ENHANCED DEBERTA-BASED HADITH UNDERSTANDING FOR ISLAMIC ASSISTANCE

Abstract

The texts of the hadith consist of the actions, sayings, and approvals of the Prophet Muhammad (peace be upon him). They comprise the second primary source of Islamic law and ethics; therefore, their significance is foundational. A major challenge for automated Question Answering (QA) systems is the multiple theologically and narratively driven challenges hadiths contain. These challenges most often result in missing answers and semantic drifting. The demand for automated, reliable QA systems has sociologically, and in the context of the projected 2.8 billion Muslims by 2050, a theologically driven demand, increased. Previous approaches have problems with the language of the domain, the specific problem of the domain, and a loss of computational efficiency. We empirically analyze the combination of the custom dataset of 42,591 QA pairs in the SQuAD (Stanford QA Dataset) format, the pre-trained transformer model DeBERTa, and other versions of the BERT model. The results for the fine-tuned DeBERTa model achieved the highest Exact Match (EM) score of 89.77%, followed by BERT (75.87%), RoBERTa (84.32%), and ALBERT (81.15%). These results suggest that DeBERTa is one of the more efficient models for QA systems in hadith and other similar domains, compared to other models.