A BERT based Ensemble Approach for Sentiment Classification of Customer Reviews and its Application to Nudge Marketing in e-Commerce

Abstract Product reviews are pivotal in guiding customer decisions, fostering trust, and driving loyalty. They offer insight into others’ perceptions, aiding in purchase choices. Understanding sentiments within reviews is crucial for e-commerce platforms, facilitating improved product understanding, positive consumer engagement, and enduring rela- tionships. Moreover, reviews present avenues for innovative marketing strategies, such as “Nudge Marketing,” subtly aiding customers in decision-making. This paper aims to construct a sentiment analysis model using deep learning and natural language processing techniques focused on specific product types on an e-commerce platform. Employing bi-directional LSTMs and BERT, we categorize opinions into positive, negative, or neutral sentiments. The study proposes stacking/ensembling methods, discovering superior prediction accuracy in sentiment classification using one of these approaches. Real-world customer review data from diverse e-commerce platforms serves as the basis for our experiments. Notably, the stacked ensemble method emerges as the most accurate in sentiment prediction. The paper illustrates how this model output can be leveraged in Nudge marketing, assisting customers in their purchase decisions. Furthermore, a k-armed Bandit Experiment employing Thompson sampling and Monte Carlo simulation explores the effectiveness of review-based Nudging Strategy. Results suggest the review-based nudge or badge as an effective standalone or combined strategy for influencing customer decisions.