Using Sarcasm to Improve Cyberbullying Detection

AbstractCyberbullying has become more prevalent over time, especially towards minority groups, and online human moderators cannot detect cyberbullying content efficiently. Prior work has addressed this problem by detecting cyberbullying with deep learning approaches. In this project, we compare several BERT-based benchmark methods for cyberbullying detection and do a failure analysis to see where the model fails to correctly identify cyberbullying. We find that many falsely classified texts are sarcastic, so we propose a method to mitigate the false classifications by incorporating neural network-based sarcasm detection. We define a simple multilayer perceptron (MLP) that incorpo- rates sarcasm detection in the final cyberbully classifications and demonstrate improvement over benchmark methods.