Google Translate changes its translation engine
Google has dumped Systran, the company that had been powering much of Google's web-based language translation service. Google had developed its own machine translation skill, but until recently, Google was only using its own technology to translate Arabic, Chinese, and Russian text and web sites. Now Google is using its own engine to translate 25 different language pairs.
Machine translation is an imperfect science. The computer translates most words literally and doesn't do a great job of capturing nuance or proper sentence structure. It doesn't appear that Google Translate is any more accurate today than it was a few weeks ago. But the move should help Google set its service apart from competing offerings from Babel Fish and Microsoft. Both companies' translation services are powered by Systran.
[via Google Operating System]
Machine translation is an imperfect science. The computer translates most words literally and doesn't do a great job of capturing nuance or proper sentence structure. It doesn't appear that Google Translate is any more accurate today than it was a few weeks ago. But the move should help Google set its service apart from competing offerings from Babel Fish and Microsoft. Both companies' translation services are powered by Systran.
[via Google Operating System]













Comments
3
Subscribe to commentsmichaelOct 23rd 2007 4:52PM
I still find Windows Live Translator to be the better of the two, especially with showing 2 dynamic website panes in different formats, but it's good to see Google step up and raise the bar.
It might encourage Live Translator and Babelfish to do something new too.
McQOct 24th 2007 1:36PM
Try this bit of doggerel: "Smart cars grow faster with jelly legs.". Nonsense, but not difficult to translate. In French, Systrans provided the reasonable result: "Les voitures intelligentes se développent plus rapidement avec des jambes de gelée.", whereas Google could only do "Smart voitures croître plus vite avec les jambes de gelée."
I'll stick to babel fish, thanks.
Vadim BermanNov 12th 2007 7:34AM
It's not like there are just two of them, Google and SYSTRAN.
The choice is much bigger, and the results are frequently better (e.g., IBM's WTS, or Promt), but most of the higher end ones are corporation oriented.
But yes, pure statistical MT is definitely not the way to go; despite the popular opinion, NIST's 2006 evaluation results do not reflect the quality accurately:
http://forum.digitalsonata.com/forums/thread/32.aspx
One thing which the result is a mysterious figure which no one sees, such as PageRank, and another is outputting human language, which requires much higher accuracy.