IBFND: An Improved Dataset for Bangla Fake News Detection and Comparative Analysis of Performance of Baseline Models

Abundant internet connectivity, improvements in bandwidth and access to communication channels have led to a culture of free-flowing information. That almost unconstrained dissemination of information helped people develop a habit of consuming news whenever desired. Nevertheless, individuals must be made aware of the nature of the information they consume. People need assistance to determine whether the news is authentic or fake. Consequently, the detection of fake news has grown in popularity. In recent times, significant progress has been made in identifying fake news in foreign languages. However, more work needs to be done in the Bangla. Initially, we used the BanFakeNews dataset, which has fifty thousand data. However, the dataset is skewed toward authentic news. The dataset's disparity between authentic and fake news makes the models' accuracy sometimes deceptive and unreliable. So, additional resources are still required for the Bangla language. In this work, we have introduced an Improved Bangla Fake News Dataset (IBFND). In this dataset, we have tackled the problem of an imbalanced dataset. We have used web scraping to increase the number of fake news. In addition, we have resorted to using Google Translate for the purpose of translating a few fake English news. Then we have implemented the linear SVC, multinomial NB, XGBoost, and BERT models on IBFND. When we ran all of our models on the BanFakeNews dataset, the highest F1-score came to 81%. Using the IBFND, we have achieved the maximum possible F1-score of 97% with the BERT deep neural network model.