.png){kind=icon}
.png){kind=icon}
Return
Multimodal progressive contrastive learning for sentiment analysis
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.132033.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Multimodal sentiment analysis (MSA) seeks to infer human sentiments by integrating diverse modalities. However, the rapid advancement of natural language processing has created a significant modality imbalance, where models extract richer information from text than from other modalities. This disparity, compounded by the inherent noise in non-textual data and limited data availability, often leads to suboptimal feature representations and challenging fusion scenarios. To address these issues, we propose the Multimodal Progressive Contrastive (MPC) learning model. The core idea of MPC is to first construct a highly discriminative textual feature space and then leverage it to guide and enhance the representations of weaker modalities before fusion. Specifically, we introduce a multi-level contrastive learning paradigm that employs hierarchically structured hard positive samples from different layers of a pre-trained language model and progressively challenging hard negative samples from other modalities. Subsequently, a conditional transform-based feature enhancement module is utilized to improve the features of non-textual modalities, facilitating a more balanced multimodal interaction. Finally, a sentiment-aware fusion module effectively integrates sentiment information from the diverse feature spaces. Extensive experiments on public benchmark datasets demonstrate the advanced performance of the proposed MPC model.