Detection of Urgent Messages Shared on Twitter during an Earthquake using the Deep Learning Method
Mücahit Söylemez, Ali ÖztürkThe ability of Twitter to provide real-time information during disasters is becoming more widely acknowledged, making it an essential forum for people to voice their concerns and ask for help during emergencies. These platforms can speed up the distribution of help, but they are also prone to false information, which might make disaster response more difficult. Using a carefully selected dataset of 10,200 tweets that have been extensively preprocessed and tokenized for reliable training and validation, this study uses deep learning models, such as LSTM, BLSTM, and BLSTMA, to classify tweets during earthquake events into two categories: “under the debris” and “not under the debris.” The model performance was further improved via hyperparameter adjustment, which included neuron counts, dropout rates, dimensions, and embedding types. The results of this study showed that while the BLSTMA model had the best accuracy (96.64%) and F1 score (0.9116), conventional machine learning techniques like XGBoost and SVM. However, in other measurements, it was shown that standard machine learning techniques like SVM and XGBoost performed better. Using Bag of Words vectorisation, SVM obtained 95.81% accuracy and an F1 score of 0.9579, whereas XGBoost earned 95.84% accuracy and an F1 score of 0.9584. By demonstrating the usefulness of the BLSTMA model in real-time disaster response and the complementary advantages of conventional approaches in the analysis of complex disaster data, these findings highlight the significance of customising machine learning and deep learning approaches to particular tasks.
