Social Media Monitoring and Rapid Response to Earthquake Emergencies: A Study on Twitter Data

Earthquakes are among the most inevitable natural disasters our country faces. Throughout history, this natural phenomenon has caused significant losses multiple times, making the management of its dangers critical for the safety and security of the community. Access to instant and accurate data is crucial for rapid actions that need to be taken immediately after an earthquake. However, the data collected during this critical period is often voluminous and complex, making only a limited amount of data labeling possible. The primary goal of this study is to quickly and efficiently classify information of critical importance for disaster management and response from the data pollution on Twitter in such a situation. The most significant challenge of this process is to achieve maximum efficiency with a minimal amount of labeled data. In this context, the study examines Bidirectional Encoder Representations from Transformers (BERT), Robustly optimized BERT approach (RoBERTa) and BERT obtained through the distillation process (DistilBERT), along with traditional machine learning methods and zero-shot as well as fewshot learning methods. The effectiveness of these models in classifying data from online social networks after earthquakes has been evaluated, and their performance rates have been compared. The results of this study demonstrate that successful classification outcomes can be achieved even with a small number of labeled data.