Sentiment-aware multimodal pre-training for multimodal sentiment analysis

Pre-trained models, together with fine-tuning on downstream labeled datasets, have demonstrated great success in various tasks, including multimodal sentiment analysis. However, most most multimodal pre-trained models focus on learning general lexical and/or visual information, while ignoring sentiment signals. To address this problem, we propose a sentiment-aware multimodal pre-training (SMP) framework for multimodal sentiment analysis. In particular, we design a cross-modal contrastive learning module based on the interactions between visual and textual information, and introduce additional sentiment-aware pre-training objectives (e,g., fine-grained sentiment labeling) to capture fine-grained sentiment information from sentiment-rich datasets. We adopt two objectives (i.e., masked language modeling and masked auto-encoders) to capture semantic information from text and images. We conduct a series of experiments on sentence-level and target-oriented multimodal sentiment classification tasks, wherein the results of our SMP model exceeds the state-of-the-art results. Additionally, ablation studies and case studies are conducted to verify the effectiveness of our SMP model.