Spark

How AI might help rescuers use social media in a disaster

Machine learning can help identify urgent tweets
A house rests on the beach after collapsing off a cliff from Hurricane Irma in Vilano Beach, Fla., Friday, Sept. 15, 2017. (The Associated Press/David Goldman)

"This is as real as it gets.

Nowhere in the Florida Keys will be safe.
You still have time to evacuate.

Please Retweet.

#Irma."

That dramatic tweet is from the National Weather Service of Key West, Florida, posted just before Hurricane Irma made landfall in Florida. It's a good example of the role social media plays in getting crucial information to people affected by natural disasters.

And of course, as we've seen in recent weeks, social media is also used by people in need of rescue.  

But social media can also create challenges for emergency responders. During Hurricane Harvey, the U.S. Coast Guard advised people NOT to report distress on social media and to call emergency numbers instead.

Part of the issue is just the incredible volume of social media messages, as well as in authenticating the information.

Mahshid Marbouti is a PhD student at the University of Calgary who's tackling the issue. She wants to use machine learning to help emergency responders identify relevant social media messages during disasters.
Mahshid Marbouti

Currently, individual people at various EMS agencies have to sift through tweets to ensure their veracity, which is a time-consuming process. But with the aid of an AI, that process could be considerably more efficient.

"Machine learning can create a predictive model," she says. It can analyze tweets, for example, looking for tone and location, and rate it for its urgency. "The amount of fear in a social media post would be a defining factor," Mahshid says. The aim would be to complement human responders, not replace them.

Mahshid also notes that social media users can help responders with the type of tweets they send, she said. Hashtags are valuable, as well as things like an address, the number of people involved. But it's critical that hashtags should be reserved for real information, and not for example, for sympathetic messages.

So far, Mahshid's AI has has only been sifting through historical data, where it's been very accurate at identifying critical social media posts.

Currently, she has a working prototype that she hopes to deploy in the field soon.