# Abusive and Threatening Language Detection in Urdu using Boosting based and BERT based models: A Comparative Approach

**Authors:** Mithun Das, Somnath Banerjee, Punyajoy Saha  
**arXiv:** [2111.14830](https://arxiv.org/abs/2111.14830)  
**Date:** 2021-11-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.14830v1/

## Key points

- The paper addresses the concern of online hatred on social media platforms
- Various social media platforms have implemented moderation policies and employed moderators to tackle this issue
- There is a significant gap in detecting abusive language in low resource languages like Hindi and Urdu
- The authors explored machine learning models including XGboost, LGBM, and m-BERT for abusive and threatening content detection in Urdu
- A Transformer model trained on an abusive language dataset in Arabic yielded the best performance
- The model achieved first place for both abusive and threatening content detection with F1 scores of 0.88 and 0.54 respectively
- Effective methods for detecting abusive language, particularly in low resource languages like Urdu, are important in addressing online hatred

## Summary

The paper titled "Abusive and Threatening Language Detection in Urdu using Boosting based and BERT based models: A Comparative Approach" addresses the growing concern of online hatred on social media platforms. To tackle this issue, various social media platforms have implemented moderation policies and employed moderators to identify and take action against posts that violate these policies. Additionally, researchers in the field of abusive language detection have conducted studies to improve the detection of such content. While there has been extensive research on abusive language detection in English, there is a significant gap in detecting abusive language in low resource languages like Hindi and Urdu. The organizers of this task provided a dataset for detecting abusive language and threatening content specifically in Urdu. In their study, the authors explored several machine learning models including XGboost, LGBM, and m-BERT based models for abusive and threatening content detection in Urdu. They observed that a Transformer model trained on an abusive language dataset in Arabic yielded the best performance. The results of their study showed that their model achieved first place for both abusive and threatening content detection with F1 scores of 0.88 and 0.54 respectively. Overall, this paper highlights the importance of addressing online hatred by developing effective methods for detecting abusive language, particularly in low resource languages like Urdu. The authors' comparative approach utilizing different machine learning models demonstrates promising results for identifying abusive and threatening content on social media platforms.