s silly.

    That’s exactly what bulbul was implying — that you’re not supposed to be using GT as a dictionary. You tacitly agreed with his underlying argument by saying “it certainly wasn’t intended as one”. Which is why I felt impelled to show that his contention was baseless. GT is designed with people who use it as a dictionary in mind.

  90. That’s exactly what bulbul was implying — that you’re not supposed to be using GT as a dictionary.

    Implying is in the eye of the beholder, and I don’t think he’s actually implying that. At any rate, I’d prefer to hear him say it explicitly before assuming it.

    You tacitly agreed with his underlying argument by saying “it certainly wasn’t intended as one”.

    Again, “tacitly agreed with his underlying argument” is unfair and in this case inaccurate argumentation. I meant exactly what I said, that it was not intended as a dictionary, it was intended as a translation service. I would have thought it was obvious that that in no way precludes people using it as a dictionary, any more than the fact that my glasses case was not intended to be used for weighing down a page precludes my using it in that capacity, one in which I am now employing it and in which it serves perfectly well.

  91. And my argument at 3:29 am is that it was intended as a dictionary.

  92. Well, you know more about it than I do, so I’m not going to argue the point!

  93. I still refuse to believe, even as my inside knowledge slips from eight to nine years out of date, that the GT maintainers are dumb enough to incorporate their own output as one side of a bilingual corpus, and I think the fear that it will degenerate into something that translates itself is unfounded.

  94. Do you have any former comrades in arms you can contact about it?

  95. I know that in (American) retail world consumer experience is everything. But what if we step back for a second from this principle and look at a situation a bit differently. GT is based on some technical principle of inferring the meaning of the words (if it comes to individual words) from bilingual concordances. It definitely provides users (aka consumers) with possibility to use it for that purpose. The question is whether it is the best method to look up meanings of individual words. Maybe not. But it is what GT does for you and if you don’t like it, use another method (like traditional dictionary). It is a bit strange to come to a kosher shop for some pickles and complain that you cannot buy pork as well.

  96. You are in the business of providing attitude, not facts.

    I had never thought of it quite like that, but it’s true. For me, facts in the internet are mere fuel for the fires of ideation. In everyday life I am more cautious, as was Scrooge with his office coals.

Trackbacks

  1. […] Hat notes the insensitivity of machine translation and examines the evolution of the Spanish […]

Speak Your Mind

*