Exploring the Impact of Deep Learning Techniques on Evaluating Arabic L1 Readability

Educational researchers are interested in the causes and effects of reading difficulties on learners because reading is one of the primary pathways for language learning and knowledge acquisition. Thus, any difficulty in reading can affect the learning and comprehension process. In view of this, several approaches have been proposed by the researchers to automatically assess the difficulty level of texts. Thanks to recent advances in linguistics and computer science, text readability analysis now relies on a variety of linguistic indicators to measure text complexity, as well as on powerful computer models. Research on Arabic as a foreign language (L2) is more advanced than research on Arabic as a first language (L1). In this study, we outline several approaches to assessing the readability of Arabic texts for L1 learners. Two approaches have been used to evaluate the readability of texts. The first one is mainly based on the evaluation of sophisticated handcrafted features, which represent the texts and allow estimating their readability level. The second approach evaluates the use of different contextual and non-contextual word vectors (CBOW, Skip-Gram and AraBert) instead of the handcrafted features. The results indicate that the AraBert model achieves the best readability prediction accuracy of about 76.93%.