Bangla Protest News Categorization Using Deep Learning

Protests are a vital form of public expression, often reflecting collective dissent towards specific policies or injustices. In recent years, the significance of protests as a medium for public opinion has escalated, necessitating effective categorization for informed decision-making by policymakers, law enforcement, researchers, and journalists. This paper presents the first study on categorizing Bangla protest-related news articles, addressing the absence of a publicly available annotated dataset in this domain. We collected and annotated a dataset of 8,039 protest-related news articles from prominent Bangla newspapers, such as Prothom Alo, Samakal, Daily Star Bangla, and Ittefaq. We employed a multifaceted approach, leveraging attention mechanisms with deep learning and transformer-based models to categorize protest news into eight predefined protest categories. The fine-tuned multilingual BERT model showcased superior efficacy, achieving an accuracy of 91.58% and an F1 score of 91.43%. This study fills a critical gap in protest categorization research and provides a comprehensive dataset and methodological approach for future exploration. The findings have significant implications for understanding the dynamics of social movements and public protests, offering valuable insights for academics and practitioners alike. This research paves the way for further studies in categorizing protest-related news, particularly in underrepresented languages like Bangla.