Building a Dataset for Combined Classification of Source Code Reviews
Pavlína Petrová, Sergey I. Markov, Vladimir Vladimirovich Kachanov · Pattern Recognition and Image Analysis · 2025
Abstract The problem of classifying source code reviews is considered. A new dataset 10 045 comments in size has been created and made publicly available, which combines four open datasets and is supplemented by 3200 hand-tagged comments. A hierarchical classification scheme has been developed, including 16 classes grouped into 5 categories. Various classification methods were tested, including FastText with F1-macro = 0.579 and CodeBERT with F1-macro = 0.775, with an assessment by computational costs. The obtained results confirm the applicability of automatic review classification for identifying discussions that require increased attention from developers.