Adversarial training and ensemble learning for automatic code summarization

Abstract

Natural language summaries of codes are important during software development and maintenance. Recently, deep learning-based models have achieved good performance on automatic code summarization, which encode token sequence or abstract syntax tree (AST) of code with neural networks. However, almost all of these models are trained using maximum likelihood estimation, which do not guarantee the quality of generated summaries. Moreover, existing models that benefit from multiple encoders lack a fined-grained selection between different encoders, and the encoders may be insufficiently optimized. To address these issues and generate better code summaries, we propose a novel code summarization framework based on adversarial training and ensemble learning. It includes two separately trained encoder-decoder models, one for source code sequence and the other for its AST. Here, an efficient approach to obtain AST node sequence is introduced. We train our models via adversarial training, where each model is guided by a well-designed discriminator that learns to evaluate its outputs. During inference, a module named mixture network is introduced to compute an adaptive combination weight of the models’ outputs. We evaluate our framework on a large Java corpus and compare it to several state-of-the-art models. Experimental results show that our approach outperforms the best baseline by 22.6% on BLEU-4, 5.7% on ROUGE-L and 7.6% on METEOR.