Do Not Have Enough Data? An Easy Data Augmentation for Code Summarization

Abstract

Code comments improve the readability and intelligibility of codes, which can help developers understand programs and improve the efficiency of the software maintenance and evolution process. Unfortunately, code comments are often mismatched, missing, or outdated in software projects, which negatively affects the efficiency of developers to infer the functionality from source code and affect the efficiency of software maintenance and evolution. To solve this problem, many source code summarization algorithms have been proposed. However, these methods usually try to collect a large data set which contains the mapping between code comments and source code to train models. Hence, the effectiveness of the models often rely on the quality of the training data. There are two limitations for the training sets: the insufficient data collection limitation (i.e., generate a large amount of noises-free training data) and data distribution bias limitation (i.e., generate training data for infrequently used methods). To address this issues, we have proposed a data augmentation method for code comments, named CDA-CS. Extensive experiments on Java and Python projects collected from GitHub are conducted to evaluate the performance of CDA-CS. Training models on the augmented dataset, the state-of-the-art algorithms can easily get a further 1.37% to 2.24% improvement in terms of different evaluation metrics (i.e., BLEU-4, METEOR, ROUGH_L) with no additional cost.