MS-CUP: A Multi-Source Code Comment Updating Method

Code comments are vital for program comprehension and maintenance, yet developers often neglect to update them during code evolution, which results in inconsistent or misleading documentation. Existing methods for automatic comment updates, such as CUP, HEB-CUP, and HAT-CUP, primarily treat code as natural language text. These methods overlook the inherent structural properties of code and struggle to align complex code changes with comment semantics. To address these limitations, we propose MS-CUP, a multi-source fusion method that integrates old comment sequences, code edit sequences, and AST difference sequences within a Transformer- based architecture. By simplifying AST differences and leveraging multi-head attention mechanisms, MS-CUP captures syntactic and semantic dependencies between code evolution and comment updates. Experiments on a dataset of 108k code-comment coevolution samples demonstrate that MS-CUP outperforms other advanced methods, achieving improvements of 5.3% in Accuracy and 2.6% in Recall@5 at least.