Classification of Abusive Comments Using Various Machine Learning Algorithms

Abstract

In the past decade, we have seen an increasing surge in the popularity of social networking sites, with Twitter and Facebook being some of the most popular ones. These sites allow you to share your expressions and views. However, as of now, there is no particular restriction applied by them to control the kind of content that is being uploaded. These uploaded contents may have obnoxious words, explicit images which may be unsuitable for social platforms. There is no predefined method for restricting unpleasant texts from publishing on social sites. To solve this problem, we propose a method that can aid human moderators as well as work independently. In this approach, logistic regression, multinomial Naïve Bayes, and random forest techniques are used to extract features like term frequency–inverse document frequency features, text features, and frequency features of the comments, respectively, to obtain a weak prediction model. Gradient boosting is applied to this model to obtain the final prediction model. We also applied a neural network using bidirectional long short-term memory and compared the accuracy rate of the two models. We believe that these models can help human moderators on various online platforms to filter out abusive comments.