Detection of Harassment on Twitter with Deep Learning Techniques

Abstract

Online harassment is a common issue since the beginning of social networks and it’s still present nowadays, causing serious consequences to victims because their gender, race, sexuality, among others. We have seen efforts to fight these behaviors creating automated systems to detect and report this kind of bad conduct. However, these solutions tend to perform well only on a specific type of data without generalizing well. In this paper, we present a new dataset of harassment detection on Twitter with four classes, presented for the SIMAH competition. Then we apply three different deep learning architectures (CNN, LSTM, and BiGRU) to classify these tweets showing that it is a hard problem to solve especially because of the lack of annotated data within some classes. The results only on the test set reach 46% in f1-score and using all data to train gives 55% using the same metric.