Classification of Hadith Authenticity Based on Sanad Using BERT

Hadith authenticity plays an important role for Muslims, as Hadith serves as the second primary source of Islamic law after the Qur'an. A Hadith has two main components: the Sanad, the chain of narrators, and the Matan, the text content. With the increasing accessibility of Hadith through online platforms, opportunities for easy distribution have grown, but so have the spread of fabricated Hadith. The abundance of data available has made Machine Learning (ML) an increasingly common approach for tackling the classification of Hadith authenticity and, more recently, the use of Deep Learning. However, the use of transformer models for Hadith classification has not been fully explored. This study investigates the application of pre-trained Arabic transformer models for classifying Hadith into three classes: Sahih, Hasan, and Da'if, and using only the Sanad. Specific transformer models used are the AraBERT, ARBERT, and QARiB compared to traditional ML models such as Linear Support Vector Classifier (LinearSVC) and Multinomial Naive Bayes (NB). The results show that the performance of the models using only the Sanad is slightly better than using the full text, with the best model being QARiB with a 75.71% F1-score in the 3-class classification setup. This score reflects the complexity of the dataset, and it can be improved by addressing misclassifications, especially between the overlapping Hasan and Da'if classes.