A Benchmark Arabic Dataset for Arabic Question Classification using AAFAQ Framework

Abstract

Arabic Natural Language Processing (NLP) is still faced with the complexity of the language's morphology and the limited availability of quality annotated resources. In this paper, we introduce an open-domain dataset of 5,009 Modern Standard Arabic (MSA) questions labeled according to AAFAQ framework that has11 linguistic and cognitive aspects, e.g., Question Particle, Question Particle Type, Intent, Answer Type, Cognitive Level, and Temporal Context. Based on the AAFAQ Framework (Arabic Analytical Framework for Advanced Questions), the dataset is designed to support semantic and cognitive understanding for Arabic Question Classification and related tasks. The dataset's effectiveness was validated by fine-tuning state-of-the-art models. AraBERT achieved 100% accuracy on Question Particle Type classification and 94.95% on Intent classification. Integration within a generative question-answering system with Alpaca + Gemma-9B Unsloth improved evaluation metrics, including BLEU (+37.6%), ROUGE-1 (+132%), and BERTScore (+17.3%), validating the dataset's value in both classification and generation tasks. Despite its broad coverage, the dataset includes underrepresented categories, e.g., Sociology and Volunteering, to be considered in future extensions. AAFAQ is a foundation benchmark for the advancement of Arabic question comprehension, with prospective applications in education, cognitive computing, and multilingual AI system creation.