Automtic detection of hate speech on social media for the algerien dialect

Abstract

Hate speech on social media platforms can lead to cyber-crime, affecting social life on a personal and national level. Hateful speech spread through social media has the potential to cause personal harm and suffering, as well as social tension. Social media platforms, on the other hand, are unable to regulate all of the contents that users post. As a result, there is demand for automatic detection of hate speech. This need is increased when the posts are written in complex languages, such as Arabic, which is known with its challenges, difficulties and lack of resources. In this work we focused on the Algerian dialects. We release a manually annotated multi-label Algerian hate speech detection dataset. Our dataset contains 14k comments extracted from Facebook, YouTube and Twitter. To confirm our annotated dataset’s practical utility, we used it to train and evaluate several classification models started with the machine leaning classification models, followed by the basic deep learning models, then we test several BERT based models built by either adding DL layers or by composing ensemble learning models constituted of BERT model and other DL models trained individually and finally reported the obtained results, as a best results we got 74% accuracy and 76% F1-Score with ensemble learning (Bert with GRU,LSTM and both of GRU and LSTM). Though the dataset is mainly designed for hate speech detection, it can also be used for cyberbullying detection, and offensive and abusive language detection