An Uncertainty-Aware Cross-Modal Graph Attention Model for Improving Aspect-Based Sentiment Analysis

Abstract

The significance of the user-generated content is crucial in influencing business strategies by capturing public sentiment across various platforms. In this context, Aspect-Based Sentiment Analysis (ABSA) is now considered a main technique to determine sentiment polarities related to definite product attributes or aspects. While early ABSA approaches relied solely on textual input, recent advancements have highlighted the benefits of incorporating multimodal data, particularly text and images to improve sentiment interpretation. However, existing models face challenges such as ineffective fusion strategies, loss of informative features, and increased computational cost due to complex architectures. To address these limitations, a novel framework Attention aware Graph Convolutional Multimodal Aspect based Sentiment analysis Network (AGCMASN) is proposed. This model introduces an Uncertainty-Aware Cross-Modal Graph Attention (UACMGA) module for enhanced multimodal aspect-based sentiment classification. Initially, the text features are extracted through Bidirectional Encoder Representations from Transformers (BERT) and the visual features are obtained through the Residual Network (ResNet-50) encoder. The UACMGA module then fuses these features while explicitly modelling uncertainty, which suppress noise and helps retain useful information during cross-modal interactions. A Graph Convolutional Network (GCN) is subsequently employed to predict sentiment polarities at the aspect level. The proposed model provides a strong answer to practical multimodal sentiment analysis problems, as shown by the experimental findings, which show that it achieves better performance in recall, accuracy, precision, and F1-score.