Wednesday, July 21, 2004

Machine-Augmented Translation

This article, from Dan Gillmor , concerns a coupling of machine translation with human interpretation. I think it's a fascinating concept, with obvious application potential.

One of the big problems in the American presence in Iraq has been a wall of misunderstanding and ignorance brought about by lack of language skills and a lack of information. Machine translation is getting better all the time, and we're seeing some of the results of that improvement here. The gear in the picture is a piece in the puzzle. From a satellite dish outside this tent at the Strong Angel II exercise, it's capturing and recording the Al-Manar TV station, a Hezbollah outlet in Lebanon. U.S. military and aid workers -- not to mention officials in Washington -- may find abhorrent what this and other Arabic stations are broadcasting, but they need to know what's being said; they need to understand what prominent Arab news outlets are saying about the U.S. occupation of Iraq and other issues. Audio is extracted from the news broadcasts, and converted to text in a speech-to-text program. Then the Arabic text is translated, also by a machine, into English. The results, twice removed from what the announcers have said, are rough approximations. But they capture the gist of the reports. The video, audio and texts (Arabic and English) go into a database. Sean Colbath, senior engineer at the BBN Technologies Speech and Language Processing Group, pulls out items -- short video quotes -- that look especially interesting or relevant to the key topics of the day. He bundles them into a file that goes to a human translator on the mainland, who gives an accurate rendering of what the broadcast snippet actually told the audience. This is a blending of human and machine translation capabilities, and it's a smart one. The machines get us part of the way. Humans capture more subtlety, but machines can winnow out a lot of the dross first. At the heart of the machine translation activity here is technology from Language Weaver, a southern California company. Impressive stuff, and it's barely at the beginning of what's possible. Colbath's gear is just one element in the translation tests here at Strong Angel, where people from the military, civilian, governmental and non-governmental spheres are testing new ways to communicate and collaborate in the aftermath of great human tragedy. Another experiment, which I have yet to see, sounds amazing: a chat in which you input in one language, someone gets your input in a different language and replies in his or her own tongue, which then comes back to you in your language.



No comments: