A Cross-Modal Sarcasm Sentiment Description Generation Model Based on Multiscale Inception Attention Network and BERT

To address the issue of inaccurate and incomplete reflection of sarcasm sentiment in generated sentences, which arises from insufficient visual information in image-based sarcasm sentiment description generation, we propose a cross-modal sarcasm feature extraction network based on a Multiscale Inception Attention Network (MIANet) and BERT. The network employs an encoder-decoder architecture, first accurately capturing image features through the Multiscale Inception Attention Network, followed by decoding the extracted features using BERT. By integrating both detailed and global semantic information from the image, the network generates corresponding descriptions. This approach not only enhances the model's ability to process multi-modal data but also improves its performance in extracting complex semantic features. Experimental results demonstrate that the proposed method outperforms existing approaches on multiple public sentiment analysis datasets, achieving superior performance across various evaluation metrics, thereby validating the effectiveness and superiority of the proposed approach.