FAAQA-QAD: A Frequently Asked Arabic Question Answering Dataset for Diseases Detection System

Domain-specific question-answering (QA) systems are a specialized field in Natural Language Processing. Their objective is to generate answers to questions within a specific domain, such as healthcare. The existing Arabic datasets are often characterized by their low quality, resulting from inadequate preprocessing or inaccurate annotation because that translated from another language. To tackle this issue, our work introduces a new Modern Standard Arabic QA dataset, entitled FAAQA-QAD. The dataset was collected in Modern Standard Arabic without the use of translation. For the preprocessing and annotation of the FAAQA-QAD dataset, we employed the Pandas library and an end-to-end comprehension CdQA annotator. We further evaluated our FAAQA-QAD dataset using various pre-trained language models, demonstrating promising results in Arabic NLP tasks. This evaluation allowed us to compare the performance of these models using our dataset and conduct an analysis to comprehend the reasons behind the underperformance of certain models. Through our experimentation, we observed that the ARBERT model exhibited superior performance when compared to other RC span-extraction datasets using the FAAQA-QAD dataset.