Putting Google to the Test in Translation | Hacker News Reader

Putting Google to the Test in Translation

nytimes.com

16 pointsseer7 comments

The Systran system is quite old, check out how Hofstadter thrashed it some years ago in his "Le Ton beau de Marot"(http://en.wikipedia.org/wiki/Le_Ton_beau_de_Marot) (awesome book on machine translation and literature in general, BTW). Looking at these I thought that (i) Google totally blows the competition out of the water in French, Spanish, and German (I even found the German translation slightly more readable than the human one - "uneasy dreams" sounds better than the human's "anxious dreams"); but (ii) the Russian and Arabic translations are not very readable. Too bad they didn't provide a comparison in Chinese.

Many old school linguists often scorn the no-model, statistical approach to machine translation that Google adopts (my advisor was one of these), opting for more semantically-based models instead. I have yet to see results comparable to Google's (i.e. in unconstrained domains) from this school.

How can we know about the linguistic models that Google uses? Are these from some papers published by Googlers?

For a simple introduction to their approach you can read Peter Norvig's explanation of how Google performs spell checking (http://norvig.com/spell-correct.html) or his video lecture (http://www.catonmat.net/blog/theorizing-from-data-by-peter-n...). You can get links to their results for NIST 2005 evaluation from here: http://googleresearch.blogspot.com/2006/04/statistical-machi...

There is something strange, in the Arabic article both the human and the Google translation say "[..] soccer great Diego Maradona" where as the original arabic there is no mention of anything resembling "great". Was the human translator influenced by the Google result ? or is Diego Maradona always referred to as "soccer great Diego Maradona" in English press and google is picking up on that?

Edit:I suppose it is the latter since the Human translator was not trying to have a correct reference translation for testing the engines. He totally skipped translating the words " in South Africa ". Imho, this makes the results much less interesting.

Also in the french translation it is interesting that "Le premier soir" translates as "The first evening" if you type it in google without the rest of the sentence, but becomes "The first night" after you put it in context.

> I suppose it is the latter since the Human translator was not trying to have a correct reference translation for testing the engines. He totally skipped translating the words " in South Africa ".

Each human translation also has the source below it, which ranges from old classics to relatively recent publications. So, yes, the human translators were not trying to have a correct reference translation, which imho makes it only more interesting.

> becomes "The first night" after you put it in context.

which makes me suspect this one happened to one of the inputs used to train the translator also.

The comparisons with Yahoo and with Bing are quite interesting.

The Russian version confuses two engines because of hyphenation (perhaps due to line-wrapping in the original article)

Reply on news.ycombinator.com