Abstract

Customer feedback, in the form of user ratings and reviews, enables businesses to gain insights about their customers and their product’s performance. The text analysis of these reviews, however, can be time-consuming and labor-intensive, therefore many companies choose to outsource the analysis to a third party. This project proposes to discover what customers are saying in mobile app reviews, by implementing a supervised feature-based reclassification of star rating to produce more accurate sentiment for the positive, negative, and neutral reviews; and the discovery of topics using the three topic models Latent Dirichlet Allocation (LDA), Gibbs Sampling for Dirichlet Mixed Model (GSDMM), and Embedded Topic Model (ETM). As a special application in this project, sarcasm in the review text is also examined. LDA and GSDMM achieved higher coherence scores but supplementing the discoveries with ETM may improve context in the topics. The topics are visualized using topic tables and word clouds.