<scp>Katana</scp> : Dual Slicing Based Context for Learning Bug Fixes

Contextual information plays a vital role for software developers when understanding and fixing bug. Consequently, deep learning based program repair techniques leverage context bug fixes. However, existing treat in an arbitrary manner, by extracting code close proximity of the buggy statement within enclosing file, class, or method, without any analysis to find actual relations with To reduce noise, they use predefined maximum limit on number tokens be used as context. We present slicing approach, which instead arbitrarily including context, we analyze statements that have control data dependency statement. propose novel concept called dual , leverages both fixed versions capture relevant ingredients. our technique tool Katana first apply slicing-based task. The results show effectively preserves sufficient model choose contextual while reducing noise. compare against four recent state-of-the-art context-aware techniques. Our fixes between 1.5 3.7 times more bugs than