Towards an Open Domain Arabic Question Answering System: Assessment of the Bert Approach

Recently, deep learning-based contextualized word representations have made substantial advancements in enhancing the efficiency of various natural language processing (NLP) applications. However, only limited efforts have been dedicated to employing these representations for the development of Arabic opendomain question-answering (QA) systems, which are an indispensable component of conversational agents such as ChatGPT. In this study, we address this gap by delving into the Bert architecture to create a pre-trained Arabic Bert model. Furthermore, we assess the performance of this model in constructing a QA system by comparing its performance with that of a multilingual Bert model. The experimental results show that our AraQA_Bert_SL model, fine-tuned on the weights of a single-language pre-trained model, outperforms existing systems, boasting an F1 score of 90.6% and a pRR score of 93.7%. This achievement surpasses the performance of the AraQA_Bert_ML model, which relies on a multilingual pre-trained model. Notably, our approach significantly reduces the computational costs associated with the process of Bert fine-tuning.