HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture
HACS-TL: Cross-Script Transfer Learning for Hausa Ajami Hate Speech Detection Using Transformer-Based Architecture
Abdulkadir Shehu Bichi,Muqaddar Ali,Prashant Sharma,Ismail Dauda Abubakar
TLDR
HACS-TL, a novel transformer-based architecture for detecting hate speech in Hausa Ajami, is proposed, which uses a linguistically motivated Latin-to-Hausa Ajami converter, cross-script attention, and dialect-specific pooling to establish a baseline for the computational processing of Ajami Hausa.
Abstract
Hausa is one of the most widely spoken languages in West Africa and uses two scripts: the Boko Latin and the Ajami Arabic-derived script. Although Ajami is culturally important, it is still very underrepresented in natural language processing. In this paper, HACS-TL, a novel transformer-based architecture for detecting hate speech in Hausa Ajami, is proposed. Our model uses a linguistically motivated Latin-to-Hausa Ajami converter, cross-script attention, and dialect-specific pooling. After conducting extensive 2-fold cross-validation on 2,000 samples, we found that HACS-TL outperformed baseline models (mBERT (73.70%), XLM-RoBERTa (67.20%), and AraBERT (56.96%)) with a macro F1 score of 75.33%. Detailed error analyses and orthographic stress tests demonstrate the robustness of this approach. The proposed Hausa Ajami conversion system achieved 63.57% character-level similarity with 51% correct conversions, thereby establishing a baseline for the computational processing of Ajami Hausa.