Automated Cyberbullying Detection in Social Media Using an SVM Activated Stacked Convolution LSTM Network

Abstract

Cyberbullying is becoming a huge problem on social media platforms. New statistics shows that more than a fourth of Norwegiankids report that they have been cyberbullied once or more duringthe last year. In the most recent years, it has become popularto utilize Neural Networks in order to automate the detection ofcyberbullying. These Neural Networks are often based on using Long-Short-Term-Memory layers solely or in combination withother types of layers. In this thesis we present a new Neural Networkdesign that can be used to detect traces of cyberbullying intextual media. The design is based on existing designs that combinesthe power of Convolutional layers with Long-Short-Term-Memorylayers. In addition, our design features the usage of stacked corelayers, which our research shows to increases the performance ofthe Neural Network. The design also features a new kind of activationmechanism, which is referred to as "Support-Vector-Machinelike activation". The "SupportVector-Machine like activation" isachieved by applying L2 weight regularization and utilizing a linearactivation function in the activation layer together with using aHinge loss function. Our experiments show that both the stackingof the layers and the "Support-Vector-Machine like activation"increasesthe performance of the Neural Network over traditionalState-Of-The-Art designs.