A Hybrid Tri-Encoder model for fake news detection in Bengali with LIME-based explainability

Abstract

The widespread dissemination of false information on digital platforms has become a critical concern, particularly for low-resource languages like Bengali. Addressing this issue, the present study introduces a Hybrid Tri-Encoder model that utilizes three transformer-based encoders DistilBERT, mBERT, and BanglaBERT, to process different components of a news article: its headline, a hybrid-summarized form of the content, and the complete article itself. The summarized version is generated using a two-step hybrid summarization strategy that combines YAKE-based extractive summarization with mT5-based abstractive summarization. This ensures the preservation of key terms along with semantic coherence for better representation. The outputs from the three encoders are combined to create a unified feature representation, which is then passed through a dense classification layer to determine the authenticity of the article. To address the class imbalance in the dataset, we employed a weighted loss function during training. Furthermore, this study incorporates LIME (Local Interpretable Model-agnostic Explanations), a technique that approximates the model's behavior locally, to generate instance-level interpretability for predictions. The experimental results confirm that the proposed method effectively distinguishes between genuine and fake news. The integration of explainability enhances model reliability, offering a transparent and robust solution for fake news detection in Bengali.