Using Active and Self-Training to Improve Sentiment Prediction on App Reviews

App marketplaces such as Google Play Store and Appstore provide a central platform for their users to download apps and submit their feedback in the form of ratings and textual reviews. In a textual review, users express their sentiments (positive, negative, and neutral) towards features delivered in an app. Summarizing users’ sentiments mentioned in user reviews of an app can help developers to understand user’s perception of their apps and hence improve the app quality accordingly. Predicting users’ sentiments mentioned in user reviews is the main task for generating such summaries automatically. Several sentiment analysis tools such as Stanford CoreNLP, NLTK, SentiStrength, and Spacy are available for predicting sentiments conveyed in a text but their performance is not satisfactory when used for the app review domain. In this context, we build a sentiment prediction model targeting the app review domain. As manual labeling of users’ sentiments in app reviews is an expensive task, we combined active learning and self-training to reduce the burden of manual labeling. Compared to the general-purpose off-the-shelf sentiment tools, our sentiment prediction model for app reviews has improved performance for negative, neural, and positive classes by 19%, 31%, and 3%, respectively.