http://www.asiatranslate.net/you-translate.

Wishing everyone luck with translation
Regards

1. Apples satisfy me
2. I love apples

here 2 is more fluent, because that’s what people usually say in English… a related measure is grammaticality (“I loves apples” would score lower than 2. on this measure).

Google Translate is great with fluency, mainly because it really is based on what people say in a certain language. They have a lot of data for eg. English and Spanish, so they can easily check if something sounds fluent.

What they score lower on, however, is accuracy. To give an example, entering the Swedish phrase “Själland kan ses her”, the Danish translation becomes “Zealand kan ses her”, when it should be “Sjælland kan ses her” — ie. it changed the name. This makes Google Translate unreliable for getting hold of factual information, since key words and phrases may be distorted (although they may be very “readable”, you’re reading something other than a translation).

Besides, traditionally, RBMT has been quite the slow-poke compared to SMT; if you want to make an argument for RBMT, that’s not the place to do it. What RBMT _is_ good at though, is producing translations that are true in meaning to the original text. You can easily make an SMT-system such as Google’s produce a translations that looks quite fluent, but has completely twisted the meaning (esp. proper names, like translating “Praha” (Prague) into “Dublin”). Also, RBMT has the advantage of being easy to change if someone does notice a problem, whereas changing an SMT system involves retraining on an improved corpus (this might be helped by using a TMX instead though).

I hear that Google, which uses statistical machine translation, is 4 times slower than rule-based machine translation. Also, Google scanned the UN corpus of translated docs going back to 1945, so at least for 6 official UN languages, its good for translating political science–that’s whats in the memory…so Transclick’s subject domain specific dictionaries solves that problem by having translation memory that is matching the connotations in a context-sensitive way, producing better linguistic accuracy. Check it out.