Hate speech and offensive language detection using naïve bayes, SVM and BERT

Abstract

Nowadays, with the rise in the number of smartphones and laptops around the world, social media has become one of the most popular and an important sphere in everyone’s lives. It has become one of the most used platforms for communication and dissemination of information throughout the world. But sometimes, along with this content, there also exists hateful and offensive content in these platforms, which is unwanted. Hate speech can be defined as any form of speech that expresses prejudice or bias against a particular group of people on the basis of race, religion, caste, creed etc. This also includes online bullying, cyber bullying, profanity etc. Offensive speech can also be defined as speech that makes the listener feel annoyed, upset, angry or experience any other negative emotions. In this paper, an attempt has been made to classify 2 datasets – one dataset has been used to classify text as containing hate-or-offensive-text or not containing hate-or-offensive-text, while the other dataset has been used to classify text as having hate text or offensive text or having neither hate-or-offensive-text. The ML models used are ensemble models using only BERT+SVM and NB+BERT+SVM. For the dataset having the 3 classifications, the accuracy, f1, precision and recall scores are respectively 87.65, 85.31, 85.23, 87.65 while for the dataset having the two classifications, the scores are respectively 70.53, 68.16,69.37 and 70.53 respectively.