The Application of Deep Learning in Qur’anic Tafsir Retrieval Using SBERT, FAISS and BERT-QA

Abstract

Accurate understanding of the Qur’an requires access to reliable tafsir, yet many classical tafsir resources remain non-digital, making search and retrieval time-consuming. This study presents a semantic-based retrieval system for Tafsir Ibn Kathir, covering 114 entries and 6,236 Verses, using SBERT embeddings and FAISS indexing. The system enables users to perform semantic queries, retrieving relevant passages in response to their questions. Evaluation was conducted using 50 representative queries spanning diverse topics, including Fiqh, Aqidah, History, and Spirituality. Relevance judgments were independently provided by three Qur’anic studies experts and reconciled through discussion, with inter-annotator agreement indicating substantial consistency. Each query included 20 non-relevant passages as negative samples to increase evaluation difficulty. Two approaches were tested: retrieval-only and retrieval combined with a zero-shot QA module for span extraction. Retrieval-only achieved slightly higher top-1 accuracy (0.72), but retrieval + QA improved ranking-oriented metrics, including Accuracy@5 (0.88), Mean Reciprocal Rank (MRR = 0.76), and normalized Discounted Cumulative Gain at 5 (nDCG@5 = 0.82), with the increase in Accuracy@5 statistically significant (p = 0.01). The zero-shot QA module enabled the system to extract more precise and contextually relevant information, enhancing overall retrieval quality and robustness. These results indicate that the proposed system effectively retrieves relevant tafsir passages and provides accurate, context-specific answers. The study demonstrates the potential and limitations of zero-shot QA for domain-specific religious texts and supports the development of web-based applications or Islamic chatbots, facilitating easier access to shahih tafsir knowledge for scholars and the broader Muslim community.