Detection of Virulent Messages Written in Code-Mixed Hindi-English Language

Abstract

Social media as a medium of communication has opened countless possibilities for individuals around the globe to express their opinions on various topics ranging from politics to entertainment, thus reducing the gap in the accessibility of information. However, this has also led to social media becoming a hub for spreading hateful messages. These virulent messages spread online result in the rise of cases of cyberbullying, leaving an enduring impact upon the victims. Manually classifying and reporting these messages is a challenging task and has therefore created an interest in the research community to incorporate machine learning-based techniques to ease up the hand-operated process. The study's goal is to play a small auxiliary role in the prevention and ultimate eradication of cyberbullying. The proposed work in this research primarily focuses on the users from the Indian subcontinent by experimenting with Code-Mixed Hindi-English language. This research achieves this by performing the task of developing a Code-Mixed Hindi-English hate speech dataset containing 4014 tweets, out of which 2000 tweets contain hate, using the publically available tweets from the Twitter platform and classified on the basis of a set of guidelines. These languages are employed in the training of the models to create a more reliable hate speech prediction system for the targeted audience. The use of pre-trained multilingual models are explored by fine-tuning on the collected dataset. This research discovered an accuracy of 80.6 on the fine-tuned BERT Multilingual Base model when trained and tested on the combination of all the three languages and an accuracy of 82.2 on theXLM-RoBERTa model upon the same dataset. The goal of the study is to play a small auxiliary role in the prevention and ultimate eradication of cyberbullying.