I played around with Google Translate in response to Daniel's post. I was surprised how quickly an English-French-English translation became stable. That's not so surprising when you think about it, but I was not expecting it. So I tried a few more languages.After being translated (something like English-French-Spanish-German-Portugeuse-Italian-English)
Surprising how good the "main" European languages are, though there's still a lot of loss of meaning there.

_________________
Occupy Babel!

Try translating through Korean as well. I've had really bad times trying to translate Korean into intelligible English with regards to some Korean go books.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

Yeah, crossing major language family boundaries really kills it. There was a hilarious site I was trying to find again that would translate back and forth between English and Japanese until it got a stable cycle, but I can't remember what it was.

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

Here's the link

I gave it the "original" expression above, and it went 25 times there and back before claiming "It is doubtful that this phrase will ever reach equilibrium."

Here's the result of 25 times there & back:

Yeah, my reason for poking around was the thought that all the entropy was coming from the Japanese, Chinese and Basque translations. Though I really should've checked the Czech in that case.

_________________
Occupy Babel!

Bah, I tried some longer sentences with a few complicated words, and got minimal loss of meaning. Then I tried "Don't fear the reaper":

http://translationparty.com/#9998014
http://translationparty.com/#9998017

Getting the opposite meaning in such a simple sentence is mind boggling.

If you translate "Don't fear me", you get equilibrium at "I am not scared" ... Google isn't afraid of me!

Awesome! That's what I was looking for!

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

I know Google Translate works statistically, by analyzing large corpora, but does anyone know more about how it works? Does it have some sort of underlying/universal grammar that it translates the text into and then uses that to translate it into the target language? (If it did, you'd have the amusing possibility of trying to translate English into English).

_________________
Occupy Babel!

People, learn languages. It is more rewarding than google translate

http://translate.google.com/about/index.html

I believe the reason that the European languages work so well is that the EU produces reams and reams of mega-multi-lingual documents.

Unfortunately, that front page doesn't answer my question, as far as I can tell (I probably should have put "grammar" in scare quotes in my question. I just meant some kind of representation that isn't part of any natural language but can be used to translate into any natural language. I bet something in the links would say.

However, elsewhere, I see this: http://en.wikipedia.org/wiki/Google_Tra ... ethodology

_________________
Occupy Babel!

i worked in automatic software tranlsation before (we were doing google translate before they did it themselves, and babelfish).
and we had a collections of the worst nonsenses. amongst the worst was "fan" translated by default as "ventilator" instead of groupie, and nut translated as in "bolt nut" instead as in nut the fruit.
also, in french "My name is bond, james bond" was translated as "Mon nom est esclavage, lien de james" which might get retranslated as "my name is slavery, chain of james"
also they was a mistake in our arabic dict so that lebanon was translated as yogourt (well there were other mistake but that was one of the funniest)

i think those errors are still on the internet in quite some places because at the time we were the only ones and some site translated their content with bablefish. At start google even reproduced some of them because they automatically extract translation form bilingual sites, and some of them where in fact automaticala translated.

as for multi translation, for the more rare language pairs google uses a pivot language (typically english) so the back anf forth effect is less than if you go through a truly weak kanguage pair ie
if you translate swedish-vietnamese it will do swedish->english->vietnamese (not sure about swedish vietnamese just an example)

_________________
In theory, there is no difference between theory and practice. In practice, there is.

Most European languages have similar grammar and are "easy" to translate more or less successfully. However, e.g. Finnish is entirely different from other European languages (except Estonian and Hungarian), and the translation doesn't work at all well. I suspect the same holds for Estonian and Hungarian too. I have never tried translating to/from those languages (and in most cases I wouldn't know if it worked well or not)

_________________
Offending ad removed

You can judge the accuracy of a translation by adding it to a cycle. So instead of English-French-English, you do English-French-Estonian-English (or English-Estonian-French-English).

_________________
Occupy Babel!

Last time i heard they were using vanilla prhase base translation, but with huuuge language models;
In those models there is no notion of grammar at all, you just have a dictionnary of sentence fragment where punctuations like. or , is treated just like grammatical words or noun .+ a "language model" which give the probability that a given sentence is good in the target language, and you run a minimizazion algo to find the les costly translation

an excellent site for explanations is http://statmt.org/

They might be doing more advanced stuff now, but at the start at least it was just the method described in this site with enormous compting power.
Those algos are notoriously bad to moves words far in a sentence (becasue to reduce computation you would typically hard-code a limit to the number of slots a translated word might move ), a big problem for example with german when you need to boot the verb at the end., So they might do something extra to fix that.

as for english-english translation we actually dit that to correct our rule base models:
you have a non statistical engine that gives you an unperfect translation, and you train a statistical model to translate from "machine translated" english to more human english. It works well in case were you want to translate text from a narrow domain with little bilinugal corpus available / the rule base translation deos the grunt work with domain specific dictionnaries and you polish the language statistically

The "universal grammar" is still a theorist dream as far as i understand. Some research try to statiscally construct a syntax tree in the source language and then transform it into the target language but i don't think that google does that (not sure though)

_________________
In theory, there is no difference between theory and practice. In practice, there is.