Hacker News Reader

Comment by Mahn | Hacker News Reader

I wonder if these "algorithmic errors" are the result of people "maliciously" training Google Translate, since there's the option to correct and give a better translation. I wouldn't be surprised if it were programmed to automatically assume that a correction is good if multiple people correct the same thing, which would then lead to these gaffes as communities like 4chan figured it out and "exploited" it.

View on HN

That was my first thought, but the article has this paragraph near the bottom:

  "Although translations are managed automatically, it is possible for users to suggest alternative translations manually.
  However, the BBC understands that this was not how the errors were introduced."

Substituting "Russia" for "Mordor" is oddly specific if this algorithm issue was caused without direct human input though.

As I understand it they're doing statistical translation - they compare versions of the same text in multiple languages and use it to automatically create a model of what things translate to. So if Ukranians are using the word "Mordor" instead of Russia when translating Russian text, then as far as Google Translate's concerned that's the correct translation of Russia.

both ways are "automatic". one is getting source from thousands of translations, the other from thousand of user corrections.

you can only say something isn't automatic if someone with authority over the system executed the manual step.

4chan also exploited reCAPTCHA when it was just bought by google and they used it to OCR books. reCAPTCHA provides you with two words, and one is much easier to recognize than the other. So 4chan users created a simple website with lots of reCAPTCHAs, typed in the easy word correctly to pass the test and typed in "nigger" instead of the second word. As many users raided at the same time, Google OCRed hard to recognize words as "nigger".

I always do the same when seeing a reCAPTCHA or other captcha from Google: Type the correct word properly, and a slightly off word for the other word, which seems plausible to naive OCR, but still is wrong.

From the ideological side, as long as Google profits from the work of users due to their monopoly, but doesn’t provide it as open dataset under a noncommercial license, I don’t wish to support reCAPTCHA.

Pro tip: the other word can be left out altogether.

I used to type aa before noticing that and this also worked like a charm.

The word you type is not going to be accepted just because it is similar to what it should be. It is double-checked multiple times anyway [0]. People like you just make OCRing slower, just as if you never typed reCAPTCHA at all.

[0] https://www.reddit.com/r/pics/comments/cygfx/4chan_is_using_...

That seems to be his goal. To not help.

Indeed. Google constantly takes work from volunteers, modifies it, keeps the modification under closed source, and heavily profits from them.

Helping a company that uses such a business model is not in my interest, as it means less volunteer power is available for actually open projects.

That’s exactly the goal. To not give Google work for free unless they also share it with others.

If Google wants actually meaningful manual OCR from me, they can have it – under GPL license, of course.

Why the quotation marks? If you were working on Google Translate, then dealing with this sort of thing would be on your TODO list as an exploit, not an "exploit." As a user of Google Translate it's not helpful, either (that a user thinks that Nazi Germany and Soviet Russia are both evil states does not mean that the user is helped by reading about a war of Mordor against Mordor.) It's funny and all, but it's definitely an exploit.

People have been doing "google bombs" like this for about a decade, haven't they?

Yes. I believe that this has happened on more than one occasion before. Maybe 4chan could be behind this.

I study Norwegian, and I've seen some odd things like that in Google Translate. For instance, one time I translated a sentence containing a prominent Norwegian university that got translated to Princeton in English. I'm inclined to think that this wasn't malicious or a joke, but simply a deficiency in the automation. I've read a little bit about how Google Translate's automated stuff works, but not enough to feel confident to suggest how or why these kinds of errors happen.

I would throw a wild guess that it tries to replace phrases with ones of equivalent meaning, so that the gist of a sentence is translated.

Things like "raining cats and dogs" doesn't probably translate well literally, for instance. It probably failed to categorise these universities as things that should not be translated in this manner.

I've seen this with measurements and currency.

Danish "... 15 meters ... 100kr" becomes "... 15 feet ... $100" which isn't helpful. (45 feet would not be helpful either.)

My favorite example (since fixed) was the word "amistad" translated from Spanish to English. It translated to "friendship," of course. You could add exclamation marks to it, like "amistad!" and it would translate to "friendship!" You could add more, and it would add more. But if you translated "amistad!!!!!" with precisely five exclamation marks, not four nor six but exactly five, it would instead translate to "murder!"

This sort of thing was mentioned in an early talk given by someone at Google Translate. (Sorry, I have neither the link nor the name of the person handy.)

As it makes use of parallel corpora, there were examples of documents mentioning (say) a famous person in the United States, but when the document was manually translated into French, the human translator selected a different famous person more well-known in France. It made perfect sense in the context of the document, but it threw a wrench into the automated translation based on those documents.

Both this and the article remind me of a strange Google Translate problem where the translation of "Austria" varied, depending on how many exclamation marks it was followed by:

http://itre.cis.upenn.edu/~myl/languagelog/archives/005485.h...

Reply on news.ycombinator.com