FAD-SA-GRU: ENHANCING HATE SPEECH DETECTION IN ALGERIAN DIALECT THROUGH FEATURE-AUGMENTED SELF-ATTENTION GRU NETWORKS

Automatic hate speech detection is a critical component of modern content moderation, yet it remains difficult for low-resource languages like Algerian Darija. The authors address the challenges posed by the dialect's linguistic complexity—specifically the mixture of Arabic, French, and Arabizi (Latin-script Arabic)—and the lack of standardized orthography, which complicate traditional NLP approaches.

The researchers developed a new hybrid architecture called FAD-SA-GRU. This model employs a multi-embedding fusion strategy that combines three distinct semantic representations: DZ FastText (static), DZ AraVec (static), and DziriBERT (contextual). These fused features are then processed by a Gated Recurrent Unit (GRU) encoder equipped with a self-attention mechanism, which allows the model to weigh the importance of different tokens in the sequence before making a final classification. The model was trained and tested on a manually annotated dataset of 22,193 Algerian Darija social media comments.

The FAD-SA-GRU model consistently outperformed all baseline approaches, including traditional machine learning models (TF-IDF based) and standard Transformer models like multilingual BERT and DziriBERT. The proposed architecture achieved an accuracy of 93.2%, a precision of 93.4%, and an F1-score of 92.1%. The results suggest that combining complementary static and contextual embeddings with attention-based sequence modeling is highly effective for capturing the nuanced, heterogeneous nature of dialectal Arabic text.

This research provides a robust framework for moderating harmful content in under-resourced dialects. By demonstrating that hybrid models can outperform even sophisticated Transformer-based baselines in specific dialectal contexts, the study offers a practical path forward for improving online safety and content moderation in regions where linguistic diversity and non-standard writing systems are the norm.