Tokenizers United at QIAS-2025: RAG-Enhanced Question Answering for Islamic Studies by Integrating Semantic Retrieval with Generative Reasoning
Résumé
This paper presents the approach and results for Sub Task 2: General Islamic Knowledge Question Answering at QIAS 2025, a shared task designed to evaluate the capabilities of Large Language Models (LLMs) in answering multiple-choice questions across diverse domains of Islamic knowledge, including theology, jurisprudence, biography, and ethics.A Retrieval-Augmented Generation (RAG) system powered by the Gemini language model was developed for this task.In the proposed system, the retriever module performs semantic search over curated classical Islamic sources to identify passages relevant to each input question, while the generator module leverages the LLM to reason over the retrieved evidence and generate a final answer.This integration of evidence retrieval with contextual reasoning enables accurate responses across diverse knowledge areas.On the official test set, the system achieved an accuracy of 87%, ranking 5th out of 10 participating teams in QIAS 2025 Sub Task 2. These results demonstrate the effectiveness of combining retrieval-based evidence with generative reasoning in specialized religious domains, highlighting the potential of RAG architectures for high-stakes, knowledge-intensive question answering tasks and confirming their robustness in the QIAS 2025 benchmark.
Citer ce document
Accès au document
Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter
Voir l'article sur le site de la revueAuteur(s)
Statistiques
Consultations : 1
Téléchargements : 0