Abstract
App reviews often reflect end-users’ requests, issues or suggestions for supporting app maintenance and evolution. Hence, researchers have evaluated several classification approaches for identifying and classifying such app reviews. However, these classification approaches are driven by manually derived taxonomies. This is a limitation given the burden of human involvement, numerous app reviews and dependency on the availability of domain knowledge to perform classification. In this study, we develop and evaluate a novel approach towards the automatic generation of a dynamic taxonomy that groups related app reviews. Our approach uses natural language processing, feature engineering and word sense disambiguation to automatically generate the taxonomy. In a pilot study, we validated the feasibility of our proposed approach with app reviews extracted from the popular My Tracks app, where outcomes revealed a 72% match with a manual taxonomy generated from domain knowledge provided by humans. We then extended the scope of this study by validating the application of the automated taxonomy generation approach on app reviews belonging to TradeMe and Flutter apps. The outcomes revealed 80% and 71% match with the manual taxonomy of the latter two apps. Thus, our approach shows promise for rapidly supporting software maintenance and evolution.