Intelligent model for target identification from Urdu threatening tweets in low-resource language using deep learning and fine-tuning of language models

The rise of social media sites like Twitter, which is one of the most widely used venues for expressing thoughts, is a result of the development of technology. However, these opinions may sometimes contain threatening text that can disturb other users. As a result, it's critical to spot harmful content on social media. Regrettably, there aren't many methods for this task in low-resource languages like Urdu, as opposed to high-resource languages like English and Dutch. To address this issue, a study proposes an intelligent model for detecting threatening language in Urdu. Machine learning (ML) classifiers such as Extreme Gradient Boosting (XGB), Random Forest (RF), and Bernoulli Naive Bayes (BNB) have been employed. Deep learning (DL) such as Convolutional neural networks (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) have been used. Finally, BERT and RoBERTa have been utilized as transfer learning approaches. FastText has been used as a feature vector for ML and DL classifiers. In comparison, BERT and RoBERTa have been fine-tuned on the benchmark dataset. The employed algorithms have been tested using Accuracy, Precision, Recall, and F1-score as evaluation metrics. The RoBERTa outperforms other models with 83% accuracy, 70% precision, 53% recall, and a 61% F1-score. This study highlights the importance of detecting target-based threatening language on social media and presents an effective approach for achieving this in a low-resource language such as Urdu, which can improve the safety and security of social media platforms.