Enhancing Urdu hate speech detection through differential transfer learning and adaptive loss functions

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Hate speech detection is complex, especially in low-resource languages like Urdu.
• Existing research predominantly focuses on high-resource languages, leaving Urdu understudied.

Purpose of the Study:

• To enhance Urdu hate speech detection using novel differential transfer learning and adaptive loss functions.
• To address linguistic and cultural nuances specific to Urdu in hate speech detection.

Main Methods:

• Leveraged pre-trained models from high-resource languages for feature extraction.
• Implemented differential transfer learning to adapt models to Urdu's unique characteristics.
• Developed an adaptive loss function to handle class imbalance and improve sensitivity to hate speech.
• Created and utilized a Nastaliq Urdu dataset of 18,058 YouTube comments.

Main Results:

• Transfer learning methods outperformed conventional machine learning and deep learning techniques, increasing F1 scores from 81% to over 89%.
• The proposed DAmBERT model, incorporating pre-trained embeddings, achieved a weighted F1 score of 91.49%.
• The adaptive loss function improved model sensitivity towards the minority hate speech class.

Conclusions:

• Differential transfer learning combined with adaptive loss functions offers a robust approach for Urdu hate speech detection.
• The DAmBERT model demonstrates significant potential for improving hate speech detection systems in low-resource languages.
• This research highlights the importance of cultural and linguistic adaptation in NLP tasks.