Fighting Cyberbullying: An Analysis of Algorithms Used to Detect Harassing Text Found on YouTube

Cyberbullying is a form of harassment that occurs through online communication with the intention of causing emotional distress to the intended target(s). Given the increase in cyberbullying, our goal is to develop a machine learning classification schema to minimize incidents specifically involving text extracted from image memes. To provide a current corpus for classification of the text that can be found in image memes, we collected a corpus containing approximately 19,000 text comments extracted from YouTube. We report on the efficacy of three machine learning classifiers, naive Bayes, Support Vector Machine, and a convolutional neural network applied to a YouTube dataset, and compare the results to an existing Formspring dataset. Additionally, we investigate algorithms for detecting cyberbullying in topic-based subgroups within the YouTube corpus.