AT&T is eavesdropping on its disgruntled customers via their Twitter messages. The project is an experiment designed to automatically pinpoint where and when people are having problems with their wireless connection. Software developed by AT&T researchers finds complaints about network problems on the social network (which now has 175 million users) and extracts the approximate time the tweet was sent and the location of its sender.
More and more firms are turning to Twitter in search of their customers’ voices, but usually they do this to understand how people perceive their brand or to respond to specific consumer problems. AT&T’s project is a novel way to mine the collective mood of the tweetosphere.
When AT&T secured exclusive rights to distribute the iPhone in 2007, the deal proved to be both a blessing and a curse. The much-desired handset brought huge customer growth and brand prestige. But along with that came soaring demand for wireless data that overwhelmed AT&T’s network in places such as New York and San Francisco, leading to dropped calls and sluggish connections.
The company has an automated network monitoring system that can detect connectivity problems, and customers can, of course, call in to report problems. But by mining messages shared on Twitter, AT&T gets extra real-time information and can prioritize fixes, says Jia Wang, a member of the company’s Internet and Systems Networking Research Center. “We are trying to identify three pieces of information: where the customer experienced problems, what type of problem, and when they experienced it,” she says.
Wang and colleagues use two levels of filtering to find tweets by frustrated customers, and they do this by tapping into the programming interface tools Twitter makes freely available. A general set of queries pulls in every tweet related to AT&T’s mobile service before a more rigorous set of rules homes in on those relating to service quality, for example messages containing words like “call dropped” or “3G.” This automated method was around 90 percent accurate at identifying genuine complaints, the researchers found.