Um estudo empírico da comparação de técnicas de recomendação de tarefas e códigos fontes similares em projetos de software livre

Abstract

Managing issues in open-source software projects is a challenging and costly task, as many developers are casual and/or novice. On one hand, maintainers need to ensure the quality of issue descriptions and their labels and create mechanisms for recommending and assigning issues. On the other hand, to successfully complete the issue, contributors need to understand it and locate the artifacts related to a given functionality or the defect to be fixed. The goal of this work was to conduct a comparative study of different models for recommending similar issues that could help developers with their contributions. For this we collected data on issues and pull requests from 35 open-source projects hosted on GitHub. We used the Term Frequency - Inverse Document Frequency (TF-IDF), Sentence BERT (SBERT), and Word2Vec techniques to recommend similar issues and source code to assist in the contributions of novice developers. We observed that the models based on the SBERT and TF-IDF techniques yielded better results in the recommendations generated compared to Word2Vec in the two evaluated scenarios (general issues and those marked as good for beginners). These models recommended issues where the code used in the solution was approximately 17% similar to the actual solution of the issue. We concluded that, based on the empirical results obtained, we hope to take the next steps in transferring the knowledge gained to software projects and developers, especially by supporting novice developers during their first contribution.