Break-It-Fix-It: Unsupervised Learning for Program Repair

2022년 11월 04일 (금) | 발표자: 윤태현

In this paper, authors proposes an unsupervised learning framework for program repair, Break-it-fix-it (BIFI), that augments a critic to the backtranslation. BIFI trains a fixer (code repairer) and a breaker (code corrupter) iteratively while generating data; use a critic (code analyzer or compiler) to check the fixer's output satisfiability. Finally, the paper shows that learning from real bad code significantly improves the code repair performance compared to learning from synthetic bad code.