AI-driven detection of hate speech on social media: a case study in the French language

Abstract In today’s digital age, the rise of online hate speech–fueled by user anonymity and widespread use of platforms like X–has become a serious issue. It is a challenge for both social media companies and authorities in their efforts to maintain a safe and respectful online environment. Hate speech (HS) is mostly seen as an intentional act of aggressive behavior towards anyone based on their personal traits. The main corpus of research work done in this area concerns the main spoken languages in the world, such as English, Arabic, or Hindi. However, in the circle of European languages, HS in the French language needs more development. Throughout this study, we present our contribution to the area by compiling and preprocessing a custom dataset by merging and cleaning existing French HS corpora. We develop binary-class Machine Learning (SVM, Random Forest, Naive Bayes, Logistic Regression), Deep Learning (LSTM, Bi-LSTM, GRU), and Transformers models (CamemBert, DistilCamemBert, DeHateBert). Experimental results show that DistilCamemBert significantly outperforms other models, achieving an F1-score of 80%. Then we compare our results to the existing state-of-the-art online. After that, we try to take a step further in the research with the implementation of interpretability studies with explainable AI and bias mitigation.