Threatening Expression and Target Identification in Under-Resource Languages Using NLP Techniques

In recent decades, hate speech on social media platforms has been on the rise. It is highly desired to control this kind of material because it initiates unrest and harms to the society. Literature describes several forms of the hate speech and it is quite challenging to differentiate between these forms and to design an automated detection system, especially for under-resource languages. In this study, we propose a robust framework for threatening expressions and its target identification in Urdu (Nastaliq style) language. The proposed methodology presents each step in detail like data collection & annotation, cleaning & pre-processing step, and fine-tuning of Robustly Optimized Bidirectional Encoder Representations from Transformer (Urdu-RoBERTa) with grid search technique for hyper-parameters optimization. The study exploits the strength of a pre-trained Urdu-RoBERTa as a transfer learning technique with grid search fine-tuning. The proposed framework is compared with state-of-the art baseline and ten comparable models and it outperformed all for both tasks (threatening expression and target identification). Furthermore, the proposed framework obtained benchmark performance and improved the f1-score with substantial margin.