NLP · Software Mining
Refactoring-Commit Intent Classifier
Multi-class intent classification over a large corpus of GitHub refactoring commit messages.
Question
Given a commit message from a refactoring history, can we reliably infer the developer intent (bug fix, functional change, code smell, internal QA, or external QA)?
Method
I led post-collection NLP processing on a pre-curated corpus of 111k GitHub commit messages: cleaning, lemmatization, TF-IDF with n-gram features for five intent classes. Models were compared under randomized grid search and 10-fold cross-validation.
Result
A Random Forest classifier reached about 0.87 macro F1, outperforming SVM, MLP, and logistic baselines on the same folds.
Limitation
Commit messages are short and noisy; labels inherit upstream annotation choices. The model does not inspect diffs, so intent that only appears in code changes can be missed.