Multilingual Emotion Classification in E-Commerce Customer Reviews Using GPT and Deep Learning-Based Meta-Ensemble Model

Abstract

As the e-commerce platforms grew exponentially, the volume of multilingual customer reviews increased, indicating that sentiment analysis is a priceless tool for finding consumer sentiment, enhancing marketing strategies, and improving customer experience. Nevertheless, emotion classification in multilingual reviews is very hard, and for one causes the variability of the language, the ambiguity of the sentiment, hierarchical word dependencies, and class imbalance which can skew traditional models. In order to resolve such challenges, this paper introduces a T5-CapsNet meta-ensembles model, which combines the T5 transformer for context embedded feature extraction and with Capsule Networks (CapsNet) for hierarchical sentiment learning. Furthermore, the model is further enhanced by a GAN-based data augmentation technique which increases the number of minority class reviews in a dataset by adding synthetic minority class reviews in an effort to correct dataset imbalance and promote classification fairness. As a metaensemble fusion strategy, weighted voting and stacking ensemble learning are used to improve sentiment prediction by making good use of the advantages of T5 and CapsNet. Experimental evaluations on the Multilingual Amazon Reviews Corpus (MARC) confirm that the proposed model surpasses the best sentiment classifier to reach an accuracy of 97.56% and an F1-score of 95.5%. It turns out that this hybrid deep learning approach very well captures the complex sentiment structures, or to put it differently, the multilingual e-commerce sentiment analysis is largely benefited from such a hybrid deep learning approach. The findings from this study will be a foundation for building more advanced emotion classification models, that can assist in improving customer sentiment analysis, automated feedback systems as well as decision-making in global e-commerce ecosystems.