AI & Society
A 2025 study found that Bangla-BERT detected political hate speech in Bangla text with 92.41% accuracy
In a 2025 study, researchers tested several machine learning, deep learning, and transformer-based models on the task of identifying political hate speech written in Bangla. A model called Bangla-BERT performed best, with 92.41% accuracy. Accuracy here means the share of examples the model labeled correctly.
The study focused on social media discussions in Bangladesh after a period of political change. The authors described the political hate speech they targeted as undemocratic language or incitement to violence aimed at political groups and leaders.
What did the study set out to do?
The researchers wanted to detect political hate speech in Bangla text. The study highlights challenges of hate speech detection in low-resource languages like Bangla.
How did the models compare?
The next best result came from a stacking approach that combined logistic regression, random forest, decision tree, and support vector machine (SVM) classifiers. It reached 91.54% accuracy. Other configurations and standalone models came close. Random forest reached 90.99% accuracy and a convolutional neural network reached 90.54%.
FAQ
What language did the study focus on?
The study focused on Bangla political text, especially discussions on social media in Bangladesh.
Which model performed best?
Bangla-BERT performed best, with 92.41% accuracy.
Why does the paper say this matters?
The authors point to possible uses in social media moderation and public sentiment analysis.
Limits and what's next
The abstract does not describe limitations or future work.
Source: Pravin More et al., “Identifying Political Hate Speech using Transformer-based Approach,” 2025.