Hate Speech Detection on Multilingual Twitter Using Convolutional Neural Networks

Accepted: 29 December 2019 Hate speech detection on Twitter is often treated in monolingual (in English generally) ignoring the fact that Twitter is a global platform where everyone expresses himself with his natal language. In this paper, we created a model which, taking benefits of the advantages of neural networks, classifies tweets written in seven different languages (and even those that contains more than one language at the same time) to hate speech or non hate speech. We used Convolutional Neural Networks (CNN) and character level representation. We carried out several experiments in order to adjust the parameters according to our case study. Our best results were (in terms of accuracy) 0.8893 for a dataset containing five languages and 0.8300 for a dataset of seven languages. Our model solves properly the problem of hate speech on Twitter and its results are, compared to the state of the art, more than satisfactory.