In today’s digital age, categorizing social media data, particularly from platforms like X, can be an effective strategy for identifying key first responders during emergencies, thereby improving overall emergency response efforts. In this study, we introduce a First Responder Classification System (FReCS), a framework that annotates and classifies disaster tweets from 26 crisis events. Our annotations cater for first reponders and their sub-layers. Furthermore, we proposed a classifier called RoBERTa-CAFÉ that integrates pre-trained RoBERTa with Cross-Attention and Focused-Entanglement components, improving the precision and reliability of classification tasks. The model is rigorously tested across publicly available disaster datasets. The RoBERTa-CAFÉ model outperformed state-of-the-art models in identifying relevant emergency communications, displaying its generalization, robustness, and adaptability. Our FReCS approach offers a pioneering technique for classifying first responders and enhances emergency management systems’ operational capabilities, leading to more efficient and effective disaster responses. FReCS annotated dataset and code are available on GitHub ( https://github.com/abdul0366/FReCS ).