Classificação de requisitos de software com processamento de linguagem natural utilizando bert

Abstract

The task of classifying software requirements can be an extensive and expensive task, since it demands time and effort when performed manually, especially in large projects, since requirements written in natural language can have different interpretations, that is, their classification depends on from the interpretation of who performs the task, this can lead to failures that can be reflected in the software life cycle. Faced with the need to classify requirements in an automated way, this work demonstrates how Natural Language Processing (NLP) through the BERT language model can support the intelligent classification of requirements. The method is applied to the PROMISE dataset, reaching an accuracy of 92.8% for classifying functional requirements and non functional requirements.