Analysis of Linguistic and Geographic Disparities Across Hurricane Related Rescue Request Social Media Data

Keywords: Natural Language Processing, Social Sensing, Disaster, GIS
Session Type: Virtual Paper Abstract
Day: Tuesday
Session Start / End Time: 3/1/2022 05:20 PM (Eastern Time (US & Canada)) - 3/1/2022 06:40 PM (Eastern Time (US & Canada))
Room: Virtual 18

Abstract

Since the burgeoning of social media platforms, social media data have infused themselves into the geospatial big data that had been widely leveraged among disaster related researches. While most researchers analyzed data only pertinent to disasters, few of them targeted at analyzing the rescue request data. However, capturing the dance of rescue request data can unveil deeper underlying issues due to it is a representative of the more severe form of disaster responses. With the help of a BERT-based rescue request classifier designed and trained from previous studies, we manage to narrow our scope to the rescue request twitter in Hurricane Harvey and Hurricane Irma. The research objective of this paper lies on two fronts. First, to certify if a universal rescue request classifier exists and provide proper guidance to social media users of how to draft rescue messages so that the classifier can be reliably applied in future events. The universality of the classifier is also validated with the tagged Twitter data from Hurricane Irma. We extract the rescue request tweets from the whole dataset to scrutinize its linguistic patterns by analyzing the keyword, keyword co-occurrence and dot product of the embedded vectors. Second, we attempt to delineate the relationship between rescue request Twitter activities and the actual damage caused by hurricane events by visualizing the geographic distribution of the Twitter data.