AI-Driven Decision Support for Disaster Management: Prioritizing Emergency Appeals using Computationally Efficient Deep Learning

Social media is an important lifeline in terms of real-time situational awareness in the aftermath of natural calamities. Nevertheless, emergency response teams are overwhelmed with information that they cannot process because of the general noise and actionable appeals. Although recent solutions in Deep Learning, namely, Bidirectional LSTMs (BiLSTM) and standard Transformers (BERT) exist, they are plagued by major drawbacks: RNNs are not always capable of capturing long-range semantic dependencies, whereas BERT models are not computationally efficient enough to be resorted to in resource-limited disaster areas. In order to fill this knowledge gap, this paper proposes Crisis-Connect, a computationally efficient decision support system that relies on a distilled Transformer architecture (DistilBERT). The model is benchmarked against traditional Machine learning (Naive Bayes), Recurrent Neural Networks (LSTM, GRU, BiLSTM), and Transformer models (BERT-Base). The experimental results on the HumAID dataset indicate that our framework gets an F1-score of 0.80, which is 5.2 percent higher than the RNN baseline without losing the performance of 95 percent of the teacher model. More importantly, the model incurs an inference latency of 1.46 ms, the model is 1.9x faster than standard LSTMs and 2.6x faster than BERT. Also, we use the Explainable AI (LIME) to confirm the model decision transparency. The present research has a role to play in the context of social innovation because it offers a low-latency, scalable tool that allows disaster management authorities to make data-driven decisions about edge devices without the use of high-end infrastructure.