Aspect-based Sentiment Analysis with an Ensemble Learning Framework for Requirements Elicitation from App Reviews

Abstract

In the past three years, numerous studies have demonstrated the effective performance of aspect-based sentiment analysis (ABSA) in eliciting requirements from APP reviews. However, in aspect category detection (ACD) and aspect category polarity (ACP), traditional supervised machine learning techniques still dominate the latest research. Inherent limitations, such as poor generalization ability, low robustness, and high dependence on feature engineering, often constrain the methods. Additionally, while most research continues to center on the efficacy of singular models or strategies, the effective amalgamation of traditional and contemporary techniques has yet to be adequately explored. Given these challenges, this study proposes an ensemble learning framework based on XGBoost and Stacking. Compared to baseline, the framework achieves a performance improvement of 22.9%-28.4% in ACD and 9.3%-13.2% in ACP tasks based on different feature engineering. Overall, the preliminary attempt in this study to apply ensemble learning in ABSA for requirements elicitation from APP reviews, providing a feasible new direction for overcoming the inherent limitations of traditional techniques in fine-grained sentiment modeling in this field.