Consider the phrase “ ki11 yrslef a$$hole” an example of hate speech , token based models/systems does not recognise this as a hate speech so features based on character level n-gram is in need.