New hybrid neural system achieves 70% recall for Quranic passage retrieval

A new research paper presented at IMSA 2026 introduces a hybrid neural architecture designed specifically for Quranic Passage Retrieval (PR). The task is challenging due to the linguistic gap between Modern Standard Arabic (MSA) used in queries and the Classical Arabic (CA) of the Quran. The proposed system, by Mohamed G. Salman, Mohammad E. Moftah, and Ali Hamdi, addresses this with four distinct phases. First, hybrid candidate retrieval combines AraColBERT dense indexing with BM25 sparse retrieval to capture both semantic and lexical matches. Second, a CAMeLBERTmix cross-encoder performs semantic reranking on the candidates. Third, a confidence gating mechanism filters out queries likely to have zero relevant answers, improving reliability. Finally, an AraT5-based generative refinement module aggregates multiple relevant verses into a coherent output.

Evaluated on an expanded version of the Quran QA 2022 dataset, the system achieves a Recall@10 of 0.7024 and a Mean Average Precision (MAP@10) of 0.4947. While its Mean Reciprocal Rank (MRR) of 0.5807 shows a slight tradeoff in top-rank precision compared to heavily optimized single models, the overall architecture provides substantially better coverage and context awareness. This makes it particularly suited for multi-verse retrieval where users need multiple relevant passages rather than a single perfect match. The system effectively handles the semantic complexity of Quranic text and reduces the issue of zero-answer queries, a common failure point in traditional retrieval systems.