Automatic Program Repair with Improved Transformer

The rise of Deep Learning (DL) has led researchers to explore techniques for Automatic Program Repair (APR) using DL. However, existing methods for APR suffer from issues such as small datasets, low accuracy, and long execution times. In this paper, we propose a method based on transformer for APR called TLinearFFN, and we augment the dataset. The results show that on the original dataset, our model achieved a repair accuracy improvement of approximately 13.81% compared to previous studies, and approximately 2.78% compared to the original transformer model. Unfortunately, our augmented dataset did not have a positive impact on accuracy. To summarize, our results demonstrate that the TLinearFFN method has excellent code repair performance, provides a more efficient and accurate solution for code repair.