Esapp: um modelo para realizar a estimativa de esforço de software a partir de requistos textuais

• Ano de defesa: 2025
• Autor(a) principal: Silva, Daniel de Oliveira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Cornelio Procopio; Brasil; Programa de Pós-Graduação em Informática; UTFPR
• Palavras-chave em Português: Engenharia de software; Processamento de linguagem natural (Computação); Software de aplicação - Desenvolvimento; Software engineering; Natural language processing (Computer science); Application software - Development; CNPQ::ENGENHARIAS; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/38861

Effort estimation in software projects plays a crucial role, directly influencing planning, cost, and development success. With the growing adoption of agile methodologies, new estimation techniques focused on agility and efficiency have emerged. At the same time, the use of machine learning techniques has enabled the development of models that seek to increase the accuracy of these estimates. However, many of these models were trained exclusively on English-language data, limiting their applicability in other languages. Therefore, this dissertation proposes the study and development of an effort estimation model based on textual requirements, supporting both Portuguese and English. The model was built with a deep learning LSTM architecture, using pretrained embeddings from a contextualized language model. For comparison purposes, the XLNet and FastText models (contextualized and context-free, respectively) were selected. To conduct the experiment, we used a database containing 23,313 user stories extracted from 16 open source projects. The experiments were conducted with data in both languages, obtaining, for English, a Mean Absolute Error (MAE) of 3.77, a Mean Squared Error (MSE) of 79.94, and a Median Absolute Error (MOAE) of 1.93, outperforming related works, such as the BERT model, with improvements of 11% in the MAE metric, 7% in the MSE, and 16% in the MOAE. For Portuguese, the results were: MAE of 3.94, MSE of 81.33, and MOAE of 1.94. These results indicate that the proposed model is effective and competitive with the state-of-the-art, in addition to being applicable to different languages. As a practical contribution, the dissertation also presents a web application that allows the use of the model by development teams, promoting faster, more accessible and accurate estimates.