4

This post on Rocket News 24 described a Google Translate bug, in which

さよなら大好きな人

was translated into "Episode 78". This bug is still live as of the time of writing, and you can replicate it independently.

Does anyone know why the machine translate created this mistranslation?

An example of machine mistranslation etymology would be "Chinese fuck" , where 干 could be translated as "dry" and as "fuck" at the same time. Is there a similar reason why this bug occurred for Japanese?

  • Recently, I came across a similar bug. Google still translates the English When the Going Gets Tough, the Tough Get Going. into German as Take On Me.. My guess as to why this happens is that google translate uses a lot statistics, ie texts translated into different languages and builds a model from these data. So I suppose this is either some bug in pre-processing, wrong match-up etc. Also, this bug disappears if you change the original even slightly. さようなら大好きな人 gives the expected non-sense. But while interesting, this is probably off-topic here. – blutorange Dec 16 '14 at 13:34
  • Could you suggest a SE on which this would not be off topic then? I do want to know why this happened, if nothing, for curiosity's sake. Japanese seemed like the best place to post this. – March Ho Dec 16 '14 at 13:37
  • Google Translate likes to compare a lot of pages that exist in multiple languages to comb for translations of various terms and phrases. As a result, sometimes it makes mistakes, especially with city names (because a lot of companies might have their English speaking offices in London and their Spanish speaking offices in Madrid and the city listed at the bottom of the page will be London or Madrid by language, so the engine "learns" that London in Spanish is Madrid). – user0721090601 Dec 16 '14 at 17:14
9

The source of this mistranslation series was identified by Japanese internet users as parallel corpus data contaminated by a Japanese TV reality show program あいのり Ainori.

According to this article, a number of "episodes" other than "78" were observed as well.

トナカイさんの贈り物 → Episode 167
星の印 → Episode 58
砂漠の決断 → Episode 60
恋する勇気 → Episode 50
愛の泉 → Episode 57
飛べない蝶 → Episode 65
涙の記念日 → Episode 66
心の鍵 → Episode 82
無添加な恋 → Episode 112

And all of them coincided with Ainori's episode titles list, with disturbed episode numbers.

| improve this answer | |
4

This is due to one of the mayor disadventages of statistical machine translation systems as the one Google is using: In general nobody knows exactly based on which information a certain mistranslation was created on - so fixing is not a trivial thing. The algorithm just states: Based on hundreds of millions of lines of translated texts in my database A seems to be the most probable translation of B.

| improve this answer | |
  • Can you improve your answer with a citation? – March Ho Dec 18 '14 at 7:49
  • 1
    Is it reliable enough if I provide my own work as a source? t-c3.org/index.php/t-c3/article/view/28 – Daniel Dec 18 '14 at 9:02
  • I presume the claim that Google uses a statistical machine translator must originate from Google, so if you cited them in your paper, then the citation should be the primary source. Of course, if you are yourself from Google or otherwise work with their systems, then it's perfectly reliable. – March Ho Dec 18 '14 at 9:09
  • 1
    I thought you were asking for a citation regarding the disadvantages of statistical machine translation (SMT). The fact that Google uses SMT is stated here: translate.google.com/about/intl/en_ALL – Daniel Dec 18 '14 at 9:18
  • Thanks! Your paper was rather interesting in this aspect as well. – March Ho Dec 18 '14 at 9:20
3

One potential source of this kind of seemingly incomprehensible corpus-poisoning could be the use of translation memories from Computer-Aided Translation tools. Human translators not infrequently have to deal with badly formatted and segmented source texts, and in particular with texts where hard line breaks (etc.) have been used for layout purposes. This results in sentence fragments being handled by the program as separate translation units, while the target text may require a completely different word order. If the translator does not re-join the split sentence in some way (which may not even be possible) but merely inserts the translation in a target-language-appropriate order across the existing units, then the memory will end up containing garbage which applications may then attempt to re-use later (particularly as these fragments are often short - perhaps even single words - and thus more easily matched).

| improve this answer | |
2

Look if the surrounding context has anything to do with Japanese pop-culture. The "花*花" duet had a song in 2012 titled "さよなら 大好きな人". "Take on Me" and "When the Going get Tough, the Tough Get Going" are also song titles.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.