RCBERT an Approach with Transfer Learning for App Reviews Classification

Abstract

AbstractNowadays, extracting valuable information from user reviews of mobile applications has been a critical channel to obtaining user requirements in the process of mobile application development. However, it’s difficult to extract accurately valuable information from massive reviews that are unstructured and uneven. Existing works classify the user reviews to assist the process of requirement elicitation. However, due to the differences in wording, sentence structure, and requirements of reviews, the generalization of existing methods is poor. Therefore, we propose a Review Classification using BERT model (RCBERT) for user reviews classification. RCBERT contains the great capability of generalization that can complete specific fine-tuning tasks only by providing less training data. We evaluate the performance on an English dataset with 4014 user reviews and a Chinese review dataset that contains 18331 user reviews. The experimental results show that the approach proposed in this paper improves the review classification accuracy effectively (F1-score of up to 88% on the English dataset and 91% on the Chinese dataset).KeywordsApplicationUser reviewsTransfer learningReviews classification