Revisiting Learning-based Commit Message Generation

Commit messages summarize code changes and help developers understand the intention. To alleviate human efforts in writing commit messages, researchers have proposed various automated commit message generation techniques, among which learning-based techniques have achieved great success in recent years. However, existing evaluation on learning-based commit message generation relies on the automatic metrics (e.g., BLEU) widely used in natural language processing (NLP) tasks, which are aggregated scores calculated based on the similarity between generated commit messages and the ground truth. Therefore, it remains unclear what generated commit messages look like and what kind of commit messages could be precisely generated by existing learning-based techniques. To fill this knowledge gap, this work performs the first study to systematically investigate the detailed commit messages generated by learning-based techniques. In particular, we first investigate the frequent patterns of the commit messages generated by state-of-the-art learning-based techniques. Surprisingly, we find the majority ( 90%) of their generated commit messages belong to simple patterns (i.e., addition/removal/fix/avoidance patterns). To further explore the reasons, we then study the impact of datasets, input representations, and model components. We surprisingly find that existing learning-based techniques have competitive performance even when the inputs are only represented by change marks (i.e., “+”/“-”/“ ”), It indicates that existing learning-based techniques poorly utilize syntax and semantics in the code while mostly focusing on change marks, which could be the major reason for generating so many pattern-matching commit messages. We also find that the pattern ratio in the training set might also positively affect the pattern ratio of generated commit messages; and model components might have different impact on the pattern ratio.