KeyTitle: towards better bug report title generation by keywords planning

Abstract

Bug reports play an important role in the software development and maintenance process. As the eye of a bug report, a concise and fluent title is always preferred and expected by developers as it could help them quickly seize the problem point and make better decisions in handling the bugs. However, in practice, not all titles filled by bug reporters are found to be of high quality; some may not carry essential bug-related information, and some may be hard to understand or contain extra noise. With the aim to reduce the burden of bug reporters and ease developers’ life in handling bugs, we propose a deep learning-based technique named KeyTitle, to automatically generate a title for a given bug report. KeyTitle formulates the title generation problem as a one-sentence summarization task. It could be viewed as a Seq2Seq generation model (which generally directly generates target text based on source text) that incorporates keywords planning. Specifically, within KeyTitle, a transformer-based encoder-decoder model is enforced to generate a chain of keywords first from the detailed textual problem description, and then generate the target title by considering both these keywords and description content. Experiments over three large bug datasets collected from GitHub, Eclipse, and Apache shows that KeyTitle could outperform state-of-art title generation models relatively by up to 8.9-18.2\(\%\), 11.4-30.4\(\%\), and 13.0-18.0\(\%\) in terms of ROUGE-1, ROUGE-2, and ROUGE-L F1-scores; the titles generated by KeyTitle are also found to be better in terms of Relevance, Accuracy, Conciseness, Fluency in human evaluation. Besides generating titles from textual descriptions, KeyTitle is also found to have great potential in generating titles based on just a few keywords, a task that also has much value in bug reporting/handling practice.