BERT-RCNN: an Automatic Classification of App Reviews using Transfer Learning based RCNN Deep Model

Abstract

App reviews considered as the major concern over the internet. It influences the user’s mind before purchasing the app. However, such user reviews might contain technical information about the app that can be valuable for the developers and software companies. Due to pervasive use of mobile apps, large amount of data is created by users on daily basis. Manual identification and classification of such reviews is very time-consuming and laborious task. Hence, automating this process is essential for helping developers in managing these reviews efficiently. Recently, traditional machine learning and deep learning oriented approaches have been exploited for classification of app review into software requirements. Most of the machine learning techniques utilizes traditional word techniques to extract the features from the textual reviews. In addition, deep learning techniques with efficient word embedding technique also improve the classification performance of the model. However, it is found that these techniques suffer from polysemic problem. To address aforementioned problem, this research paper presents novel neural language based framework, BERT-RCNN to classify app reviews into specified categories of functional requirements such as Bug report, Feature Request and Relevant etc. The presented framework has two main modules: Bidirectional Encoder Representation for Transformer (BERT) module and Recurrent-Convolutional Neural Network (RCNN) module. BERT module extracts the contextual relationship between the textual reviews. Then, these features are input to RCNN module to capture more important features with deep semantic information. The output of RCNN is fed to fully connected layer for classification purpose. Extensive experiments are conducted on five datasets to investigate the effectiveness of the presented BERT-RCNN model. For evaluation purpose, standard performance measures are used to obtain the results. Moreover, hyperparameter such as BiLSTM memory unit, CNN filters and CNN kernel size are tuned to ensure the quality of prediction. Comparative analysis of proposed model is conducted with existing state-of-the-art models. It shows that BERT-RCNN outperforms other state-of-the-art models with respect to precision, recall, f-measure and accuracy.