Evaluating Hate Speech Detection in Roman Urdu Using BERT and Various Learning Approaches: A Comparative Study

Social media and technology have made it easier for people to engage in hate speech violations, which makes it challenging to filter out such content due to the amount of data and possible psychological impacts. Hate speech that targets particular groups is frequently unlawful and may be disseminated on websites like Facebook, where it can influence others. The development of automated systems for detecting hate speech in European languages is a trend, and many datasets are available, while creating digital material in Urdu presents unique difficulties. It is challenging to train a model to correctly detect hate speech in Urdu. To tackle this problem, machine learning models are compared with transformer-based models to identify hate speech in Roman Urdu. A publicly available benchmark dataset (HS-RU-20) consisting of Hostile and neutral comments from social media is used. To vectorize the dataset, count vectorizer is employ to convert posts to a vector representation and evaluate various machine learning and transformer-based models. Results show that the BERT model outperformed the other classification models on neutral/hostile classes and had the highest F1-Score of $\mathbf{9 1 \%}$. On the other hand, it achieved an F1-score of 87% for the Hate/Offensive classification task.