Researchers have been working on machine translation for half a century, yet only in the past few years have microchips grown powerful enough to even approach the kind of processing human ears and brains perform effortlessly.
One simple way to do the job is called example-based translation. It distills a few million words worth of translated conversations into a giant set of syntax rules.
A computer program containing those rules works like a tourist with a phrase book, taking each word and plugging its translated equivalent into the right place in a sentence in the target language. The approach is crude but efficient.
In 1996, when the Department of Defense needed translation help for English-speaking forces in Bosnia, Carnegie Mellon scientists developed a rudimentary Serbo-Croatian-to-English system in a matter of months, relying largely on example-based techniques.
Waibel and other researchers are exploring a more sophisticated approach called Interlingua-based translation, which breaks down utterances into an internal, language-neutral representation of their meaning. Then, rather like what goes on in the head of a multilingual person, it generates a translated sentence that says the same thing.
Interlingua requires a lot of programming skill, but it provides a lot of flexibility in return.
"It's interpretation rather than translation," Waibel says.
When the computer hears "It would please me very much if you would be so kind as to reserve me a room in your fine hotel," it extracts something like "Reserve (object "room") (user "me") (form polite)."
Once the computer has captured the meaning of a sentence, repeating it in another language may not even be necessary. If someone has asked a question, the computer may be able to provide an answer from its own database.
Next Page > Hand-Held Wireless Devices > Page 1, 2, 3, 4, 5
