Does Machine Translation Affect International Trade?
nber.org
nber.org
Oh fascinating, do you have any idea of what things machine translation still consistently gets wrong?
As to sibling comment, IIRC Yandex translate has a nice feature where they'll provide a number of meanings for words when clicked upon in the source text.
"Invisible, Insane"
Here's part of the Google translation of the front page of ja.wikipedia.org into English:
> 非行少年の再非行の抑止や更生を目的としており、決定までの過程として、「非行事実を家庭裁判所に送致・通告 - 家庭裁判所調査官（以下「調査官」と略称する）等による調査 - 調査結果をふまえた審判 - 必要に応じて保護的措置あるいは保護処分を決定」という流れを経るのが通例である。
> And for the purpose of deterrence and rehabilitation of re-delinquency of juvenile delinquents, as the process of until a decision, " the fact delinquency to the family court -up Reel-notice - family court investigators (referred to hereinafter as" investigators "), or the like by the survey - survey It is customary to go through the flow of " judgment based on the result- decide protective measures or protective measures as necessary ".
That's barely recognizable as English. The translation is hot garbage. Perhaps something you might use if you had an emergency.
Just to point out here... the text on the front page of Wikipedia tends to be very neutral, matter-of-fact, and well written. There are no characters (like in novels), no use of slang, and nothing else that should be tricky for translation, and yet we get this garbage.
It's a miracle that machine translation works at all, and when it does work, it is usually because you are translating between languages that are fairly similar to begin with. For example, English and Spanish.
Then there are other things besides pronouns that require inference. For example, adding な to the end of a sentence could mean either a modifier expressing admiration or confirmation, or it could be indicate a demand or order not to do something. In machine translation it almost always seems to be translated to "don't" no matter the context. That only makes the translation less accurate.
And I’d assume that could be the most common language people would want to translate from online platforms in this day and age.
Of course the details of how English is turned into incomprehensible garbage by machine translation always depend on the target language in question.
> The purpose of this process is to deter and rehabilitate the juvenile delinquent, and the process of making a decision on the delinquency is usually conducted in the following manner: referral and notification of the fact of the delinquency to the family court; investigation by an investigator of the family court (hereinafter referred to as "investigator"); trial based on the results of the investigation; and, if necessary, a decision on protective measures or protective measures. There is.
Translated with www.DeepL.com/Translator (free version)
DeepL is doing a better job of generating text that seems to make sense, but this just helps to more effectively communicate the wrong meaning, and in a manner less detectable to people who would use it. This is a much more insidious failure mode.
This is something I don't understand about the use of autocorrect. Autocorrect takes misspelled words, which are essentially always easy to understand, and replaces them with random other words. It's frequently impossible to guess what original word got mangled by autocorrect into appearing in a suddenly unintelligible sentence. This is a problem that misspellings just don't have.
Autocorrect is actively making communication worse. What's it supposed to be doing? Why is it there at all?
The first has your stated effect. The second does not.
Also, it has "bugs" that sometimes it will just drop an entire sentence or half-sentence for no reason.
"Its purpose is recidivism prevention and rehabilitation for youth offenders; the usual process leading up to a decision is: Notice or report of the fact of the youth crime to the family court > Investigation by the family court investigating officer (the "Investigator") and/or other parties > Judgment based on the results of the investigation > Protective measures or a disposition for rehabilitation, as necessary."
>Just to point out here... the text on the front page of Wikipedia tends to be very neutral, matter-of-fact, and well written
This is true in general, though perhaps not always of Japanese wikipedia, and certainly not on legal topics. I have no love lost with machine translation (hard agree on naniwaduni's comment about "insidious failure modes"), but there are very few humans who would be able to translate this particular excerpt effectively. Legal translation is my job and I still found this sentence to be reasonably challenging.
"The juvenile protection procedure aims to deter and rehabilitate juvenile delinquents, and as a process up to the decision, "send and notify the facts of delinquency to the family court-family court investigator (hereinafter abbreviated as" investigator "). It is customary to go through the flow of "investigation by, etc.-a referee based on the investigation results-determine protective measures or protective measures as necessary"."
Not incredible (especially when it starts listing the process), but not as god awful as whatever you got. I'm curious if you omitted some part of the sentence when pasting in.
It's also worth noting that this is a pretty out-of-domain legal Japanese sentence.
My biggest pet peeve with MT is the incredibly brain-dead choice of using English as an intermediate language.
This effectively prevents MT from ever becoming useful for any A→B translation where neither A nor B are English.
Example: put "пружи́на" into Google translate (or Bing translate, doesn't matter)and translate to German. You'll get "Frühling" as the suggested translation.
This is due to the fact that internally "пружи́на" is translated to English first, resulting in "spring". This is then translated to German with the statistically highest probability being "Frühling", even though "Feder" would be the only correct match ("Frühling" would be "весна" in Russian).
If you now say "But wait! This should clear itself up once you add some context, right?" you'd be wrong.
Try "Die Sprungfeder kam gerade rechtzeitig." ("The coil spring arrived just in time,") and you'll get "Весна пришла как раз вовремя." ("Springtime came just in time").
It's ridiculous!
EDIT: happens with DeepL as well; only Yandex gets it right (presumably because they actually use an actual German/Russian-model.
For N languages you need to train N models if you go through English as an intermediate language, but N^2 if you don't use an intermediate language model. OK, computer time is cheap these days, but more importantly, for each language pair you need D high-quality documents translated in each of your two languages.
There are far fewer pieces of writing that exist in both German and Russian than in both German and English or both Russian and English.
There is cool-looking research on zero-shot language translation where the model learns its own internal "intermediate language". But you can be sure that if it produced reliably better results with the data available, the likes of Google would already be using it.
That's quite literally a cheap excuse. It'snot lack of training data (remember, you still need translations in the target language anyway!), it's a result of Anglo-Saxon cultural hegemony.
> There are far fewer pieces of writing that exist in both German and Russian than in both German and English or both Russian and English.
Right. And that's why Russia-based Yandex gets excellent results for German/Russian translation, I see...
> But you can be sure that if it produced reliably better results with the data available, the likes of Google would already be using it.
Nah, you're being way too optimistic there. Google is a business and as such, they don't chase SOTA outside of research (which doubles as advertisement). Good enough is the name of the game and it's simply infinitely cheaper to just train and maintain N-models (i.e. English/X and X/English) than to do the same for all combinations of X/Y-pairs (which would be N² models).
So as long as there is no incentive to do better, they simply won't, because they rely on the fact that English is the de-facto Lingua Franca and people translate to and from English more often than directly between other languages.
This would be an actual chance for companies from other countries (especially European ones) to step in and do exactly that, but again - "good enough" and reliance on English stops that from happening anytime soon. It's a missed opportunity especially with low-resource languages as demonstrated by the recent "Scottish" Wikipedia fiasco...
Do you have a source that that is actually what is happening behind the scenes? I work in this field and would be surprised if that is what Google is doing. My guess is this phenomenon is more due to a model that was trained on an Russian->English corpora first and then trained on Russian->German due to low resource, or something like that.
Someone like Yandex is going to have better translations because (shocker) they have larger russian corpora than google.
Indirectly yes, I do: https://codesachin.wordpress.com/2017/01/18/understanding-th...
> What this basically means is: If during training you provide it examples of English->Japanese & English->Korean translations, GNMT automatically does Japanese->Korean reasonably well! In fact, this is the biggest achievement of GNMT as a project.
It's exactly what Google have been doing for years now: train on X-to-English and English-to-Y only to get X-to-Y for free. Since the intermediate language only ever sees English as a source or target language, ambiguities like "spring" literally get lost in translation.
1. It is pretty out of date (ML is a fast moving field) - I doubt Google is using LSTMs for translation in 2020.
> Since the intermediate language only ever sees English as a source or target language, ambiguities like "spring" literally get lost in translation.
2. This article is trying to dumb down what Google is doing - but you're right that this is why ambiguities get lost in translation, due to pre-training on a different language pair.
That said, there isn't literally a process of "translate into english" and then "translate english into german". These models are trained on Russian-German corpora, but because there is little resources for that, they are supplementing with Russian-English and English-German.
Sure, but that doesn't change the fact the training data is focused on English-to-X and X-to-English corpora. The underlying architecture of the model is just an implementation detail that doesn't really affect this as demonstrated by my example.
> These models are trained on Russian-German corpora, but because there is little resources for that, they are supplementing with Russian-English and English-German.
This is exactly what I'd argue isn't the case at all. Otherwise words that have a direct 1:1 translation wouldn't be mistranslated and companies like Yandex wouldn't be able to deliver so much better results.
German-Russian isn't low-resource at all, given 95M and 150M native speakers respectively and a close history for the past 150 years. [edit]The rich cultural history of both countries resulting in a vast library of literature, theatre plays, news publications, films and the general cultural relevance of both languages is even more important.[/edit] It's simply (quite comprehensible) bias towards English for research taking place in the USA and the fact that it's much easier to compile English-to-X and X-to-English corpora in a predominantly English-speaking country.
There are tons of translated books, films, news paper articles, scientific papers, etc. available for Russian-German and Yandex, being a Russian company, naturally has no problem compiling a Russian-German corpus (since they're not biased towards English).
This also happens between e.g. Spanish and Bulgarian.
Definitely anything where a word has two or more different meanings.
Paper jam being my favourite example.
https://translate.google.com/#view=home&op=translate&sl=en&t...
It is translated as "(fruit) jam made out of paper"
This is, I think, the best way to look at MT.
As many people here have commented, machine translations are often incorrect, misleading, or just incomprehensible garbage. However, in situations with enough context, people can use MT to accomplish tasks that they would not be able to do without it. Some examples, and a discussion of implications for foreign-language education, appear in a paper I wrote last year [0].
[0] https://researchmap.jp/multidatabases/multidatabase_contents...
And I understand Chinese pretty well.
I bought babies, got good deals on babies, and even tried to return a defective baby once, but it didn't work out.
It looks like this has now been fixed, but it was good fun for a long time.
It's getting to be within spitting distance of being completely adequate (though not excellent) for translating long-form articles, which is amazing considering where things stood ten years ago.
In a sense, I feel that this is an even worse situation, because the word-salad signal that it isn't an accurate-meaning translation isn't as strong.
Maybe this was true 4-6 years ago, but no longer.
I'm curious what the domain is and what languages you're talking about, it's hard for me to evaluate your claims otherwise. I would buy it for English -> Tibetan for instance.
My observation of the failure rate (if you define failure as "meaning has been lost") is well above 80% with Google Translate for these languages and this type of text.
It's literally unusable because you have no confidence that the translation has not completely reversed or garbled the meaning of the text, and the failure mode is particularly unpleasant because you may be unaware that it has failed to translate correctly, since it produces text that appears to have meaning - just the wrong meaning.
Your comment about Chinese feeling like English written in Kanji is really interesting to me, perhaps it's a case of different perspectives - I tend to focus on grammar when learning a language, and so Chinese feels very different to English from my perspective.
But Asian languages are nevertheless related in more mysterious ways. Compare the intro to https://en.wikipedia.org/wiki/Mainland_Southeast_Asia_lingui... :
> The Mainland Southeast Asia linguistic area is a sprachbund including languages of the Sino-Tibetan, Hmong–Mien (or Miao–Yao), Kra–Dai, Austronesian and Austroasiatic families spoken in an area stretching from Thailand to China. Neighbouring languages across these families, though presumed unrelated, often have similar typological features, which are believed to have spread by diffusion.
In a European context, this is similar to the unusual closeness between the "unrelated" English and French, known to be due to intensive contact between the languages post-William-the-Conqueror. (In a world context, English and French are closely related regardless.)
It's not "literally unusable." Even for language pairs with poor translation, it considerably saves on manual translation time if you can have a translator look over sets of already translated (but lower-quality) sentences and correct those with errors.
Plus, there are actual studies out there comparing human translations to machine translations in a double-blinded evaluation by translators and the gap is not as high as your comment suggests.
For chinese-english (the other direction), bilingual evaluators have had difficulty differentiating between the machine translation and the human translation.
Having attempted to use CN->EN machine translation many times, I find this very surprising. Have you got a reference I can have a look at?
https://news.ycombinator.com/item?id=24302564
Bonus clip: https://www.youtube.com/watch?v=0E4aYgKmIko
-> Human evaluation does not agree that PBSMT is more accurate than NMT. Sure, the fluency/"sounds good" gains from NMT are much larger than the accuracy gains, but evaluations I've seen put NMT ahead on both metrics.
Before, I could skim outputs, and get a fair idea of which areas were poorly translated by where it was garbage. Now, I get the impression that (like with GPT or style transfer) something always comes out, whether it's related (or even necessarily has the right negation parity) or not.
And it is also worth noting (as you've sort of mentioned) that appearance of fluency can bias translators who are instructed to just evaluate accuracy, or to evaluate accuracy and fluency separately. There is ongoing research into that problem.
The neural networks are great at faking it, and they'll make up text that looks completely reasonable all day long. If that means inventing subjects and choosing random verb tenses, that's what they'll do. If you pick closely-related languages, you might be able to get a near perfect translation just by putting each word through a dictionary, and the neural networks can disambiguate things and smooth over the problems in the output by looking at local context. If you pick distantly-related languages, you might routinely need context from paragraphs away, or context which isn't present in the text at all but relies on some kind of human experience.
low resource translation is definitely still bad, you're right.
> near perfect translation just by putting each word through a dictionary, and the neural networks can disambiguate things and smooth over the problems in the output
Funnily enough, this is not a terrible explanation of how the current sort of approaches to low-resource MT work.
> low resource translation is definitely still bad, you're right.
No, you've misunderstood. Languages that are closely related to each other are fairly easy to translate between. Languages that aren't related are not, regardless of whether you have a large corpus to work with. English - Mandarin is not an example of "low resource translation".
The size of the corpus makes a huge difference. I don't see how can both be true that a. there is this massive failure of translation engines on these language pairs,
b. human translators were unable to differentiate between human translated chinese-english sentences and machine translated sentences.
It's not incredible - but it's also not awful or unusable.
If your use case is (1) being vocally excited about machine translation, or (2) having fun in your spare time, then machine translation is good enough for you.
If you have a genuine need to know what something in a foreign language says, machine translation is unusable.
This isn't really a good usability profile.
“- Milk tea or regular tea? -Yes!” Yes is the proper response in some languages. You can either change the language or your mindset.
And it’s not like these things are one offs, all languages are basically a bunch of quirks.
No doubt it’s the way of the future but I’d wait for the researchers to improve translation a lot.
Baidu recently added Saami to their system. They manage to make a Wikipedia article about wolverines into one about spreadsheets, depressed buttons and famous chickens: https://imgur.com/a/m91SbRw
A rough translation of what Baidu gave:
> No one is greater than zero. There are new things for spreadsheet[here Baidu used the English word for some reason] and all the things in it.
> English is an active wolf, but the background as wolf indicates that the active button is depressed.
> Too many places to install. Sweden was released from 1968 and July 1973. Probably number of times the alarm is to be reapeated within 39 minutes. 2010 is in total 66 chickens, but later a famous chicken and chickens 12th of march 2010, with a total of 54 chickens
Apertium's translation is often stilted, but it gets the meaning across: Wolverines are the largest animals in the marten family, etc. Google Translate isn't half bad at Norwegian→English, so you can even combine them to get a sort of Saami→English: https://translate.google.com/#view=home&op=translate&sl=no&t....
I tried Baidu again now with the same input, just to see what they would give into English, and now it's about old ports, finished volcanos and civic computers:
https://fanyi.baidu.com/#sme/en/Geatki%20(Gulo%20gulo)%20lea....
(but at least "Berners Convention" is slightly close to the input …)
It's amazing how low a bar big companies have for publishing ML models that just completely fail in the most absurd and spectacular ways, as if GM said "hey marketing department it looks like people want electric cars now, that seems easy just strap a battery on a chicken and hammer some wheels onto it and call it a day"