Multiclass Classification for Toxicity Detection in Urdu Comment

This paper addresses the growing issue of toxic language on social media platforms, particularly in the Urdu language, which remains underexplored in existing research. Toxic comments, including hate speech, abusive language, and cyberbullying, have become prevalent on social media, leading to severe consequences such as mental health issues and, in some cases, suicide. While much research has focused on detecting toxicity in English and other major languages, the lack of datasets and benchmarks for Urdu presents a significant gap. With over 170 million speakers globally, Urdu’s complex morphology and syntactic structure make it particularly challenging for machine learning models. In this study, we propose a multi-label, multiclass classification framework for detecting toxic content in Urdu using machine learning techniques. We developed a dataset of user comments collected from platforms such as Facebook, YouTube, and Twitter, focusing on four categories: positive, negative, rude and offensive. The proposed framework integrates models like Support Vector Machines (SVM), Naive Bayes, Random Forest, and Gradient Boosting to classify toxic content. Our results demonstrate the effectiveness of these machine learning models in handling the complexities of Urdu, marking a significant contribution toward the automatic detection of toxicity in this widely spoken language. This research also provides a valuable public benchmark for future studies in this domain.