Authors: Morteza Karimzadeh*, Purdue University, Luke Snyder, Purdue University, Yi-Shan Lin, Purdue Univeristy, David S. Ebert, Purdue University
Topics: Geographic Information Science and Systems
Keywords: social media, deep learning, situational awareness, natural language processing, geo-text
Session Type: Paper
Start / End Time: 4:30 PM / 6:10 PM
Room: Virginia B, Marriott, Lobby Level
Presentation File: No File Uploaded
Many first response groups or emergency operation centers are tasked with real-time monitoring of events or day-to-day activities to identify potential disruptions. In recent years, such organizations have been increasingly using public social media posts for situational awareness, essentially leveraging social media users as human sensors on the ground. Systems built for this purpose rely heavily on filtering and classification of social media posts to eliminate large amounts of noise (e.g. irrelevant posts, robot posts, or job advertisement) to provide relevant information to the analysts for the type of event or disaster under study. Linguistic (semantic) ambiguities, however, make this filtering and classification a challenging task. For instance, “My team is on fire tonight” may trigger a false positive alarm for a “fire event” classifiers. In this paper, we present a novel interactive deep learning-based classification system for identifying social media posts (tweets) that are relevant to the analysts’ tasks in situational awareness. We evaluate our approach on a test dataset. Our results show significant improvement over baseline filtering approaches. Further, we present the integration of our deep classifiers within SMART (“Social Media Analysis and Reporting Tool”), our geovisual analytics platform that has been in use by more than 20 organizations across the U.S. for daily use or event monitoring. This integration allows analysts to train the system’s classifiers in real-time for any event of interest.