How ChatGPT Could Help First Responders During a Disaster

When 911 systems become overloaded, like during Hurricane Harvey in 2017, those who need help from first responders are typically out of luck and on their own until things calm down.

But new research involving ChatGPT — generative AI technology — and some of the lessons from Harvey may have solved a piece of the response for future situations when 911 systems go down and stranded residents turn to social media for help.

During Harvey as countless residents became stranded, victims took to social media and posted their locations and pleas for someone to come and help.

But the number of social media posts and various locations was too much of a puzzle for first responders to solve. They simply didn’t have the resources to follow every tweet and hashtag and extract actionable location information.

Researchers at the University at Buffalo, with support from the National Science Foundation, published a study in which they developed ChatGPT models with carefully constructed prompts and found that they could extract from myriad tweets sent during Harvey location data that would help first responders find an exact location for stranded residents.

ChatGPT, developed by OpenAI, is built on a large language model and “pretrained” on a vast amount of information from across the Internet. Through this process, ChatGPT learns language patterns used by humans and can respond in a human-like way. For instance, prompt it with “What is the capital city of California?” and ChatGPT can generate an answer close to, “The capital city of California is Sacramento.”

That’s significant when you consider what transpired during Harvey when stranded residents turned to Facebook and Twitter and posted their locations with the hope of someone coming to the rescue.

“Many of those social media posts contained location descriptions, such as door number addresses, road intersections and highway exits,” said Yingjie Hu, associate professor in the University at Buffalo Department of Geography, within the College of Arts and Sciences, and lead author of the study.

Under such conditions, there was no way for first responders to monitor social media and target stranded residents. Previous research, using Named Entity Recognition (NER), had sought ways to extract locations from similar text, but a message such as “Family needs rescue at 1280 Grant Street, Cypress, Texas 77249,” would prompt the model to focus on each word instead of the entirety of the sentence, yielding unreliable responses.

“Those errors could make first responders arrive at the wrong locations and waste rescue time,” Hu said.

The generative AI model can quickly be adapted to a particular task when provided with specific prompts. In the experiments at the University at Buffalo, the researchers selected 22 examples of location descriptions from Hurricane Harvey posted to Twitter. The 22 tweets covered 11 categories of location descriptions, such as addresses, road intersections, nearby neighborhoods and highway exits.

“We then used the default GPT models not informed by the 22 examples and typical NER approaches and tested the geo-knowledge-guided GPT models on the same data set of tweets. The result was that the geo-knowledge-guided models performed much better than the other models.”

The geo-knowledge-guided models knew which multiword phrases to use as a location description instead of picking just one word out of an entire address such as “Grant” in Grant Street.

“Implementing this research in emergency response might involve several software components,” Hu said. “We may need a component that can use the application programming interface (API) of social media platforms to retrieve disaster-related posts [and then] we could use the methodological framework developed in this research to identify location descriptions.”

He said misinformation, given its potential to show up during a disaster, may also need to be “handled” with an additional software component as well.