Optimizing Sentiment Classification Using Dynamic Weighted Stacking Ensemble of Pre-trained Models

In the era of information explosion, sentiment classification, as a crucial task in natural language processing, is widely applied across various fields, including e-commerce, media, government, and finance. This paper proposes a dynamic weighted stacking ensemble method, integrating three pre-trained models: NeZha, XLNet, and ERNIE.The approach leverages Bayesian optimization to dynamically adjust the weights of the models. Therefore, robustness and generalization in sentiment classification tasks are improved. Experiments conducted on the SMP2020 Weibo sentiment classification dataset and the ChnSentiCorp sentiment analysis dataset confirm the effectiveness of this method. Results show that the ensemble model significantly outperforms individual models and traditional ensemble methods in terms of classification accuracy and F1 score, particularly excelling in handling complex emotional expressions. This study provides a new solution for sentiment classification tasks and demonstrates the potential of ensemble learning to enhance model performance .