Evaluating Bert Variants For Disaster-Related Information Classification: Identifying Situational Vs Non-Situational Data

Abstract

          Natural language processing (NLP) plays a critical role in disaster management, particularly in distinguishing between situational and non-situational information shared during emergencies. This paper presents an evaluation of various BERT (Bidirectional Encoder Representations from Transformers) variants for classifying disaster-related information. The study focuses on identifying the most effective approach for accurately categorizing data into situational (relevant to disaster response) and non-situational (irrelevant or secondary) information. Several BERT models, including BERT-base, RoBERTa, and DistilBERT, are fine-tuned on a dataset containing real-world disaster communications such as social media posts, news articles, and emergency updates. We compare their performance based on metrics such as accuracy and processing time. Our findings highlight the strengths and limitations of each variant, with a particular focus on how each model handles the nuances of situational information during crises. The results demonstrate that RoBERTa outperforms other models in accuracy, while DistilBERT offers a faster alternative with minimal trade-offs in performance. This comparative analysis provides insights into selecting the optimal BERT model for real-time disaster response systems, ultimately contributing to faster and more accurate decision-making during emergencies.