Classification of Humanitarian Crisis Response Through Unimodal Multi-Class Textual Classification

In the realm of disaster response it is crucial to emphasize the significant role that social media plays as an immediate and dynamic source of information. In times of both natural and human- created calamities citizens resort to sites like Twitter to broadcast updates in the form of text and multimedia. These updates vary from reporting on wounded or killed individuals to damage assessments and the hunt for missing or discovered folks. The value of this online information for humanitarian organizations is well-established, a large chunk of research has typically concentrated on textual content analysis. This research study seeks to offer a unique technique in textual classification of a vast and diversified multi-modal dataset dubbed "CrisisMMD," which is collected from Twitter after significant natural disasters. This dataset offers three separate types of annotations, each geared to fulfill a wide variety of crisis response and management needs, responding to the various requirements of humanitarian organizations. In our attempts to increase the efficiency of humanitarian operations, we argue for a approach that leverages the power of textual material inside social media. To achieve this objective we apply cutting-edge machine and deep learning techniques, to construct a framework with a common representation successfully breaking down the conventional boundaries of textual modes. Our study done utilizing real world catastrophe datasets indisputably proves the superiority of our suggested method when compared to other models. The results of this study not only underline the untapped potential of textual aspects in crisis response but also offer an appropriate framework for categorizing the Multi-Class CrisisMMD dataset textually. This method signals a new era of strong and effective catastrophe management, promising to have a big influence on the humanitarian scene. Experiments conducted on the CrisisMMD dataset, which simulates real-world catastrophes, have yielded validation results demonstrating the consistent and significant outperformance of our technique. The findings reveal a notable advancement compared to the previously published works by 3% and the current state-of-the-art surpassing them by over 6.73%. The performance improvement observed ranges from 3% to 6.73%, underscoring the effectiveness of our approach in addressing the challenges posed by real-world catastrophe scenarios.