A Recursive Prompt-Driven Reasoning Approach for Multi-Hop Arabic Fatwa Question Answering

Answering Arabic fatwa inquiries represents a highly critical yet challenging task in Natural Language Processing, requiring consideration of multiple jurisprudential conditions, interpretation of interconnected concepts, and synthesis of evidence distributed across multiple religious sources. Existing Arabic question answering systems, including conventional single-step question answering (QA) models and retrieval-augmented generation (RAG) frameworks, typically perform retrieval only once and therefore struggle to capture the intermediate reasoning steps required for complex religious inquiries, often resulting in incomplete reasoning and factual hallucinations. To address these limitations, this study proposes a recursive prompt-driven reasoning approach for multi-hop Arabic fatwa question answering that integrates instruction-based supervision, task-specific language models, hybrid retrieval, and recursive reasoning to support end-to-end multi-hop question answering. The proposed method employs an instruction-based supervision strategy constructed from the MAFQA dataset to train specialized question decomposition and question answering models. To support evidence grounding, a hybrid retrieval module combines dense semantic retrieval, sparse lexical matching, and cross-encoder reranking to retrieve both few-shot demonstrations and supporting evidence at each reasoning hop, while a relevance-guided context update mechanism retains only the most informative contextual evidence throughout the recursive reasoning process. Comprehensive experiments conducted on the MAFQA dataset demonstrate the effectiveness of the proposed approach. The hybrid retrieval module consistently outperforms individual dense and sparse retrieval methods, while task-specific fine-tuning significantly improves end-to-end performance across Arabic sequence-to-sequence models and large language models under realistic inference conditions. The fine-tuned GPT-4.1 configuration achieves the best overall performance, obtaining an F1-token score of 31.88, a ROUGE-L score of 26.64, and a Relevance score of 92.39. Furthermore, the proposed approach surpasses representative state-of-the-art Arabic question answering systems, retrieval-augmented generation methods, and reasoning-based approaches, achieving a 36.5% relative improvement in F1-token over the strongest reasoning baseline. These results demonstrate that integrating recursive question decomposition, hybrid retrieval, and dynamic context updating improves performance under the evaluated multi-hop Arabic fatwa question answering setting.

@article{Al-Qahtani2026,
  title     = {A Recursive Prompt-Driven Reasoning Approach for Multi-Hop Arabic Fatwa Question Answering},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {8},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Manal Al-Qahtani and Bader Alkhamees and Mourad Ykhlef},
  doi       = {10.14569/IJACSA.2026.0170847},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170847}
}