Bengali social media comments analysis to prevent cyberbullying

Abstract

One of the significant wonders of modern technology is social networking platforms. This gives a major acceleration to the field of communication. With such good advantages of this technology this has a darker side too. Nowadays people are often seen trolling each other online. There is also a new roasting culture in online platforms which arrived a few years ago, where people are often downgrading other people without any moral consideration. Cyberbullying nowadays has become a trend. People are now just waiting for a new topic to arrive in social media so that they can make fun of each other. So we are doing research on this field so that we can make an automated system which would be able to detect cyberbullying instances on social platforms using textual analysis with the help of machine learning and artificial intelligence. We are using social media comments in Bangla language from Facebook for this research along side we are also taking a dataset from kaggle. Our personal Bangla comments dataset contains 5053 Bengali comments. These comments are initially categorized as bully or non-bully. And the bully comments are further categorized as Sexual, Religious, Threat, and Troll and Political comments. The kaggle dataset also contains Bengali social media comments categorized as Neutral, Political, Sexual, troll and threat. We have run experiments on both the dataset individually and combining them together. We have used both machine learning and deep learning architecture for our experiments. We have achieved highest accuracy of 88% on binary class classification on the kaggle dataset. And highest 65% on multiclass classification on the kaggle dataset as well.