The effectiveness of data augmentation in code readability classification

Qing Mi

doi:10.1016/j.infsof.2020.106378

The effectiveness of data augmentation in code readability classification

2021-01-01

17

PRE

AI

DOI：10.1016/j.infsof.2020.106378

原文链接

原文求助

摘要

En 中文

Context: Training deep learning models for code readability classification requires large datasets of quality pre-labeled data. However, it is almost always time-consuming and expensive to acquire readability data with manual labels. Objective: We thus propose to introduce data augmentation approaches to artificially increase the size of training set, this is to reduce the risk of overfitting caused by the lack of readability data and further improve the classification accuracy as the ultimate goal. Method: We create transformed versions of code snippets by manipulating original data from aspects such as comments, indentations, and names of classes/methods/variables based on domain-specific knowledge. In addition to basic transformations, we also explore the use of Auxiliary Classifier GANs to produce synthetic data. Results: To evaluate the proposed approach, we conduct a set of experiments. The results show that the classification performance of deep neural networks can be significantly improved when they are trained on the augmented corpus, achieving a state-of-the-art accuracy of 87.38%. Conclusion: We consider the findings of this study as primary evidence of the effectiveness of data augmentation in the field of code readability classification.

Keyword:

Code readability classification

Data augmentation

Generative adversarial network

Deep learning

Empirical software engineering

The effectiveness of data augmentation in code readability classification

摘要

AI总结

期刊

机构

引用论文