Extracting Arguments Based on User Decisions in App Reviews

Abstract

Review mining from app marketplaces has gained immense popularity from researchers in recent years. Most studies in this area, however, tend to focus on improving the performance of classification prediction. In this study, we consider review mining from a different perspective, that is, mining user actions/decisions along with their respective arguments/reasons. Our motivation is to obtain a deeper understanding of users' decisions regarding applications and their underlying justifications, e.g., why users give ratings or recommendations. These information abstractions can benefit app developers, especially in planning app updates, by providing data-driven requirements from users' points of view. We utilized a supervised learning approach and built a machine-based annotator to set the ground truth. Seven classifiers and different feature configurations were trained and evaluated on two app review datasets. We then extracted relations between user decisions and arguments based on functional and nonfunctional requirement attributes. The results show an improved performance over the results of the baselines and favorably acceptable performance compared to the results from a human assessment.