A HYBRID MBERT–LSTM ARCHITECTURE WITH SPATIAL ATTENTION AND SMOTE FOR BILINGUAL DENGUE SENTIMENT CLASSIFICATION

Abstract

Dengue fever is a major public health concern in tropical and subtropical regions, affecting millions of people each year and continuing to spread rapidly across different parts of the world. At the same time, social media has become an important source of real-time public opinion during disease outbreaks, providing valuable information that can support disease surveillance and public health decision-making. In this study, a hybrid deep learning framework, referred to as mBERT–Stacked LSTM with Spatial Attention (SSA-LSTM), is developed to classify dengue-related sentiments expressed in both English and Urdu tweets. The proposed framework employs Multilingual BERT (mBERT) to obtain context-sensitive representations and capture linguistic relationships across the two languages. These representations are subsequently processed through a Stacked LSTM network to learn deeper sequential patterns within the text, while the spatial attention mechanism assigns greater importance to words and tokens that contribute strongly to the expressed sentiment. To address the class imbalance observed in the manually collected Urdu dataset containing 16,000 tweets, SMOTE (Synthetic Minority Over-sampling Technique) is applied to generate additional minority-class samples. Experimental evaluation demonstrates that the proposed approach achieves 93.33% accuracy on the benchmark English dataset and 96.48% accuracy on the Urdu dataset after SMOTE-based balancing. The model also demonstrates superior performance compared with conventional and deep learning baselines, including LSTM, CNN, SVM, Naive Bayes, Gaussian Naive Bayes (GNB), and K-Nearest Neighbors (KNN). In addition to quantitative evaluation, the study provides a comprehensive visual interpretation of the experimental findings using grouped bar charts, radar plots, and other comparative visualizations.  The results highlight the potential of multilingual deep learning and social media sentiment analysis as effective tools for supporting dengue-related public health monitoring.

Keywords: Dengue; Sentiment analysis; mBERT; SSA-LSTM; Spatial attention; SMOTE; Urdu NLP; Public health surveillance