An automated approach to aspect-based sentiment analysis of apps reviews using machine and deep learning

Apps reviews hold a huge amount of informative user feedback that may be used to assist software practitioners in better understanding users’ needs, identify issues related to quality, such as privacy concerns and low efficiency, and evaluate the perceived users’ satisfaction with the app features. One way to efficiently extract this information is by using Aspect-Based Sentiment Analysis (ABSA). The role of ABSA of apps reviews is to identify all app’s aspects being reviewed and assign a sentiment polarity towards each aspect. This paper aims to build ABSA models using supervised Machine Learning (ML) and Deep Learning (DL) approaches. Our automated technique is intended to (1) identify the most useful and effective text-representation and task-specific features in both Aspect Category Detection (ACD) and Aspect Category Polarity, (2) empirically investigate the performance of conventional ML models when utilized for ABSA task of apps reviews, and (3) empirically compare the performance of ML models and DL models in the context of ABSA task. We built the models using different algorithms/architectures and performed hyper-parameters tuning. In addition, we extracted a set of relevant features for the ML models and performed an ablation study to analyze their contribution to the performance. Our empirical study showed that the ML model trained using Logistic Regression algorithm and BERT embeddings outperformed the other models. Although ML outperformed DL, DL models do not require hand-crafted features and they allow for a better learning of features when trained with more data.