TBCUP: A Transformer-based Code Comments Updating Approach

Good code comments are of great value for program maintenance. However, during the development process, developers often neglect to update the corresponding comments when changing the code, which results in inconsistent comments and affects the maintainability of software. Studies have shown that even in widely used programs, there is a large number of outdated comments. Existing code comment update approaches use long short-term memory (LSTM) model based on encoder-decoder to capture the relationship between code changes and comment updates, to automatically update code comments. However, due to the complexity of code changes, the corresponding comments update results do not perform well. Thus, a Transformer-based automatic update approach for code comments called TBCUP is proposed in this paper. TBCUP uses the multi-head attention mechanism to learn the relationship between code change sequences and comment updates to update comments more accurately. In addition, TBCUP uses the Byte Pair Encoding (BPE) algorithm to build a unified vocabulary for code and corresponding comments to alleviate the problem of out–of-vocabulary (OOV) words while updating comments. With BPE, TBCUP can split or combine words to handle OOV words. The experimental results on the dataset containing 108k sets of code-comment co-evolution samples show that the accuracy of TBCUP is improved by 3.2% in comparison with CUP.