Identifying Political Hate Speech using Transformer-based Approach

In recent times, Bangladesh has experienced significant political changes, sparking debates and discussions across various platforms, particularly on social media. As political opinions continue to proliferate, the rise of political hate speech targeting political groups and leaders has become increasingly evident. Such hate speech often includes undemocratic language or incitements to violence, which can destabilize social harmony. This study focuses on detecting political hate speech in Bangla text, providing valuable insights for social media moderation and public sentiment analysis. It highlights the challenges of hate speech detection in low-resource languages (LRLs) like Bangla, particularly in the context of recent political unrest. We employed a range of machine learning (ML), deep learning (DL), and transformer-based models, achieving significant results. Notably, a stacking approach combining Logistic Regression, Random Forest, Decision Tree, and SVM yielded an impressive accuracy of 91.54%. Additionally, other stacking configurations and standalone models, such as Random Forest and CNN, achieved accuracies of 90.99% and 90.54%, respectively. Most of the models demonstrated strong performance; however, Bangla-BERT surpassed all previous approaches by achieving the highest accuracy of 92.41%.