A Study of Text Representations for Hate Speech Detection

The pervasiveness of the Internet and social media have enabled rapid anonymous spread Hate Speech content on microblogging platforms such as Twitter. Current EU US legislation against hateful language, in conjunction with large amount data produced these has led to automatic tools being a necessary component detection task pipeline. In this study, we examine performance several, diverse text representation techniques paired multiple classification algorithms, abusive language discrimination task. We perform an experimental evaluation binary multiclass datasets, significance testing. Our results show that simple hate-keyword frequency features (BoW) work best, followed by pre-trained word embeddings (GLoVe) well N-gram graphs (NGGs): graph-based which proved produce efficient, very low-dimensional but rich for A combination representations Logistic Regression or 3-layer neural network classifiers achieved best performance, terms micro macro F-measure.