Dear Harald
thank you for the clarification, but i still don't understand.
There are no grammatical rules here, I mean if you translate word by
word you will get the right translation, if you add a single word you
get the right translation, even if you add a single letter (to make it
female instead of male) you get the correct translation.
There is a very simple test, take the sentence
إسرائيلى يقتل فلسطيني
which means "israeli kills palestenian" which is the exact map of the
Arabic sentence word to word (arabic from right-left, english from
left-righ)
إسرائيلى = israeli
يقتل = kills
فسطينى = palestenian
this sentence is translated to "Israeli killed by Palestinians" which
is the total opposite meaning.
Try to switch the words israeli and palestenian in the original Arabic
sentence, you will get the correct translation of "Palestinian kills
Israeli"
I was wondering how does "contribute a better translation" option in
Google translate work, and if it can be used to force this misleading
translation?
thank you.
On Jan 4, 7:34 am, Harald Korneliussen wrote:
> Remember, it does not know any grammatical rules the way we do. When
> figuring out a sentence, it relies on statistics. My guess is that
> phrases like "Israelikills Arab" are less likely to turn up in the
> English corpus, and more likely to turn up in the Arabic, and the
> other way around for "Arab killsIsraeli". So, when it for some reason
> becomes uncertain of what is the subject and object of a sentence, it
> decides on what it thinks is "most likely".
>
> What is certain is that it is automatic, and not a product of racism -
> except possibly racist bias in the corpus the translator has been
> trained on. Which could be the whole web.
>
> Try translating "Israelisettler kills Palestinian man". I would guess
> there are somewhat more mentions of settler violence in the English
> corpus, so maybe that additional word is enough to push it to the
> correct translation.
>
> (I also agree with SPARKILL: If you use the translator for anything
> important, you should always go through its output and correct obvious
> mistakes - and be aware that since it is a statistical translator, it
> will make errors such as these.)
