Roman Urdu Hate Speech Detection Using Transformer-Based Model for Cyber Security Applications.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Social media platforms facilitate global communication but are exploited for hate speech dissemination.
• Hate speech online can escalate to real-world harm, including hate crimes and cyber violence.
• Effective hate speech detection requires context-aware mechanisms for real-time intervention.

Purpose of the Study:

• To develop a robust model for detecting Roman Urdu hate speech.
• To evaluate the efficacy of transformer-based models in capturing contextual nuances of hate speech.
• To establish a baseline for Roman Urdu hate speech detection using advanced NLP techniques.

Main Methods:

• Employed a transformer-based model for Roman Urdu hate speech classification.
• Developed and trained the first Roman Urdu pre-trained BERT model (BERT-RU) from scratch on a large dataset (173,714 messages).
• Compared performance against baseline models (LSTM, BiLSTM, BiLSTM + Attention, CNN) and transfer learning approaches.

Main Results:

• The transformer-based model achieved superior performance, with accuracy (96.70%), precision (97.25%), recall (96.74%), and F-measure (97.89%).
• The model demonstrated robust generalization capabilities on a cross-domain dataset.
• Outperformed traditional machine learning, deep learning, and pre-trained transformer models.

Conclusions:

• Transformer-based models are highly effective for Roman Urdu hate speech detection due to their contextual understanding.
• The developed BERT-RU model and the transformer approach offer significant advancements in combating online hate speech.
• The findings highlight the importance of context-aware NLP for maintaining online safety and social harmony.