Data Science and Management | 2026
Authors: Hicham N.; Nassera H.
DOI: 10.1016/j.dsm.2025.06.003
Journal: Data Science and Management
Year: 2026
Publisher: KeAi Communications Co.
Document Type: Article
Open Access: All Open Access; Gold Open Access
Cited by: 1
This study introduces stacked deep learning for multilingual opinion mining. This framework incorporates RoBERTa-GRU, RoBERTa-LSTM, RoBERTa-BiGRU, and RoBERTa-BiLSTM hybrid models, optimized using the Adam optimizer. The methodology can handle three languages: French, English, and Arabic. This research explores the challenges of imbalanced datasets in opinion mining, employing oversampling approaches like advanced easier data augmentation, synthetic minority over-sampling technique, and generative pre-trained transformer to balance the datasets and improve classification efficiency. We used the assessment metrics of Cohen’s kappa, the receiver operating characteristic-area under curve, accuracy, and mattheus correlation coefficient, along with k-fold validation, to evaluate the sentiment analysis performance across three languages and six datasets. Moreover, we computed performance metrics for all models while scaling the dataset for training and testing. We also determined the memory usage and execution time for each model. Utilizing a stacked deep learning algorithm, the suggested methodology for multilingual opinion mining demonstrated high efficacy in extracting meaningful insights from social media data across several languages. The technique produced significant outcomes, rendering it a potentially valuable instrument for enhancing performance and customer satisfaction by identifying patterns and trends in public sentiment. © 2025 Xi’an Jiaotong University. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC BY-NC-ND license. http://creativecommons.org/licenses/by-nc-nd/4.0/
GPT; Imbalanced datasets; Natural language processing (NLP); Stacking deep learning