Improving emotion classification in e-commerce customer review analysis using GPT and meta‑ensemble deep learning technique for multilingual system

In recent years, the field of natural language processing has placed a significant emphasis on multilingual emotion classification, particularly within social media analysis. This study area is crucial as it enables the extraction of invaluable insights from user-generated content. The research introduces an innovative strategy for multilingual emotion classification, utilizing a stacked hybrid deep learning framework. This framework combines cutting-edge models like RoBERTa-LSTM, RoBERTa-GRU, RoBERTa-BiGRU, and RoBERTa-BiLSTM, fine-tuned with the Adam optimizer over 15 epochs using a batch size of 64. It effectively handles multiple languages, including English, Arabic, and French. The study tackles the complexities of dealing with imbalanced datasets in emotion classification, exploring oversampling and undersampling techniques, including methods like Random Under-Sampling (RUS), Synthetic Minority Over-sampling Technique (SMOTE), and Generative Pre-trained Transformer (GPT). Comprehensive assessment metrics evaluate emotion classification performance across languages and datasets, including accuracy, F1-score, ROC-AUC curve analysis, and Matthews Correlation Coefficient (MCC). The evaluation considers both with and without dataset preprocessing to analyze the impact of different techniques. Additionally, the study assesses memory usage and execution time, providing insights into resource efficiency. Utilizing a stacked hybrid deep learning approach in this novel methodology for multilingual emotion classification yields impressive results, serving as a valuable tool for organizations seeking to enhance performance and customer satisfaction by unveiling sentiment patterns and trends across languages and domains.