\(A^{2} H^{2}\) for multimodal emotional data analysis

期刊论文
发表日期：

2025年8月30日

查看全文

Jun Wu,
Jinyu Liu,
Tianfeng Zhang,
Shuai Guo,
Yu Chen,
Jiahui Huang,
Fang Deng

摘要

As multimodal sentiment analysis technology continues to evolve, it is increasingly important to accurately capture and interpret sentiment information from different modalities. However, the traditional attention mechanism model employs a fixed number of attention heads across different attention levels, which makes it difficult to flexibly capture feature information at varying levels in text. This paper has proposed a deformable improvement strategy, comprising two key elements: Adaptive Attention Hierarchy and Head Optional Strategy, collectively referred to as the \(A^{2} H^{2}\) strategy. The experimental results demonstrate that the sentiment analysis model shows improved performance, with an accuracy of 87.97% using the \(A^{2} H^{2}\) strategy. In comparison to the prevailing sentiment analysis models, this deformable improvement strategy not only enhances the accuracy rate, but also reduces the information redundancy, improves the extraction efficiency of important features, and maintains the computational efficiency of the model in large-scale tasks. The codes are available at https://github.com/mmm587/A2H2.