Enhancing Cross-Modal Alignment in Multimodal Sentiment Analysis via Prompt Learning

Multimodal sentiment analysis (MSA) aims to predict the sentiment expressed in paired images and texts. Cross-modal feature alignment is crucial for models to understand the context and extract complementary semantic features. However, most previous MSA tasks have shown deficiencies in aligning features across different modalities. Experimental evidence shows that prompt learning can effectively align features, and previous studies have applied prompt learning to MSA tasks, but only in an unimodal context. Applying prompt learning to multimodal feature alignment remains a challenge. This paper employs a multimodal sentiment analysis model based on alignment prompts (MSAPL). Our model generates text and image alignment prompts via the Kronecker Product, enhancing visual modality engagement and the correlation between graphical and textual data, thus enabling a better understanding of multimodal data. Simultaneously, it employs a multi-layer, stepwise learning approach to acquire textual and image features, progressively modeling stage-feature relationships for rich contextual learning. Our experiments on three public datasets demonstrate that our model consistently outperforms all baseline models.