Abstract:
Social media has seen a boom in recent years, with its users growing worldwide.The
ubiquitous nature of social media has made it a useful tool to extract useful information
regarding disasters. As in case of any disaster, people upload data
related to the disaster on social media which can be very useful in the timely detection
of disasters and preventing loss of human lives and infrastructures. This
research aims to identify disaster-related tweets from bulk data, classify them to
the type of event they belong to i.e. to perform multi-classification, and identify
type of humanitarian aid-related information posted in a tweet during disasters,
with improved performance in terms of accuracy, F1, etc. To perform this multiclass
classification i.e. to detect disastrous events and further detect the type of
humanitarian information present in the tweets a DistilBert+CNN+LSTM-based
framework has been proposed in this paper. The framework is based on Distil-
Bert pre-trained embeddings and for multi-class classification, CNN+LSTM with
a self-attention layer has been used. After applying the proposed framework the
F1 score of 98 % was achieved in disastrous events classification and an F1 score
of 88% was achieved in information classification of tweets. These results obtained
from the proposed framework have shown improvement over other deep learning
models that were used as part of a comparative study.
