Abstract

The rapid growth of social media as a platform for communication and information sharing has raised concerns about its negative impact on social cohesion and peace. Harmful online content can fuel intergroup hatred, violence, mass killings, and deepen social and political polarization that promote prejudice and hostility. Therefore, detecting and mitigating harmful content like hate, offensive speech, and harassment is critically important. Amharic, Ethiopia’s official working language, is widely spoken across a country rich in diverse religions, ethnicities, and cultures. However, research on harmful content detection remains limited due to scarce linguistic resources, small datasets, and limited adoption of advanced technologies. This study aims to improve Amharic mal-information detection through robust multi-class classification tasks. A dataset of 13,683 Amharic texts was used to train and evaluate on machine learning models such as (RF, SVM, LR, NB), as well as DL models such as (CNN, LSTM, BiLSTM, CNN-LSTM, RNN, and ensemble models). Feature engineering, i.e., Bag-of-Words, TF-IDF, GloVe, FastText, and Word2Vec, was applied. The results show that CNN, stacking ensemble, and RF achieved 94% accuracy, followed by BiLSTM, SVM, and LSTM. Future research should focus on scalable, well-annotated, multimodal, and explainable AI to address the dynamic nature of social media.

Keywords: Mal-information, Hate speech, Social media, Offensive speech, Machine learning, Deep learning, Amharic