Developing Hate Speech Detector for Afan Oromo Content on Social Media Platforms Using Deep Neural Networks

Abstract There has been a massive shift toward social media communication in Ethiopia, and a major problem has emerged with online hate speech that negatively affects social cohesion and human dignity. This work is aimed at addressing the need for a reliable and automatic detection system for Afan Oromo, a low-resource language spoken by over 40 million people in East Africa, whose content has largely been under-moderated as the social media ecosystem grows and existing automated moderation remains skewed toward high-resource languages, leading to political and ethnic polarization. The primary motivation behind this investigation stems from the spread of ethnic hatred on platforms like Facebook and X, as well as campaigns of violence and dehumanization in the physical world. The main objective is to fine-tune a deep learning-oriented model for the detection of hate speech information from text. This research utilizes an innovative XLM-RoBERTa model and an Afro-XLMR transformer model. A dataset of 22,000 samples was used, comprising 20,000 benchmark cases and 2,000 specially gathered social media posts from Facebook and X. The main findings show that the XLM-RoBERTa model outperforms over Afro-XLMR transformer model, achieving F1 scores of 95.3% and an accuracy of 95.8%. For underserved linguistic populations, this work provides a scalable solution to improve online safety while setting a standard for Afan Oromo content moderation.