Every armchair policy analyst who's ever read the front page of the New York Times knows, just knows, that those in power should have seen the latest crisis coming, whatever or wherever it may be.
But can you really predict grave conflicts simply by reading the newspaper? It may indeed be possible, but only if you can track many newspapers over time.
Digital Government researcher Devika Subramanian, professor of Computer Science at Rice University, is leading design of a computer program that predicts potential conflicts by analyzing the cumulative reporting in newspapers and wire services, in a process analogous to the work of a professional policy analyst. Her work enables the intelligence and diplomatic communities to make maximum use of non-classified print and Web-based information.
In close collaboration with her Rice colleague political science professor Richard Stoll, an expert in international relations, Subramanian is not looking to replace human professionals with automation and algorithms, rather to create a tool that would expedite and augment their work. "I collected six years worth of data in three minutes," says Subramanian, explaining how quickly the computer can sort through an archive. "I want to help the poor staffer with a sheaf of papers on his desk to summarize - and the person up the command chain who has to make a decision on the basis of that summary."
Subramanian hopes that the program will be used to create reports that can be used alongside those prepared by staffers. "Even the best-intended humans bring their biases into which parts of the complex information landscape to focus on. You get different answers if you weight the factors differently," she says. "Since the computer doesn't know what patterns to look for, it looks for patterns that are much more agnostic."
Her program allows a "double-check": An objective, machine-created analysis that can be used to help verify the thoughtful analysis prepared by human beings. "The computer can look for patterns in non-stationary time series and social networks that humans are not prepared to see." Eventually, the program will also allow analysts to change the weighting of the input data, in order to do scenario testing.
The program relies on dynamic Bayesian networks to extract events pertaining to interactions between countries from news reports (i.e., who did what to whom and when).
The computer learns the significance of certain words and phrases and calculates from their occurrence the probability that a news story is relevant. Subramanian follows the "-10" and "+10" conflict/cooperation weighting familiar to political scientists for rating the severity of events. The computer is trained on a small collection of stories marked by a human as being relevant or not. Subramanian says that one unanticipated challenge was the number of war metaphors that show up in sports reporting.
Once the computer extracts events from news stories for a certain period of time, it aggregates data at a regional level to compute the overall level of conflict (a number between +10 and -10) for that time period. A chart of average weekly conflict level in a region over an extended period of time, for example, twenty years, reads like an electrocardiogram, (another interest of the polymath Subramanian, who says that some of the algorithms in this project derived from her work on modeling protein-to-protein interactions.) Even more graphically, a wavelet analysis of this time series data creates charts where conflicts stand out with terrifying clarity. In one test example, the algorithm turns macabre artist: The Gulf War is a vivid yellow plume, eerily resembling an oil well fire.
But the point is less the intense color that represents a conflict at flashpoint, than the surrounding penumbra of cooler colors - they indicate a build-up, when there still might be time to divert events. "The model is particularly good at noting discontinuities-and you can go back and ask what was it that triggered sudden acceleration," she says. Currently, the model can predict about six weeks to eight weeks into the future. Subramanian has published a paper that validates the method in the context of events data extracted from 20 years (1979-1999) of Reuters articles on the Middle East, as well as an existing collection of events pertaining to the Cold War.
Those who are fluent in the languages of areas of continuing conflict, such as the Middle East and the Indian subcontinent, would be especially welcome. "We would like to do more with local newspapers," explains Subramanian, very much aware of what can be missed without that input.
"There's a bias we've found with the American papers and large, international news services. Incidents have to hit a high pain threshold before a story is considering important enough to run, so local skirmishes that are early indicators of regional conflicts may not be reported. In addition, once a conflict has been going on for awhile, it may start to be underreported because the attitude is, ‘Oh, it's just another bombing and it's too far away to interest our readers.' We would love to include local newspapers so we can include this essential reporting, but we don't have the manpower. We need more people who can read these languages. We have the algorithms; we're now trying to focus on getting the data."
Her work will not substitute for operatives blending into local populations nor is it intended as a replacement for experienced policy analysts - but she hopes it can be one more tool in the information arsenal.
|This site is maintained by the Digital Government Research Center at the University of Southern California's Information Sciences Institute.|| CONTACT POLICIES|