AraQA: An Arabic Generative Question-Answering Model for Authentic Religious Text
Yousef Adel, Mostafa Dorrah, Ahmed Ashraf, Abdallah ElSaadany, Mahmoud Mohamed, M Khedr Wael, Ghada Ahmed Khoriba · 2023
Recently, the internet has become a vast repository of religious texts and sources. The quest for valid and dependable Islamic Q/A that provides accurate substantiation from the Holy Quran (Muslim holy book) and Hadith (Prophet Muhammed teachings) has become challenging, given the abundance of misleading answers lacking credible evidence and proper sources. Concurrently, transformer-based architectures have demonstrated remarkable efficacy in language modeling and comprehension. Notably, applications in religious Arabic generative question answering have remained underdeveloped, primarily due to the lack of available Arabic religious datasets. In this paper, we present an Arabic Islamic generative question-answer model named AraQA, which has been fine-tuned using Arabic Islamic question/answer pairs meticulously gathered and extracted from reputable open-source web sites on the internet. The model is initially designed to operate exclusively in the Arabic Language. Our model attains an impressive perplexity score of 2.3 when evaluated on held-out question-answer pairs. We have made the model publicly accessible via GitHub, anticipating it will pave the way for research in Arabic and religious Natural Language Processing (NLP).11https://github.com/Marje3na/AraQA-An-Arabic-Generative-Question-Answering-Model-for-Authentic-Religious-Text