Semi-Supervised Based Bangla Fake Review Detection: A Comparative Analysis
Nurul Absar, Tanjim Mahmud, Abubokor Hanip, Mohammad Shahadat Hossain · 2025
This study investigates the application of supervised and semi-supervised methods to enhance pre-trained language models for distinguishing between fake and genuine Bengali reviews, using limited annotated data. As fraudulent and deceptive reviews on social media and e-commerce platforms proliferate, accurate detection is essential to prevent consumers from being misled. However, identifying fake news is par-ticularly challenging for low-resource languages like Bengali, where labeled data is scarce. In this work, we present a comparative bengali transformer model with semi-supervised approaches for fake review detection, with a focus on a proposed semi-supervised architecture combining Generative Adversarial Networks (GANs) and the FixMatch algorithm. GANs achieved 84% accuracy, while FixMatch outperformed other approaches with 85%, establishing a new state-of-the-art for Bengali fake review classification. Additionally, the bengalaBert showed supe-rior performance compared to other models. Experiments were conducted on a manually annotated dataset of 6,014 real and fake Bengali food reviews collected from various social media groups. Our proposed methodologies address the challenge of fake review detection in low-resource settings and offer a promising solution for overcoming the limitations posed by insufficient labeled data. This research provides valuable insights for enhancing classification tasks in similar contexts.