Multi-Aspect Hate Speech Analysis for Roman Urdu Text

Abstract

We live in the age of technology where a large amount of information is produced daily on social media sites as it becomes a source for expressing their opinions and sharing ideas with other people, it also becomes a place for abusive language, personal attacks, and hateful comments. Determining the nature of the suspension is difficult and time-consuming. Automating the process of hate speech analysis in online conversations is the best way to ensure user security and improve online conversations. In this study, we have produced our dataset for Roman Urdu containing more than 3k comments which were annotated by NLP experts with the following aspects: Hostility, directness, target and group. The dataset is trained using various deep learning & machine learning algorithms for figuring out which model is the best at classifying multi-aspect hate speech. The results showed that logistic regression and bi-LSTM are the best algorithm in determining the toxicity of Roman Urdu text.