The response is drilled into us from childhood. Whatever the emergency—fire, earthquake, or riot—we reach for our phones and… send out a tweet? It may sound like a luddite’s social-media dystopia, but if researchers at the Indian Institute of Technology-Bombay (IIT-B) get their way, Twitter might be as important to emergency response as 911 calls.
The researchers have developed a tool to turn tweets into emergency alerts, in real time. Called Civique (pronounced “civic”), it searches for tweets about emergencies, then sends information about those emergencies to responders and bystanders. Though its creators don’t anticipate fully replacing telephone reporting systems like 911, they expect it to supplement such calls with instantaneous, geo-located updates. The tool is described in a conferencepaper posted on arXiv.
“The problem with 911 is you’re asking questions,” which can take up valuable time, said Krithi Ramamritham, an IIT-B professor of Computer Science who led the team. With Civique, he said, “the emergency is detected automatically,” collecting information more quickly and from more people than any human emergency response could manage.
But Civique faces an uphill fight for widespread adoption. “People will not change their habits in emergencies,” said Tomer Simon, an emergency management expert for Israeli Home Front Command who researches how first responders should use social media. Simon points to a previous emergency crowdsourcing system, called Tweak, which he says failed because it asked people to use an invented grammar and vocabulary to tweet about language.
To track where emergencies are occurring, Civique would also require that users change their mindset to allow Twitter, and the world, to view their location. “99% of tweets are not geo-located,” said Simon, “so you really cannot know where the tweet is originating from.”
But this problem should be surmountable, said Dave Yates, an assistant professor of business information and analytics at the University of Denver who was not involved with the project. “With some training, you could develop the mindset that you should enable geo-locating,” he said.
When Civique reads a tweet, it uses a machine learning algorithm called a classifier to decide if the tweet describes an emergency, and if so, what sort of emergency. Ramamritham and his team trained the system on tweets they designed to be misleading to prevent the system from sending firemen after someone frustrated about being fired.
But before Civique can decide what the words mean, it has to decide what the words actually are. “People tend to tweet using a sort of SMS language,” said Diptesh Kanojia, a PhD student at IIT-B who worked on developing the tool. So before the tweet is fed to the classifier, words like “fiiiiiirrrreeee” or “hlp” passes through several layers of compression, normalization, and spell-checking.
“They did a pretty good job of trying to stress their system,” Yates said. But he noted that no matter how good the training, the number of tweets—and thus the number of mistakes—would be vastly greater in a real-world context, leading to false or missed alerts. “You would still have to overcome that or be aware of that,” he said.
Yates also worries that people could use the tool maliciously; Imagine a criminal who wanted to draw police away from one part of town, so tweeted about an active shooter elsewhere. “Things like that aren’t really a problem with the tool itself, but in how the tool would be implemented,” he said, noting that a false 911 call could achieve the same end.
Accuracy isn’t just important for its own sake, but for ensuring first responders believe the information they are getting. “They get used to their information coming in from certain sources they can verify,” Yates said. “This throws that model on its head.” First responders make life-and-death decisions quickly, and they cannot make the right decision if they have too little—or too much—confidence in the tool.
Field tests may help build confidence—the team has already developed a modified version of Civique for the IIT-B campus to alert the school to maintenance problems, and they are pushing the city of Mumbai to deploy it to track local emergencies.
The next step, Ramamritham said, is to correlate the tweets to provide first responders with richer information—to, say, combine one tweet geo-located in a building with another indicating that the building had collapsed, which would tell emergency crews that there were people in the collapsed building. “You can do that almost right now. We have that information,” he said.