Q-Bully: A Reinforcement Learning based Cyberbullying Detection Framework

With the ever-increasing involvement of people into social media and online gaming, Cyberbullying has become a serious issue affecting almost all parts of the demographic. Cyberbullying can cause severe mental and emotional impacts on a person, especially on minors; hence, there is a requirement of having intelligent automated systems to detect questionable content present on social media platforms and remove it. In this paper, we introduce our novel algorithm Q-Bully which can automatically detect cyberbullying on various social media and online gaming platforms using Reinforcement Learning along with Natural Language Processing techniques. Previously the techniques used to detect cyberbullying have a good accuracy related to the text they have been trained on and do not incorporate new word patterns without complete retraining of model. In this paper, we incorporated the use of Reinforcement Learning and have conducted an experimental study in which we feed the messages and posts of bullies as well as victims to a Reinforcement Learning Agent for classification. We compare our model with the other baseline models on based on F1 scores (0.86 a benchmark dataset of 16K annotated tweets) and are able to infer that our model outperforms other state-of-the-art models when the dataset is highly dynamic and populated with words which are deliberately misspelled to trick the conventional detection systems.