Robust NLP for Social Good: Transformer-Based Classification of Noisy Disaster-Related Social Media Messages

Abstract In the wake of terrible occurrences, social media has become the preferred source for real-time information. Social media has become a significant instrument for victims of catastrophes to express their needs, document incidents, request support and share information related to their situation. However, automated analysis may be fairly difficult in the case of casual, abbreviated, grammatically inconsistent, spelling-variant and other textual noisy communications. Artificial intelligence may be used to this data to enhance the response to catastrophes provided by technology and the well-being of humans and help in the improved detection and organization of communications related to disasters. The current research is a study on the feasibility of the use of Natural Language Processing (NLP) based on transformers to classify the social media postings about the tragedy with background noise. A social media dataset related to catastrophes is pre-processed adequately to develop classification models. For the comparison of the performance of the handling of textual content on catastrophes a transformer-based model and a conventional machine-learning baseline are taken into consideration. Some of the main performance indicators used to evaluate the models include accuracy, precision, recall and F1-score. The problems caused by casual and distracting social media text are given specific attention. This project aims to (1) describe the challenges of collecting and evaluating noisy user-generated data and (2) empirically evaluate transformer-based text categorisation in a disaster-response setting. The research explores the use of AI-driven natural language processing on social media platforms connected to catastrophes to assist in the timely comprehension of crisis-related communication and human well-being. Moreover, it adds to the greater issue of Artificial Intelligence for social benefit.