Hate Speech Detection in Roman Urdu using Machine Learning Techniques
Sarah Nasir, Ayesha Seerat, Muhammad Wasim · 2024
In recent years, the spread of hate speech on social media has been a major source of discomfort. Hate speech may be extremely harmful, resulting in violence, discrimination, and even genocide. The problem is particularly acute in the case of the Roman Urdu language, where the use of hate speech is widespread. To address this issue, there’s a growing demand for effective methods to identify hate speech on social media. While many researchers have focused on European languages, few have worked on South Asian languages, such as Roman Urdu, which are widely used in the subcontinent. In this study, we contribute by developing a methodology to detect hate speech in Roman Urdu at two levels. First, we classify the social media content into neutral and hostile categories. Secondly, we classify the hostile content as offensive and hate speech. We use a benchmark corpus (HS-RU-20) to evaluate the proposed methodology and the two-level classification. Furthermore, we analyze the word and character level features along with six supervised learning models (logistic regression, multinomial bias, KNN, random forest, SVM, and convolutional neural network (CNN)). The results show that logistic regression performed better in terms of accuracy, at 81% on neutral-hostile, and outperformed offensive-hate speech classification with an accuracy of 87%.