Detecting Toxicity in a Diverse Online Conversation Using Reinforcement Learning

Abstract

In today's world, we have many online social media sites like Twitter, Facebook, Reddit, CNN, etc. where people actively participate in conversations and post comments about published articles, videos, news, and other online content. These comments by users may be toxic. The threat of abuse and harassment online means that many people stop expressing themselves and give up on seeking different opinions, so there is a need to protect voices in conversations. This theses aims to implement a self-learning model using reinforcement learning methods to detect toxicity in an online conversation. We have designed and implemented the model in the following phases: pre-processing of data, designing the scope of the problem in reinforcement learning, detection of toxicity, and evaluation with comparison to a baseline. We show in our results that the proposed model gets competitive results in terms of F1 score and accuracy when compared to the baseline models, but has computational advantages.