[Proposal-ML] Efficient Deep Learning for Disaster Tweet Classification

Isak Tønnesen, André Moreira Leal Leonor, Guilherme Félix Diogo

Keywords: Natural Language Processing, Disaster Detection, Social Media Analysis, Resource-Efficient Deep Learning, Emergency Response

TL;DR: [Proposal-ML] A lightweight deep learning system for real-time disaster detection in tweets, achieving near-SOTA performance with minimal compute requirements for deployment in resource-constrained settings.

Abstract: Social media platforms have become crucial communication channels during emergencies and natural disasters, with Twitter emerging as a primary source of real-time situation awareness. While large-scale deep learning approaches have shown promise in disaster detection, there remains a critical need for efficient, deployable solutions that can operate within practical computational constraints. We propose a lightweight deep learning architecture for disaster tweet classification that combines pre-trained word embeddings with carefully engineered linguistic features. Our approach aims to achieve near state-of-the-art performance while requiring minimal computational resources, making it suitable for real-world deployment in resource-constrained environments. The system will be developed and evaluated using the Kaggle Natural Language Processing with Disaster Tweets dataset.

Submission Number: 40