Software Feature Request Detection in Issue Tracking Systems

Communication about requirements is often handled in issue tracking systems, especially in a distributed setting. As issue tracking systems also contain bug reports or programming tasks, the software feature requests of the users are often difficult to identify. This paper investigates natural language processing and machine learning features to detect software feature requests in natural language data of issue tracking systems. It compares traditional linguistic machine learning features, such as "bag of words", with more advanced features, such as subject-action-object, and evaluates combinations of machine learning features derived from the natural language and features taken from the issue tracking system meta-data. Our investigation shows that some combinations of machine learning features derived from natural language and the issue tracking system meta-data outperform traditional approaches. We show that issues or data fields (e.g. descriptions or comments), which contain software feature requests, can be identified reasonably well, but hardly the exact sentence. Finally, we show that the choice of machine learning algorithms should depend on the goal, e.g. maximization of the detection rate or balance between detection rate and precision. In addition, the paper contributes a double coded gold standard and an open-source implementation to further pursue this topic.