Algorithmically supported moderation in children's online communities

Abstract

The moderation of harassment and cyberbullying on online platforms has become a heavily publicized issue in the past few years. Popular websites such as Twitter, Facebook, and YouTube employ human moderators to moderate user-generated con- tent. In this thesis, we propose an automated approach to the moderation of online conversational text authored by children on the Scratch website, a drag-and-drop programming interface and online community. We develop a corpus of children's comments annotated for inappropriate material, the first of its kind. To produce the corpus of data, we introduce a comment moderation website that allows for the review and label of comments. The web-tool acts as a data-pipeline, designed to keep the machine learning models up to date with new forms of inappropriate content and to reduce the need for maintaining a blacklist of profane words. Finally, we apply natural language processing and machine learning techniques towards detecting inappropriate content from the Scratch website, achieving an F1-score of 73%.