Deep Learning for Somali Hate Speech Detection: A Comparative Analysis of Sequential, Attention-Based, and Graph Neural Network Models

Social media platforms facilitate quick communication, information sharing, and the expression of opinions. However, their misuse for hate speech targeting race, religion, and political differences has become an increasing concern, particularly for underrepresented languages like Somali. This study tackles the challenge of detecting hate speech in Somali text by analyzing posts and comments from Facebook. We collected 10,000 cleaned Somali posts and comments from suspicious public pages of organizations and individuals on social media. Vital preprocessing steps, such as data cleaning and tokenization, were performed based on the requirements of the language to obtain a cleaned corpus. The dataset consists of two classes, named hate speech and non-hate speech, defined by experts based on the prepared Somali annotation guidelines. We experimented with state-of-the-art DL algorithms such as optimized LSTM, enhanced BiLSTM, CNN LSTM hybrid, Multi-Head Attention, Hierarchical Attention Networks, and GCNs. The optimal GCN model attained a 94.11% F1-score and 94.33% accuracy. While competitive with established traditional machine learning baselines, our deep learning methodology offers distinct advantages, enhanced generalization capabilities and improved capacity for modeling complex linguistic structures inherent in Somali text. This research contributes the first comprehensive evaluation of DL architectures for Somali hate speech detection, establishes performance benchmarks complementing existing traditional baselines, and provides a methodological framework suitable for extension to other under resourced languages facing similar computational challenges.