Open Access

ARTICLE

Attention-Enhanced Hybrid Deep Learning for Disaster-Related Tweet Classification

Sheraz Ali Hassan¹, Hamid Masood Khan¹, Muhammad Javed¹, Mohd Faizal Bin Yusof², Jamil Abedalrahim Jamil Alsayaydeh^3,*, Fida Muhammad Khan⁴, Inam Ullah^5,*

1 Gomal Research Institute of Computing (GRIC), Gomal University, Dera Ismail Khan, Pakistan
2 General Education and Foundation Program, Faculty of Resilience, Rabdan Academy, 65 Al Inshirah Street, Abu Dhabi, United Arab Emirates
3 Department of Engineering Technology, Fakulti Teknologi dan Kejuruteraan Elektronik dan Komputer (FTKEK), Universiti Teknikal Malaysia Melaka (UTeM), Melaka, Malaysia
4 Department of Computer Science, Qurtuba University of Science and Information Technology, Peshawar, Pakistan
5 Department of Computer Engineering, Gachon University, Seongnam, Republic of Korea

* Corresponding Authors: Jamil Abedalrahim Jamil Alsayaydeh. Email: ; Inam Ullah. Email:

Computers, Materials & Continua 2026, 89(2), 64 https://doi.org/10.32604/cmc.2026.084547

Received 24 April 2026; Accepted 22 July 2026; Issue published 15 September 2026

Abstract

The increasing frequency of natural and human-induced disasters has intensified the need for reliable methods to identify crisis-relevant information from Twitter/X streams. However, tweets are often short, noisy, informal, ambiguous, and context-dependent, making disaster-related tweet classification challenging. This study proposes a lightweight attention-enhanced hybrid deep learning framework for binary disaster-related tweet classification. The framework integrates CNN-based local feature extraction, recurrent contextual modeling, static pre-trained word embeddings, class weighting, training-only data augmentation, and learned neural attention. Two architectures, CNN–LSTM–Attention and CNN–BiGRU–Attention, are evaluated on the labeled Kaggle Disaster Tweets dataset using a leakage-aware protocol in which augmentation is applied only after dataset partitioning. Conventional machine-learning, standalone deep-learning, and Transformer-based baselines, including BERT, RoBERTa, DistilBERT, and CrisisBERT, are evaluated under the same experimental setting. Experimental results show that CNN–BiGRU–Attention achieves the highest accuracy of 94.81%, while CNN–LSTM–Attention provides slightly higher ROC-AUC and disaster-class recall. Compared with Transformer-based baselines, the proposed models achieve competitive performance with substantially fewer trainable parameters. These findings indicate that lightweight attention-enhanced hybrid models can provide an effective accuracy–efficiency trade-off for disaster-related Twitter/X monitoring, although broader validation across unseen crisis events, platforms, languages, and real-time streams remains necessary.

---

Keywords

---