Attention-optimized vision-enhanced prompt learning for few-shot multi-modal sentiment analysis

To fulfill the explosion of multi-modal data, multi-modal sentiment analysis (MSA) emerged and attracted widespread attention. Unfortunately, conventional multi-modal research relies on large-scale datasets. On the one hand, collecting and annotating large-scale datasets is challenging and resource-intensive. On the other hand, the training on large-scale datasets also increases the research cost. However, the few-shot MSA (FMSA), which is proposed recently, requires only few samples for training. Therefore, in comparison, it is more practical and realistic. There have been approaches to investigating the prompt-based method in the field of FMSA, but they have not sufficiently considered or leveraged the information specificity of visual modality. Thus, we propose a vision-enhanced prompt-based model based on graph structure to better utilize vision information for fusion and collaboration in encoding and optimizing prompt representations. Specifically, we first design an aggregation-based multi-modal attention module. Then, based on this module and the biaffine attention, we construct a syntax–semantic dual-channel graph convolutional network to optimize the encoding of learnable prompts by understanding the vision-enhanced information in semantic and syntactic knowledge. Finally, we propose a collaboration-based optimization module based on the collaborative attention mechanism, which employs visual information to collaboratively optimize prompt representations. Extensive experiments conducted on both coarse-grained and fine-grained MSA datasets have demonstrated that our model significantly outperforms the baseline models.