Bash comment generation via data augmentation and semantic-aware CodeBERT

Yiheng Shen, Xiaolin Ju, Xiang Chen, Guang Yang

Abstract: Understanding Bash code is challenging for developers due to its syntax flexibility and unique features. Bash lacks sufficient training data compared to comment generation tasks in popular programming languages. Furthermore, collecting more real Bash code and corresponding comments is time-consuming and labor-intensive. In this study, we propose a two-module method named Bash2Com for Bash code comments generation. The first module, NP-GD, is a gradient-based automatic data augmentation component that enhances normalization stability when generating adversarial examples. The second module, MASA, leverages CodeBERT to learn the rich semantics of Bash code. Specifically, MASA considers the representations learned at each layer of CodeBERT as a set of semantic information that captures recursive relationships within the code. To generate comments for different Bash snippets, MASA employs LSTM and attention mechanisms to dynamically concentrate on relevant representational information. Then, we utilize the Transformer decoder and beam search algorithm to generate code comments. To evaluate the effectiveness of Bash2Com, we consider a corpus of 10,592 Bash code and corresponding comments. Compared with the state-of-the-art baselines, our experimental results show that Bash2Com can outperform all baselines by at least 10.19%, 11.81%, 2.61%, and 6.13% in terms of the performance measures BLEU-3/4, METEOR, and ROUGR-L. Moreover, the rationality of NP-GD and MASA in Bash2Com are verified by ablation studies. Finally, we conduct a human evaluation to illustrate the effectiveness of Bash2Com from practitioners’ perspectives.