Automatic software code repair using deep learning techniques

In the multi-hundred-billion-dollar industry of software development, the debugging process is an expensive task for developers. So, much effort has been put into debugging automation. In the last decade, researchers have repaired codes according to predefined rules which are only effective in limited types of bugs. Through a lot of experiments, researchers have found that deep learning models are useful in code repair automation similar to the revolutionary results these approaches have produced in various other fields. To solve such a difficult problem, recent works focus on debugging bugs that appear on one line of code. It has been reported that this type of bug occurs at an average rate of 1 out of every 1600 lines of code in a software project, which is significant. The current research follows these approaches and introduces a novel automatic code-repair system. We have employed the transfer learning technique to reuse a pre-trained model on the problem. The proposed system is designed based on the encoder-decoder architecture. In the encoder, a new pre-trained Bert model named JavaBert is used. Then, the model was fine tuned. The decoder is a transformer with an autoregressive structure. ManySStuBs4J [1] dataset is used for evaluation purposes. The results of the evaluations show that the proposed system has higher accuracy and BLEU criteria than CodeBert and the baseline model. Baseline is a simple model that acts as a reference in machine learning studies and CodeBert is one of the most similar models to the proposed model. The bilingual evaluation understudy score (BLEU) improvement is between 0.04 and 0.16%, the accuracy improvement is between 0.64 and 5.81%, the recall improvement is between 1.08 and 9.2%, and the F-score improvement is between 3.27 and 6.18%.