Classification of Politically Aggressive Bengali Texts Using Transformers

The widespread use of the Internet and social media has enabled unrestricted communication among individuals with diverse backgrounds. However, interactions between individuals from opposing political groups can lead to physical violence. In this study, we introduce a new Bengali dataset on political aggression, classifying texts into aggressive or non-aggressive. Aggressive texts are further categorized into inciting assassination, rebellion, or communal violence. By utilizing advanced BERT transformer models such as multilingual-BERT, distil-BERT, Bangla-BERT, and XLM-RoBERTa, our approach achieved an impressive 0.87 F1 score at the first and the second level using Bangla-BERT and Multilingual-BERT respectively. This research contributes to our comprehension of political aggression in online discourse, offering valuable insights for mitigating its real-world consequences.