FILI: Syntax Repair By Learning From Own Mistakes

Sahil Bhatia, Navneet Potti, Rishabh Singh

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Automatic Program Repair, Software Engineering, Neural Syntax Fix

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Automatically fixing syntax errors in programs is a key challenge in Software Engineering community. Although, there are millions of programs on the web, both syntactically correct and incorrect, finding a large number of paired examples of <correct, incorrect> programs is difficult. This makes training a program fixer using supervised learning difficult. Recently, BIFI, an unsupervised approach for learning a syntax fixer was proposed, in which an additional model (Breaker model) is used to augment data in each learning iteration to match real-world error distribution. In this paper, we propose a novel approach, FILI (Fix-It-Learn-It) for learning a syntax fixer without having to train any additional models for data augmentation. In each iteration, FILI carefully selects examples from the fixer's own predictions, both correct and incorrect, and uses those to fine-tune the fixer. We also show that gradually increasing the complexity of the examples during training leads to a more accurate fixer. Our evaluation on the Github-Python dataset shows that FILI outperforms BIFI by 1% while being significantly easier to train. Moreover, FILI avoids training the breaker model training a 13 million parameter breaker model in each iteration, which can take about 2 days on a modest DNN accelerator.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9014