Instead of doing literal translations (word X in English is word Y in some other language), I read somewhere that Google Translations uses its massive processing power to analyze words, phrases and other elements of language in different versions of the same works that are provided on web sites, in scanned poems, essays and books that have multiple translations and so on.
This super-crunching does tend to provide a better “first pass” translation than traditional online word matching techniques. But it's still not perfect.
A game that people like to play is to send a phrase into Google Translate and have it returned in another language. Then send it back to the original language. Then repeat this process until the phrase sees no more variations as its passed back and forth. (Here's an automated way to do it using English -> Japanese -> English)
And here's an example using a famous Bob Dylan lyric (before you click, can you guess which one just by looking at the subject line of this post?)
(via Reddit)
