Fake News Detection in Bengali Language using Transfer Learning Approach

Abstract

The rapid proliferation of fake news poses grave consequences for civil discourse, political environments, and social cohesion. From public elections to mob vio lence, fake news has been leveraged to achieve personal and political gain. The influence of fake news in this era of information is undeniable. Misinformation can cause mass disruption, and we need a way to stop that from happening. This study experiments with the widely used multilingual pre-trained transformers XLM-Roberta and Multilingual-BERT, along with Bangla-BERT. It also explores the impact of stemming in Bengali and demonstrates the effectiveness of combin ing deep neural network (DNN) layers with pre-trained transformers. A major setback faced in this research was the lack of a well-balanced dataset, which led to inconsistent performance from the models. We undersampled two datasets from the original one[1], one with the ratio fake:authentic = 1:1, another with fake:authentic = 1:3. We were able to achieve 0.95 precision and a 0.92 F1 score in a heavily undersampled but well-balanced dataset derived from the original one. XLM-Roberta and Bangla-BERT based models achieved recall scores of 0.94 and 0.93 respectively on the dataset where ratio of fake:authentic is 1:1. Overall, the models trained on the 1:1 dataset delivered consistent scores across all the metrics, which emphasizes the importance of collecting more fake news data for future research. The best model, based on Bangla-BERT, achieved an accuracy of 96.2% which sets a new benchmark accuracy for transformer based models in fake news detection in Bengali.