Scholarlink

<p>Question Answering (QA) over the Holy Quran remains a challenging task due to the complexity of Classical Arabic, the density of metaphorical expressions, and the deep contextual dependencies inherent in the scripture. Existing QA systems—including those submitted to the Qur’an QA&#xa0;2022 shared task—often struggle to generate precise and contextually coherent answers, especially under low-resource conditions. In this study, we utilize the <b>Quranic Reading Comprehension Dataset (QRCD) v1.2</b> and introduce <b>Advanced Quranic Retrieval-Augmented Generation (AQRAG)</b> for low-resource Quranic QA. The proposed framework integrates retrieval-augmented generation with a reranking mechanism designed to improve verse selection and strengthen the factual and contextual grounding of generated responses. The methodology is carefully adapted to handle the linguistic richness and contextual depth characteristic of Quranic discourse. Experimental results demonstrate that AQRAG yields substantial improvements over both the QRCD v1.2 baseline and previously reported systems. The fine-tuned AraELECTTRA-discriminator-QuranQA baseline achieves <b>64%&#xa0;pRR</b>, while our RAG-based models—evaluated using <b>BERTScore</b>—obtain markedly higher performance: Gemini−2.5-flash with Naive RAG reaches <b>89.9%</b>, Qwen−2.5-72B with RAG achieves <b>89.8%</b>, and GPT−4.1-nano with Advanced RAG and reranking attains <b>90.5%</b> in Islamic-context evaluation. These results indicate that advanced RAG strategies, particularly those incorporating reranking, establish a new performance benchmark for Quranic QA on the QRCD v1.2 dataset.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Aqrag: advanced quranic retrieval-augmented generation for low-resource question answering

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Mostafa G. Abd Elhakeem,
  • Amira O. Abdullah,
  • Sara W. Elhag,
  • Ensaf Hussein Mohamed

摘要

Question Answering (QA) over the Holy Quran remains a challenging task due to the complexity of Classical Arabic, the density of metaphorical expressions, and the deep contextual dependencies inherent in the scripture. Existing QA systems—including those submitted to the Qur’an QA 2022 shared task—often struggle to generate precise and contextually coherent answers, especially under low-resource conditions. In this study, we utilize the Quranic Reading Comprehension Dataset (QRCD) v1.2 and introduce Advanced Quranic Retrieval-Augmented Generation (AQRAG) for low-resource Quranic QA. The proposed framework integrates retrieval-augmented generation with a reranking mechanism designed to improve verse selection and strengthen the factual and contextual grounding of generated responses. The methodology is carefully adapted to handle the linguistic richness and contextual depth characteristic of Quranic discourse. Experimental results demonstrate that AQRAG yields substantial improvements over both the QRCD v1.2 baseline and previously reported systems. The fine-tuned AraELECTTRA-discriminator-QuranQA baseline achieves 64% pRR, while our RAG-based models—evaluated using BERTScore—obtain markedly higher performance: Gemini−2.5-flash with Naive RAG reaches 89.9%, Qwen−2.5-72B with RAG achieves 89.8%, and GPT−4.1-nano with Advanced RAG and reranking attains 90.5% in Islamic-context evaluation. These results indicate that advanced RAG strategies, particularly those incorporating reranking, establish a new performance benchmark for Quranic QA on the QRCD v1.2 dataset.