The COVID-19 pandemic triggered an unprecedented wave of misinformation on social media, particularly in under-resourced languages such as Bangla. To address this, we present a multi-task BanglaBERT framework for joint sentiment (positive, negative, neutral) and truthfulness (real vs. fake) classification of COVID-19-related social media posts. We curatea dataset of 35,526 Bangla posts annotated for both tasks, creating one of the first large-scale dual-labeled resources in the language. Our model employs a dual-head architecture with tailored loss functions to mitigate class imbalance, achieving 75% accuracy (macro F1: 0.70) for sentiment classification and 88% accuracy (macro F1: 0.85) for truthfulness classification. Results show strong performance on polarized sentiment and real news detection, though neutral sentiment remains challenging due to semantic ambiguity, and sensational real news is sometimes misclassified as fake. To bridge research and application, we provide a Gradio interface for real-time inference. This work establishes a new benchmark for multi-task NLP in Bangla, demonstrating the feasibility of jointly modeling emotional framing and misinformation detection in a low-resource setting. The proposed framework has implications for public health monitoring and combating misinformation during crises.

Disclaimer: This is the author’s version of the manuscript submitted to the International Journal of Information Technology and Computer Science (IJITCS).