MIT Professor Gives Language Lessons to Computers
There’s a scene in the 2008 movie Iron Man where Tony Stark, the film’s inventor-superhero, threatens to donate one of his robots to a city college. You can tell by its cowed response that the computerized assistant understands the connotation is decidedly negative. In real life, software can’t yet comprehend that kind of abstract scolding. Programmers refer to such banter as “natural language,” and it’s tricky for computers to get because of its ambiguity and dependence on context.
Regina Barzilay, an associate professor at the Massachusetts Institute of Technology, is trying to make computers better listeners by making them play Civilization, a 20-year-old strategy game in which players build a city into an empire by vanquishing and absorbing neighboring cultures. A member of MIT’s Computer Science & Artificial Intelligence Lab, Barzilay, 40, developed a software program that begins with no grasp of the game. The computer “reads” the manual and then keeps returning to it while playing. As it races through thousands of simulations, the computer learns to connect words in the directions (“attack,” “build,” “capture,” and “revolt”) as the game unfolds.
The computer gets positive reinforcement—a higher score and a win—when it makes correct guesses about the meaning of words. When the computer loses, it traces back through its reading of the manual to see where its interpretation went wrong. A similar program without access to the manual won the game 46 percent of the time; after reading the instructions, Barzilay’s computer won 79 percent of the time.
Barzilay grew interested in natural-language processing in the early 1990s, as an undergraduate at Ben-Gurion University in Beersheba, Israel. She was inspired in part by her own experience as a young emigrant from Moldavia who had to learn Hebrew and English. Just as she struggled at first to understand the use of articles such as “the,” which have no equivalent in her native Russian, logic-based computers have difficulties with the inconsistencies of natural language.
Research like Barzilay’s may help computers eventually interact with humans in a more normal way. “You’d like to be able to ask for the largest state bordering New York and have it come back with the answer, ‘Pennsylvania,’” says Dan Roth, a computer science professor at the University of Illinois at Urbana-Champaign who does work similar to Barzilay’s. “And what happens inside the computer is none of your business.” Barzilay has been pushing this line of work forward, he says, in part by using a more interesting and complex game. She has a grant from the Defense Advanced Research Projects Agency to help robots understand natural language, not unlike those in Iron Man. As she puts it: “I want to see the computer benefit directly from human knowledge, without having a person in the middle who does a translation.”