A multimodal meme sentiment analysis method based on contrastive method with metaphorical information
Yuxuan Sun, Sheng ua Gao, Yufan Jiao, Zhehuan Zhao, Linlin Tian, Bo Xu · 2025
In today’s network environment, memes have become one of the most prominent means of communication and are ubiquitous on social platforms such as Twitter and Facebook. As the popularity of memes continues to grow, the study of metaphorical information in memes has gained momentum. However, metaphorical information is often overlooked in meme analysis tasks. Additionally, the small size of the meme dataset and the imbalanced distribution of data labels pose significant challenges to existing methods for achieving satisfactory performance in a timely manner. To address these issues, we propose a contrastive method framework that incorporates metaphorical features for performing sentiment classification of memes. Our approach involves the fusion of encoded multimodal visual and textual features, followed by a contrastive method to learn the associations between memes and their corresponding sentiment information. Our experiments demonstrate the superior performance of our proposed method in achieving better performance even with limited training epochs.