Toward Robust Offensive Language Detection in Urdu YouTube Discourse

This study addresses the task of detecting offensive language in Urdu text on social media, where harmful and offensive comments can lead to emotional distress. Using a dataset of Urdu comments collected from YouTube news channels, both deep learning (DL) and machine learning (ML) models were employed for offensive language detection. Feature extraction was performed using Term Frequency–Inverse Document Frequency (TF-IDF) and the Count Vectorizer approach to capture unigrams, bigrams, and trigrams. Four ML models—Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Gaussian Naive Bayes (NB)—were evaluated, with SVM achieving comparatively stronger performance among the ML approaches. For the deep learning framework, Bidirectional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN) architectures were implemented using custom Word2Vec and FastText embeddings. The experimental results indicated that the CNN model outperformed the other evaluated approaches, suggesting its effectiveness for detecting offensive language in Urdu text. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i1.1240 Dimensions.Open Alex.