Propaganda Detection Techniques for Social Media Texts us-ing Lifelong Learning

Abstract

Most people misuse social media to influence the thoughts of large masses. The use of propaganda mainly targets to play with the emotions of people and communities and thus influence them easily. In today’s society, where social media plays a crucial role in society, it becomes necessary to detect such propaganda in social media texts like tweets, online articles, Facebook posts, and advertisements to protect people from being easily misguided. Our work focuses on detecting the propaganda techniques employed in the social media texts of various domains. To achieve this, we use the approach of Lifelong machine learning. When the model is trained sequentially on different domains, most paradigm ML algorithms undergo catastrophic forgetting of previous domains when the model encounters a new domain. Our model overcomes this drawback by continuously learning the new propaganda domain without catastrophically forgetting the previous domain when a new domain is introduced. We use the Elastic Weight Consolidation(EWC) regularization technique to prevent catastrophic forgetting of previous domains when the model is trained on the new domain and to ensure good performances across all domains.