Authors: Kylen Solvik*, University of Colorado, Boulder, Lise Ann St. Denis, University of Colorado, Boulder, Jennifer Balch, University of Colorado, Boulder
Topics: Geographic Information Science and Systems, Hazards, Risks, and Disasters, Spatial Analysis & Modeling
Keywords: Natural language processing, topic modeling, social media, wildfires, geographic information retrieval
Session Type: Paper
Start / End Time: 11:10 AM / 12:25 PM
Room: Virtual Track 9
Presentation File: No File Uploaded
Machine learning has led to tremendous advancements in our ability to leverage unstructured social media data during disasters to help emergency management planning. However, much of the information shared on social media is redundant and/or already known to the emergency management team. It is a needle-in-the-haystack problem that requires informed, thoughtful approaches to augment powerful deep learning natural language processing methods. To address this challenge, this research investigates information flows between the public and the official management team. We matched Twitter data from ten of the most socially disruptive wildfires of the last decade in the US to the daily Incident Command Summaries (ICS-209) from those same fires, creating two timelines of wildfire: one from Twitter and the second from the official record. For each timeline, we extract place names from the text to find when locations are first mentioned and to identify cases where a place is mentioned on Twitter before it appears in the official management record, demonstrating a possible failure in public-to-official information streams. We will also perform topic modeling on both timelines to identify cases where the conversation on Twitter lagged behind the official record (a possible failure of official-to-public flows) or vice versa (failure of public-to-official flows). This research will identify the specific types of information that are not communicated effectively between the public and the official management team, and can aid in the design of highly targeted information extraction approaches for future disasters.