V. AUTOMATED CLASSIFICATION OF COMMENT USEFULNESS
In the second phase of the study, we built upon the findings from the interviews to create an automated classifier to distinguish between useful and not useful comments. The purpose behind this is that an accurate automatic classifier allows us to classify a large number of review comments, enabling a large scale quantitative study of review comment effectiveness (as we conduct and describe in Section VI), as manual classification of code review comments is time consuming and does not scale. Note that our goal in building a classifier is not to predict comment usefulness at the time that the comment is written. Such a classifier would not be useful to developers (an author of a change under review likely wants