MAM: Multimodel Attention Mechanism for Social Media Natural Disaster Management Tweet Classification

• Published in: 2023 20th ACS/IEEE International Conference on Computer Systems and Applications (AICCSA) pp. 1 - 6
• Main Authors: Sangeetha, M., Manjula Devi, R., Sharma, Bhisham, Chowdhury, Subrata, Dhaou, Imed Ben
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.12.2023
• Subjects: Disaster management; Disasters; Earthquakes; Machine learning; Media; Multi-modal Data Fusion; Natural Disaster Management (NDM); Social Media Analysis; Social networking (online); Transformers; Twitter Analysis
• DOI: 10.1109/AICCSA59173.2023.10479344

People have been using social media as a category for exchanging content for decades. It has revolutionized communication and enhanced the sharing of information during emergency situations. The key features of social media are collective action, connectivity, comprehensiveness, and clarity. Consequently, it performed a significant function in natural disaster management by keeping track of and reporting disaster-related incidents. The volume and diversity of the data acquired from social media during a natural disaster pose the greatest challenge. For natural disaster management it is extremely difficult to derive actionable information from the data collected from social platforms. Various strategies have been presented in the literature to address the difficulties posed by social media for natural disaster management. The proposed work is a semi-supervised machine learning model for detecting and classifying tweets. The proposed work centre's on preparing the data by performing cleansing and various transformations, generating word embedding vectors with DistillBERT, using the vision transformer, the image model was constructed. The attention mechanism is utilized for text and image model integration. In the proposed work the data pertaining to seven distinct natural disasters, such as cyclones, floods, and earthquakes are analyzed. A novel decision diffusion technique is proposed for classifying them into informative and non-informative groups and evaluating the accuracy of the results. The MAM model increases the accuracy to 97% for crisisMMD dataset.