Fine-Grained Sentiment Analysis of Public Opinion Videos Based on Conformer and Multi-Layered Interaction Attention

With the rapid development of internet technologies and the widespread adoption of smart devices, social media platforms have become significant channels for information dissemination and public sentiment expression. In particular, new media formats such as short videos have shown a substantial impact on public opinion guidance and emotional transmission, making sentiment analysis of short video content highly meaningful. However, existing research has limitations in modality interactions, often employing weighted summation or self-attention mechanisms for deep fusion of extracted features. These approaches fail to fully account for the complex local dependencies and hierarchical structures among modalities. To address these issues, this paper proposes a fine-grained sentiment analysis model for public opinion videos based on Conformer and multi-layered interaction attention mechanisms, termed DW-MIACon. The model first utilizes DeBERTa, CLIP, and Wav2Vec models to extract features from text, images, and audio, respectively. Subsequently, the extracted multimodal features are fused using a Dynamic Weighted Multi-layered Interaction Attention (DW-MIA) mechanism, generating rich fusion feature representations. Finally, a Conformer model is employed to deeply integrate the fused features, capturing complex interactions and local dependencies between modalities. Experimental results demonstrate that the proposed model significantly outperforms existing approaches in multimodal sentiment recognition tasks, notably enhancing the accuracy of fine-grained sentiment classification and the ability to identify subtle emotional nuances.