Machine translation is one of the holly grails of artificial intelligence. The first approach to this problem was to construct a set of rules by studying the grammars of the two languages (called symbolic approach), but, as with other problems in AI, statistical tehniques (inferring "rules" from many translations of documents from one langauge to the other) seem to be more succesful now.
Each year NIST (National Institute of Standards and Technology in US) organizes a contest of machine translation software. The results for this year shows Google as the front runner. I'm proud to see University of Maryland (where I got my master degree) as a top contestant. I know they have a good research group in this area. Go Terps! :)