Project Teams Humans and Artificial Intelligence to Scan Social Media Posts for Risk Data to Aid Disaster Response

When a disaster strikes, social media can be helpful in alerting the public quickly. Social media posts can also offer useful information for emergency response and decision making.

Given the immense amount of data in social media posts, only some of which may be important to emergency managers, researchers are using artificial intelligence, or AI, to make the process more efficient. Such computer systems use a process called machine learning, in which the computers are “trained” by humans, who help identify the characteristics of relevant posts in different situations.

Virginia Tech professor Chris Zobel is part of a multidisciplinary project that seeks to involve community members in working with expert systems to help uncover disaster-related risks posted on social media. The project team recently won a $50,000 grant from the National Science Foundation to fund initial work on the project.

Besides Zobel, a professor of business information technology in the Pamplin College of Business, project participants are researchers from the University of Texas at Austin, Brigham Young University, and George Mason University, as well as members of Maryland’s Montgomery County Community Emergency Response Team (CERT).

The group has demonstrated the potential of its “human-AI teaming” approach in work on social media posts related to COVID-19 risks and attitudes in the metro Washington, D.C., region. It hopes to show the wider applicability of its approach in providing situational awareness for emergency management and disaster response.

Situational awareness “is an understanding of what is happening and where it is happening during a crisis,” Zobel explained. “The idea is that when individuals or groups tweet about what they are seeing and experiencing in different locations in the surrounding area — whether it’s about a large gathering of people without masks, rising flood waters, or a large tree leaning over a road due to strong winds during a hurricane — emergency managers can use this information to understand what is happening in specific places and respond to the situation more effectively.”

The task of sifting through millions of posts for useful material would be daunting if not for computers. “Because so many people use social media, there is a lot of data to dig through,” Zobel said. “Furthermore, because most people aren’t necessarily just tweeting about the crisis, you have to actively search through this data for the subset of information that is actually relevant.

“This is why artificial intelligence is needed — a computer can do this large-scale searching through all the data much more quickly and effectively than a human can. But, the machine needs to be trained — it needs to learn what to search for — to do this correctly.”

Zobel said humans are thus used as part of machine learning, the process by which computers develop the ability to recognize patterns and correctly classify new information. “Machine learning is a form of artificial intelligence in that it replicates the process that humans go through to learn things.”

In the human-AI teaming project, trained humans would review the tweets (the project is focusing only on Twitter for now), categorizing them for relevance and the type of information they contain.

“This can be particularly effective,” Zobel said, “if the humans are local experts who can understand the context of the messages and assess their true relevance, like the CERT members we are working with.”

This expert classification of the messages is then fed to a computer algorithm, which iteratively learns to recognize the patterns inherent in them that make them relevant, he said. “The idea is ultimately to enable the trained AI system to analyze large numbers of new posts in real time and to identify whether they are relevant and how they should be classified.”

When the computer becomes an expert system, it can take the place of a human expert and store such knowledge as: “If we see a tweet with these particular characteristics, then most likely it is indicating that someone needs help,” Zobel said. “This information can then be provided to emergency managers to give them a very focused view about what is actually happening and where, as the crisis evolves.”

The group’s project brings together such disciplines as computer science and social science and such specialty areas as human-computer interaction, crisis informatics, and disaster operations. It draws on its members’ expertise on such topics as how people use social media in crises, how teams use complex technologies in disasters, and how volunteers can be motivated.

One of the project’s goals is to show that its human-AI team approach can help reinforce community resilience. “My role is to look at how the extra knowledge from social media can be used to enable emergency managers and their communities to become increasingly more resilient,” said Zobel, whose primary research interests are disaster operations management, humanitarian supply chains, supply chain resilience, and sustainability and environmental decision making.

Disaster resilience, he said, is the ability to prepare and plan for, absorb, recover from, and more successfully adapt to adverse events. “My research focuses on quantifying how much resilience, and what type of resilience, is exhibited by different organizations in different disaster situations, in order to understand how that resilience can be improved. Having access to more and better information about what is going on can help with various aspects of preparing for, resisting, and recovering from a disaster.”

Zobel would be tackling two key questions: What is needed to implement such a system in a CERT organization teaming up with volunteers to identify actionable data from social media? Within this process, how can the human-technology interaction be facilitated effectively?

He noted that the NSF award, part of its Civic Innovation Challenge, will be used for planning efforts “to convene a series of focus groups to assess the feasibility of working with different CERT organizations to implement the approach and to evaluate the receptiveness of emergency managers and government organizations to the idea.”

The team plans to pursue another grant from the NSF for the next stage of its work, to show the effectiveness of its approach and to further demonstrate that it can be used for other types of disasters and communities across the United States.