Violent Views Detection in Urdu Tweets

Abstract

The widespread use of social media has led to substantial increase in the global connectivity. Consequently, the content shared on social media has the potential to become viral in a short span of time. While some content is desired to become viral, there is a high risk that the inappropriate messages can also become viral which could be disastrous for the society. For instance, spreading violent views may lead to riots and unrest in the society. Therefore, it is desired to detect violent views to ensure stopping them from spreading. To that end, this study has scrapped Twitter to develop and publicly release the first-ever Violent Views Detection corpus for Urdu (VVD-21). The corpus is composed of 3297 Urdu tweets which are manually classified into Violent and Non-Violent views. Furthermore, experiments are performed using six traditional and two deep learning techniques to evaluate their effectiveness of these techniques for automatically detecting violent views in Urdu text. The results of the experiments show that Logistic Regression is the most effective technique as it achieved the highest F1 score of 0.881.