Exploring Transformer and Recurrent Neural Models for Sentiment Analysis of Mobile App Reviews

Sentiment analysis is one of the important aspects of discovering user opinions and enhancing the quality of products like mobile applications. This study assesses the effectiveness of six prominent Natural Language Processing (NLP) models-BERT, RoBERTa, XLNet, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)-in classifying sentiments from a large dataset of Google Play Store reviews. We collected 73,483 reviews from 45 apps, preprocessed the data, and employed two labeling methods: a pre-trained DistilBERT model and K-means clustering. Our results compare transformer-based models, BERT and XLNet, with traditional recurrent models for sentiment analysis of mobile application reviews. The results show that transformer-based models achieve higher accuracy, precision, recall, and F1 scores with DistilBERT pre-trained model labeling. However, no single model strikes a balance between recall and precision using K-means clustering labels. These findings suggest that the choice of labeling method, in addition to the use of advanced transformer models and embeddings, significantly affects model performance.