CleanSpeak: Energy-Efficient Pipeline for Detoxifying Online Comments

Up to 40% of American social media users encounter online harassment, with toxic interactions leading to serious mental health issues and social withdrawal. This emphasizes the need for tools and methods that maintain the original intent of messages while filtering out the toxic content. Addressing this challenge, we present an energy-efficient, machine learning pipeline for the detoxification of online comments, which we call CleanSpeak. Initially, our pipeline classifies comments into three categories: non-toxic, less toxic, and more toxic. Comments in the latter two classes are then detoxified using fine tuned Llama models. CleanSpeak achieves a similarity score of 0.65 and joint score of 0.46 for less toxic comments, surpassing the previous state-of-the-art similarity score of 0.48 and joint score of 0.31. For more toxic comments, it reaches similarity score of 0.55 and joint score of 0.36, exceeding the prior benchmark of 0.39 and 0.19. Also, our pipeline operates using a small fraction of GPT-4’s energy consumption.