Bimodal Sentiment Analysis Based on a Pre-Trained Model and Masked Attention Fusion

In recent years, sentiment analysis based on deep learning and the fusion of multimodal data has become a research focus in the field of natural language processing (NLP). However, previous studies have proposed neural networks with complex internal designs but without a corresponding performance improvement. Additionally, the lack of effective fusion and interaction between modalities has limited the ability of these models to make sentiment decisions. In this study, we introduces a bimodal sentiment analysis model, GTA-BERT, which is based on pre-trained models and masked attention fusion to tackle these problems. This model leverages technologies such as RoBERTa, HuBERT, the graph convolutional network (GCN), and the bidirectional gated recurrent unit (BiGRU) to extract audio features and enhance text semantic features. Subsequently, it dynamically adjusts the attention weights for different modalities using masked attention to achieve optimal sentiment decisions. Finally, it combines the masked attention matrix with the sentence dependency parsing information retained by the GCN for classification. Experimental evaluations of this model are conducted on three publicly available Chinese and English datasets, and it is found to outperform mainstream models. The research findings can contribute to sentiment analysis decision-making in fields such as personalized recommendations, psychological counseling, and intelligent chat systems.