Welcome to the upgraded BRAC University Institutional Repository. We are currently organizing collections after a recent system upgrade. Homepage category counters may temporarily show lower numbers while syncing, but over 27,000 repository items remain safe and accessible. Please use the search bar to find theses, scholarly outputs, and institutional documents.

Interpretable Bangla fake news classification using BERT and traditional machine learning approaches

Citation

Abstract

Fake news is a type of content that is inaccurate or misleading and it is usually published with the intention of damaging a person or organization’s reputation. It has recently grown significantly in the online forum and on social media platform like Facebook, Reddit, Twitter etc. Because of its falsified statements, people are often persuaded by false news, which has serious consequences in the real world. As a result, there is a growing interest in the field of fake news identification, even though the majority of fake news identification studies are for English language whereas just few of them are for Bangla language. In our study, we come up with a BERT-based system that uses Stratified K-fold cross validation that can achieve 98.45% test accuracy, whereas only the Random Forest can achieve 86.83% accuracy among all the traditional machine learning models. Furthermore, we used Local Interpretable Model-Agnostic Explanations to provide explainability to our system. In this research, we have used the existing BanFakeNews dataset to identify Bangla Fake News. The primary focus of this paper is to develop a model that can recognize fake news in natural language processing so that the developed model can decrease the time it takes individuals to extract fake news from social media.

Description

Cataloged from PDF version of thesis.
Includes bibliographical references (pages 40-43).
This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science and Engineering, 2022.

Publisher Link

Type

Thesis