An Improved Weighted Average Ensemble Method for Textual Sentiment Analysis with Transformer and FNet

To address the limitations of existing sentiment analysis methods—specifically, the suboptimal performance of equal-weight ensemble averaging—this paper proposes an improved weighted average ensemble method for textual sentiment analysis. The ensemble method determines model weights by averaging validation accuracies over the final ten training epochs (41–50). Critically, the weights are normalized to ensure that the fused probability lies in [0,1]. The Transformer and FNet models are fused as base classifiers, leveraging Transformer’s global semantic capture capability and FNet’s high efficiency in sequence processing. Experimental results on a dataset of 8000 online comments related to the employment prospects of text data-related majors demonstrate that the proposed method achieves an AUC of 92.4%, an accuracy of 91.2%, and an F1-score of 91.0%. The method also achieves an inference speed of 32 samples/son CPU, offering a favorable trade-off between performance and efficiency for resource-constrained scenarios.