Evaluating the Impact of Dataset Size on the Accuracy of Arabic Software Requirements Classification Using AraBERT

Requirement’s engineering is crucial in software development, underpinning all the later stages, and heavily impacting whether a project succeeds or fails. A standard software requirements document includes functional and non-functional requirements — both are necessary. But, to classify and represent these requirements in natural language takes much human effort. It's laborious and takes a lot of time to manually classify; expensive and error-prone. Misclassification may cause misunderstandings or ambiguities and consequent incompleteness of the product not meeting customer expectations. Though the classification of software requirements has been studied extensively in English, Arabic still lacks such research. Moreover, Arabic datasets specific to this field are not publicly available: in our prior work, we used some data to classify Arabic software requirements and get results. This paper presents an extension of our work that investigates how dataset size affects classification performance. We propose a method to gather Arabic software requirements datasets and use deep learning techniques for automatic classification into functional and non-functional requirements. The study is intended to assist software developers in saving time, cost, and manual classification effort — thus making requirements engineering more efficient. It also creates a new path for researchers in this field.