Automatic Bug Triage Using Hierarchical Attention Networks

Bug triage, which plays a critical role in software maintenance, mainly refers to the process of assigning a bug report to an appropriate developer who could fix it. Manual bug triage, especially in open-source projects, is quite a burden to developers. To address this problem, many methods have been proposed for automatic or semi-automatic bug triage, which can be observed as a task under text classification in natural language processing. In this article, we present an end-to-end approach using hierarchical attention networks to establish an automatic bug triage system. Considering syntactic and semantic information in bug reports, two methods, Word2Vec and GloVe, are provided to pre-train word vector presentation on untriaged reports to attain high speed and accuracy, respectively. Fine-tuned word vectors achieve better results during training. We validate the performance on five large-scale public datasets. From the results of the experiments, our approach achieves higher accuracy amongst existing methods based on deep neural networks.