Multimodel fusion sentiment analysis based on adversarial training and contrastive learning

This study aims to address challenges in sentiment text classification, such as emotional shifts, contextual dual meanings, and the limitations of traditional models in robustness and feature extraction. We propose a neural network model that integrates Bert, BiLSTM, TextCNN, and Attention mechanisms, along with adversarial training and contrastive learning methods. Bert is used for word embedding, BiLSTM captures sentence dependencies, TextCNN extracts local features, and Attention mechanisms finely adjust semantic relationships. Fast Gradient Method (FGM) enhances robustness, and a Dual Contrastive Learning (DualCL) strategy improves feature representation. Experimental results show that the model significantly outperforms traditional methods and other deep learning models in accuracy and robustness.