Detecting Violence Inciting Texts based on Pre-trained Transformers

In response to the alarming surge in communal violence incited through social media texts, this study aims to devise an intelligent solution for the classification of violence-inciting texts in Bengali, categorizing them into three distinct classes: direct violence (DVio), passive violence (PVio), and non-violence (NVio). To address this challenge, this study employs a variety of machine learning (ML), deep learning (DL), and transformer-based methodologies. We assessed many models, including LR, DT, RF, MNB, SVM, CNN, BiLSTM, BiGRU CNN+BiLSTM, CNN+BiGRU, M-BERT, XLM-R, BanglaBERT, and BanglaBERT
base
, via rigorous testing and fine-tuning. The experimental findings show that the proposed ensemble model (BanglaBERT + BanglaBERT
base
) gained the highest macro F1-score of 73% for detecting violence-incited Bengali texts. This outcome highlights the competitiveness and viability of the developed approach, underscoring its potential to combat violence-inciting content on social media platforms.