Disaster and Urgency Detection from Social Media Messages using Natural Language Processing

Abstract

The widespread use of social media can be exploited to disseminate timely information
and updates to disaster response agencies by the people who are on the field when a
crisis strikes. Micro-blogging platforms such as Twitter and Facebook serve as apt
datasets to derive insights about the situation and target response and recovery. In this
paper, we use approximately 10,000 tweets from Twitter that have been hand-labeled to
indicate whether or not they are an emergency. The dataset is then run through machine
learning and natural language processing models to see how well they perform tweet
classification. These models can then be used to identify calls for help and allow search
and rescue team to manage their resources better, targeting specific areas to aid first.
Pretrained models such as Bert, DistilBert, and Ernie and also conventional models such
as Logistic Regression, k-Nearest Neighbors, and Random Forest are run on seven
extracted features. It is seen that pre-trained deep NLP models give the best F1 score
among other classifiers, which is better than current state-of-the-art models.