Identifying firsthand accounts of emergencies from social media

With the recent explosion in availability and use of internet social media, citizens now utilize these resources to transmit information quickly about events as they unfold. However, for responding personnel in emergency situations, it is often difficult to sift through the enormous quantity of data within such sources to find the most pertinent information. The ability to filter messages is critical to, for example, identify firsthand accounts from persons within the direct vicinity of events. Nevertheless, on social media platforms such as Twitter, location-based information is often missing or unreliable. This paper outlines an approach to probabilistically identify the likely locations of individuals on Twitter based on their content, and from socially connected users with more reliable geographic information. We utilize measurements of user content similarity and Gaussian mixture modeling to infer “hotspots” of the likely location of users in emergencies. We are able to achieve upwards of 70% accuracy of Twitter user home cities without using any prior knowledge of geographic boundaries to look within.