VictimFinder: Harvesting Rescue Requests in Disaster Response from Social Media with BERT
Topics: Geographic Information Science and Systems
,
,
Keywords: social sensing, tweets, rescue, disasters, natrual language processing, BERT
Session Type: Virtual Paper
Day: Friday
Session Start / End Time: 4/9/2021 03:05 PM (Pacific Time (US & Canada)) - 4/9/2021 04:20 PM (Pacific Time (US & Canada))
Room: Virtual 14
Authors:
Bing Zhou, Texas A&M University
Lei Zou, Department of Geography, Texas A&M University
Binbin Lin, Department of Geography, Texas A&M University
Mingzheng Yang, Department of Geography, Texas A&M University
,
,
,
,
,
,
Abstract
Social media platforms are playing increasingly significant roles in disaster responses and rescue operations. Users could post rescue requests along with their addresses on social media during emergencies while responding agencies search for those messages and send help. However, efficiently leveraging social media in rescue operations remains challenging because of the lack of tools to accurately and rapidly process social media data. Analyzing social media data, such as Tweets, relies heavily on natural language processing (NLP) algorithms to extract information from the textual contents. The introduction of bidirectional Transformers models, such as BERT (Bidirectional Encoder Representations from Transformers), has significantly outperformed previous models in numerous NLP tasks, providing new opportunities to automatically and precisely understand and classify social media data for diverse applications. This study examined the performance improvement of BERT-based models in classifying rescue requesting tweets. We selected eight trained BERT and BERT-based models and compared them with a general-purpose word embedding baseline model. A total of 3,191 manually labeled tweets posted during Hurricane Harvey were used as the training and testing datasets. Experiment results show that all BERT-based and optimized BERT models have significantly increased the rescue requesting tweet classification accuracy. The f1-score of the best model is 0.919, which outperforms the baseline model by 10.6%. The developed model will advance the efficiency of finding victims from Twitter and promote social media use for rescue operations in future events.