Stratify Mobile App Reviews: E-LDA Model Based on Hot "Entity" Discovery

Abstract

Recent literatures have illustrated approaches that can automatically extract informative content from noisy mobile app reviews, however the key information such as feature requests, bug reports etc., retrieved by these methods are still mixed and what users really care about the app remains unknown to developers. In this paper we propose a novel model SAR: Stratify App Reviews, providing developers information about users' real reaction toward apps. SAR stratifies informative reviews into different layers, grouping the reviews based on what users concern, and we also develop a method to compute the user general sentiment on each entity. The model performs user-oriented analytics from raw reviews by (i) first extracting entities from each review, identifying hot entities of the app that users mostly care about, (ii) then stratifying all the reviews into different layers according to hot entities with a four-layer Bayes probability method, (iii) and finally computing user sentiments on hot entities. We conduct experiments on three genres of apps i.e. Games, Social, and Media, the result shows that SAR could identify different hot entities with respect to the specific categories of apps, and accordingly, it can stratify relevant reviews into different layers, the sentiment value of each entity can also represent users' satisfaction well, we also compared the result with human analysis, with the similar accuracy, the SAR can speed up the overall analysis automatically. Our model can help developers quickly understand what entities of the app users mostly care about, and how do they react to these entities.