Yingjie Hu, University at Buffalo – Can ChatGPT Help First Responders During Disasters

How do we cut through the noise of social media posts during a disaster to find those who are desperate for help?

Yingjie Hu, associate professor of geography at the University at Buffalo, says the answer might be artificial.

Yingjie Hu is an associate professor of geography at the University at Buffalo. His research interests include geographic information science and geospatial artificial intelligence. His research on social media’s role in disaster response is supported by the National Science Foundation. Prior to joining the University at Buffalo in 2018, he was an assistant professor at the University of Tennessee. He received his PhD in Geographic Information Science and Cartography from the University of California, Santa Barbara.

Can ChatGPT Help First Responders During Disasters

In recent years, disaster victims have turned to social media when 911 systems become overloaded, pleading for help on sites like Facebook and X. These posts usually have important information about addresses and other detailed location descriptions necessary for first responders to reach victims.

Yet first responders often don’t have the bandwidth to monitor social media feeds during a disaster.

My colleagues and I wondered if ChatGPT could help. While this large language model has triggered controversies, it also has the unique ability to potentially help recognize location descriptions.

Whereas other existing methods, such as named entity recognition, require a large set of training data, ChatGPT only needs a handful of examples when guided by a good understanding of the typical location descriptions used by people during natural disasters.

So here is what we did: First, we supplied GPT models with 22 real tweets from Hurricane Harvey, which we had collected for our previous study of location descriptions. These tweet examples provide necessary geoknowledge by teaching these GPT models which words in a tweet described a location and what kind of location they were describing. Next, we tested the models on nearly a thousand other Hurricane Harvey tweets, and asked these models to extract location descriptions.

The results: Our geoknowledge-guided GPT models performed 76 percent better than default GPT models, as well as 40% better than the typical named entity recognition tools.

Our hope is that AI could help emergency services better identify victims from social media posts and help save more lives.

Read More:
[Buffalo] – How ChatGPT could help first responders during natural disasters
[Taylor & Francis Online] – Geo-knowledge-guided GPT models improve the extraction of location descriptions from disaster-related social media messages