Analyzing the Computational Efficiency of LLM Models for NLP Tweet Classification During Emergency-Crisis

Social media platforms which include Twitter, Reddit, Instagram, and YouTube have emerged as key information sources during emergencies or natural disasters. However, in the case of Twitter, it might be challenging to distinguish the tweets that are genuinely informative due to the overwhelming number of tweets that are sent during these occurrences. This study proposes a method for classifying the tweets as “informative” and “non-informative” using natural language processing. The technique considers the tweets’ subjects, hashtags segmentation, and locations for pre-processing. BERT, RoBERTa, LSTM, and DistilBERT are four different NLP models that have been trained and tested. The findings will indicate the highest classification accuracy for determining if a tweet is informative or not. Our findings reveal that DistilBERT achieved the highest accuracy of 92.0% in testing and 89% in precision of the Twitter dataset.