Enhancing Fake News Analysis in Bangla: A Hybrid Model with LDA, and LIME for Interpretable Classification

Abstract

The proliferation of fake news presents a critical challenge in ensuring information authenticity, especially in the context of Bangla text. This study addresses the issue by developing a robust pipeline for detecting and categorizing fake news using advanced machine learning (ML) and natural language processing (NLP) techniques. A dataset comprising 5,000 fake news articles was prepared, collected from 20 different newspapers in Bangladesh between 2020 and 2024. Topic modeling was performed using Latent Dirichlet Allocation (LDA), whose results in topic categorization (World, Business, Entertainment, Bangladesh, and Sports) were significantly better compared to other models. The proposed model, consisting of Bidirectional Encoder Representations from Transformers (BERT), Support Vector Machine (SVM), and Bidirectional Long Short-Term Memory (BiLSTM), outperforms the traditional models in all evaluation measures, yielding a mean accuracy of 97%. Moreover, the integration of Local Interpretable Model-agnostic Explanations (LIME) adds interpretability to the model by explaining individual predictions, thereby enhancing transparency. This holistic approach sets a new benchmark for Bangla fake news detection with high accuracy, interpretability, and reliability, opening avenues for further research in combating misinformation in Bangla.