An Automatic Analysis of User Reviews Method for APP Evolution and Maintenance

Application distribution platforms such as Google Play Store or Apple App Store allow users to submit feedbacks to download applications in the form of ratings or reviews. These feedbacks can directly or indirectly reflect users' intention, and it can greatly help mobile developers (or app provider) to continuously maintain and improve their applications, such as fix the existing bugs, add or refining the app features, etc. and so as to better satisfying user expectations continuously. App reviews provide an opportunity to proactively collect user complaints and promptly improve apps' user experience, in terms of bug fixing and feature refinement. However, for many popular applications, since the large amount of user review data, unstructured review data, and inconsistent review quality, identifying the valuable review information becomes a challenging task. Therefore, classification of user reviews into specific topics and automated analysis to reduce the workload of manual analysis has become a new idea for app review mining analysis. In this paper, we propose a method named ARICA(Automatic Review Intention Classification Analysis) to automatically analyze crowd user reviews to efficiently provide developers with software maintenance and evolution suggestions. Firstly, ARICA classifies the reviews into different categories according to the user's feedbacks, and then uses the LDA topic model to classify the reviews under each user's intent category. This allows a preliminary screening of user reviews to obtain review information under each intent category. Secondly, ARICA clusters user views with similar semantic expressions under each review topic to further filter the redundant information in reviews, so that can easier and intuitive to understand the user's original feedback and capture the user's true intention more accurately. Afterwards, ARICA uses the sentiment analysis tool called SentiStrength to obtain user sentiment, and then analyzes the sentiment distribution of user reviews to identify the user's significant intentions. Finally, the multidimensional information such as user intentions and sentiment preferences are considered comprehensively for calculating the review score and then ARICA prioritizes reviews for realizing the opinions recommendation for the developers. We use real app review data from Google Play to verify the performance of review intent classification and sentence clustering of ARICA. The experimental results show that ARICA has precision of 80% in the process of user review intention classification, compared with the state-of-the-art existing automatic user intentions mining method TextCNN which based on Convolutional Neural Networks (CNN), the F-Measure of ARICA is improved by 19.1%. Meanwhile, ARICA achieves 86% of the precision during the clustering of review sentences, which provides effective support for subsequent developers to recommend app update tasks. Further, we use the official app changelog as a ground truth, and empirically analyzed whether our recommended user reviews can be truly adopted by developers, the results show that ARICA can efficiently help developers better understand the user's real requirements, which is of great significance for developers to perform subsequent app maintenance and evolution tasks. In addition, we also publish the original data set, the manually labeled data set and the source code of ARICA on github which provide materials for other relevant researchers. © 2020, Science Press. All right reserved.