Fine-Tuning Pre-Trained Model to Extract Undesired Behaviors from App Reviews

Mobile application markets usually enact policies to describe in detail the minimum requirements that an application should comply with. User comments on mobile applications contain a large amount of information that can be used to find out APP's violations of market policies in a cost-effective way. Existing state-of-the-art methods match user comments with the violations of market policies based on well-designed syntax rules, which however cannot well capture the semantics of user comments and cannot be generalized to the scenarios not covered by the rules. To address this issue, we propose an innovative method, UBC-BERT, to detect undesired behavior from user comments based on their semantics. By incorporating sentence embeddings with attention, we train a classification model for 21 groups of undesirable behaviors based on the fine-tuning of a pre-trained model BERT-BASE. The experimental results show that our solution outperforms the baseline solutions in terms of a higher precision(up to 60.5% more).