A Comprehensive Dataset and Deep Learning Approach for Misinformation Detection on Social Media in Bangladesh

Abstract

In an effort to address the growing issue of misinformation on social media, particularly in the context of the Covid-19 pandemic, we have diligently developed a comprehensive dataset on Bangla misinformation. This dataset was scraped from FactWatch, a leading fact-checking organization in Bangladesh, and annotated with fact ratings. It includes a meticulously curated collection of 1014 fact-checked reports spanning from October 4, 2021, to May 25, 2023. These reports encompass a diverse array of summaries, categories, and reliable correctness labels, providing samples of the original fake news content along with investigative descriptions of the fact-checking processes employed. The dataset represents a significant contribution to Bangladesh's participation in the global effort to combat fake news and serves as a crucial resource for ongoing research in misinformation studies, natural language processing, and automated fact-checking, particularly for content in the Bengali language. Addressing the issue of misinformation within the under-researched Bangla language context, our study also leveraged this dataset for deep learning analysis, employing advanced techniques such as Long Short-Term Memory (LSTM) networks and Bidirectional Encoder Representations from Transformers (BERT) with a Bangla base model. The BERT model, with its robust Transformer architecture, excelled in linguistic analysis, achieving an accuracy of 98.77%, while the LSTM model, adept at handling sequential data, recorded an accuracy of 88.92%. The Bangla BERT base model demonstrated exceptional performance in precision, recall, and F1-score, marking a substantial advancement in misinformation detection for the Bangla language.