Detecting Features Concerning Privacy From App Reviews
share
Privacy requirements not only relate to legal compliance but also influence user satisfaction. Massive rapidly increasing App reviews have been proved a valuable requirements knowledge repository. Existing studies on App reviews mining have made much effort to automatically extract various requirements related information, e.g., feature request, bug report, and user opinions. However, less attention has been paid to privacy requirements refinement based on user reviews mining, which is beneficial for addressing users' privacy concern. In this work, we aim to detect privacy related features from App reviews to facilitate software maintenance activities. To that end, we design a semi-automatic framework to identify privacy related reviews from which App features are extracted and mapped to those listed in App descriptions. Firstly, we combine information retrieval and supervised text classification to identify privacy related reviews. Then, we design a dependency parsing method to extract App features from those privacy related reviews. Finally, those automatically extracted features are matched with those manually annotated ones in App descriptions based on phrase similarity. We quantitatively evaluate the three components of our framework on the reviews of Apps from multiple categories. For privacy related reviews identification, Gradient Boosting classifier achieves the highest F1 score of 93.77% among other competitive algorithms including deep learning ones. On App feature extraction, our dependency parsing based method can achieve a recall of 85.63%, more than 20% higher than the baseline. For feature matching, the phrase embedding cosine similarity shows the best matching result among four types of similarity methods, obtaining an average accuracy of 57%. We finally discuss the potential applications of our framework in detecting feature problems that may cause privacy threats.
READ FULL TEXT