Determining Functional User Comments of Mobile Applications with Semi-Supervised Learning

Abstract

We can group user comments for mobile applications into two categories: First, only mentioning likes or dislikes, second having detailed feedback and valuable requests that steer the development path of the application. The developers count and weigh the latter comments; however, they are very few. In this study, we focus on discriminating these two groups of comments. In order to solve the problem that includes a vast amount of data with tremendous labeling costs, we conduct experimental studies on word embeddings, classifier selection, supervised and semi-supervised approaches. Results show that semi-supervised methods dramatically boost performance, especially for the class with too few samples. We favor FastText over BERT for the classification performance versus computational complexity dilemma in word embeddings. Our study also shows that we can successfully transfer the learning model for categorization between various mobile application markets.