Pré-processamento inteligente de requisitos textuais de software para refinamento de estimativa de esforço

Abstract

The present work focuses on the application of Artificial Intelligence (AI) methods and techniques, with an emphasis on Natural Language Processing (NLP), aiming to improve software effort estimation by analogy (EESA). The analysis of requirements texts (e.g., user stories) enables the generation of features for text classification, allowing for the automation and improvement of EESA accuracy. Effort estimation by analogy requires a deep understanding of historical data when the requirements are presented in textual format. In this context, this study proposes an intelligent process for pre-processing textual software requirements to refine effort estimates. The goal is to efficiently classify user requirements and use this categorization to infer effort estimates. The methodology adopted is inspired by previous studies, which explored pre-trained models GPT-3 and BERT for EESA. However, this study enhances the approach by focusing on the textual pre-processing method, including the application of GPT-3.5. Two distinct pre-processing strategies were outlined. The first uses a conventional method with Regular Expressions (RegEx). The second strategy explores the GPT-3.5 API to obtain a pre-processed database, using the model to produce clean and normalized texts. After pre-processing, two experiments were conducted for EESA inference, using two different methods, one via a proprietary neural network architecture and the other via the GPT-3.5 API. Metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and Median Absolute Error (MdAE) were extracted. After obtaining the metrics from the two experiments, the results were compared with those obtained in previous studies (Z-SE2 by Baratto (2022) and SE3M by Fávero, Casanova e Pimentel (2022)). Notably, the MAE value of 0.56 ± 0.15 came from the experiment that used inter-repository partitioning from a proprietary neural model.