Multilingual Toxic Comment Classification

Abstract

With the growth of the Internet and data collection in the last twenty years, we have seen a rise with Internet Relay Chats (IRC) as well, especially in 2020 due to COVID. Now more than ever, it is important that we stay connected despite being physically separated and to make sure communications systems are as inclusive as ever. Our goal in this project is to categorize toxic comments used in IRC chat rooms and forum pages via a machine learning model in order to keep communications safe and welcoming to the public. We define toxicity as any message that is rude, disrespectful, or otherwise likely to make someone leave a discussion. If these toxic contributions can be identified, we could have a safer, more collaborative internet. As a result of our models, we will be able to identify a dataset of toxic comments with at least a 70% success rate. Additionally, we want to be able to successfully identify both new and previously existing comments as toxic or non-toxic in multiple languages that others have previously label incorrectly. Using a dataset provided by Kaggle, we demonstrate how we are able to detect if a Wikipedia talk page comment is toxic (scored as a 1) or non-toxic (scored as a 0) by using statistical methods, data analysis, and various data science approaches.