[4Rin1] Interactive Session 2

[4Rin1-42] Neural commit message generation integrating multiple work context encoders

〇Tomohiro Iwanaga¹, Yuki Managaki¹, Koichi Tanigaki¹ (1. Fukui University of Technology)

Keywords:text generation, seqence to seqence, version control system

This paper proposes a new neural network model for commit message generation. Previous studies simply
apply machine translation models to convert code diffs to commit messages, which lacks the mechanism to learn
how to write intentions of code modications, although it is often required in desirable commit messages. The
proposed model introduces multiple encoders to efficiently capture different levels of working contexts, such as
code modications and original bug reports. Experimental results using GitHub dataset showed degradation when
adopting issues as additional contexts, which suggests the commit style specicity of the dataset.