Mil Tech — Teaching Machines to Read
Nov 02 2009
Can machines be taught to read? The Defense Advanced Research Projects Agency (DARPA) intends to find out with its Machine Reading Program.
DARPA and the Air Force Research Laboratory have awarded BBN Technologies of Cambridge, Mass., $29.7 million in funding to develop an automated reading system that bridges the gap between naturally occurring text and the artificial intelligence reasoning systems that need such knowledge.
The initial use of such an intelligent learning system would be to automate military intelligence analysis, but BBN Technologies maintains the system would have civilian applications, such as providing access and automated analysis for libraries that have converted to digital texts.
Ralph Weischedel, principal investigator for the program, says, “One of our goals is to make knowledge available to machines much more cheaply, and the way to do that is by developing a machine that can read text and transform it into a formal representation the machine can understand.”
Weischedel notes the program is essentially a software project, using off-the-shelf hardware, that focuses on three subject matter areas — natural language understanding, knowledge representation and reasoning, and machine learning.
“Our approach involves all three area and involves the premier scientists from those disciplines,” he says. “While we’ve just begun, we want to grow into a full-fledged reading system so a machine can solve problems and transmit knowledge, building on actual reasoning capability.”
Weischedel points out that because there are substantial software components to be built into the system, “we’re at the nuts and bolts stage of where they can communicate. And we’re coming up with applications program interfaces that allow the components we build to talk to customers’ system, so we’re at the level of basic software engineering integration.”
The Machine Reading Program has a five-year contract life.
The artificial intelligence systems that might benefit from machine reading are used in logistics, diagnostics, planning, prognostics, and situation awareness areas. However, because much of the material in such military systems is in natural language text, which artificial intelligence does not process, DARPA wants to bridge the gap that exists between those two information states.
“What we’re doing is groundbreaking,” Weischedel says. “There’s no software that can do this right now.”
About the author: Alan M. Petrillo is a Tucson, Ariz., freelance writer who works in a wide variety of fields, writing for national and regional magazines and newspapers. He’s also the author of the historical mystery, Full Moon.
