Sistema inteligente de busca de issues em repositórios de código

Abstract

The concern about how software developers use their time and attempts to automate manual processes have taken on a greater proportion every year. In Software Engineering, recommender systems help developers find information that should be known, evaluate alternatives and aid data navigation. Currently, Git-based shared code repositories are part of everyday life for most software developers. These repositories make several functionalities available to their users, among them, issues are a way to document problems, request some documentation, provide feedback and track the workflow on a problem, however, active code repositories tend to generate many issues, which which in turn ends up making the task of recovering relevant issues something more costly for its users. The purpose of this work is the elaboration of a recommendation system capable of analyzing issues, using unsupervised machine learning, through the Word2Vec natural language processing algorithm, with the objective of facilitating the search for issues and assisting in decision making based on stored issues. in a code repository. To achieve this goal, it was necessary to extract the issues from the code repository, pre-process the previously defined fields, create unigram, bigram and trigram dictionaries and finally train the neural network used by the Word2Vec algorithm, using as input previously created dictionaries. After the model has already been trained, the user will be able to search for issues through a data entry, this entry may contain one, two or three words and should return a set of issues more semantically similar to the entry. The evaluations showed that the recommendation system makes relevant suggestions, helps with navigation and facilitates decision-making regarding issues in code repositories.