A Comprehensive Investigation on Multimodal Sentiment Analysis with Emoji

Abstract

: Nowadays, with the integration of multiple information carriers into daily communication, emotional transmission shows a trend of cross-modal fusion. Multimodal sentiment analysis has become a cutting-edge direction in Artificial Intelligence (AI). According to the study, there are three basic categories into which the current mainstream models can be divided: Attention-based multimodal model, Contrastive Learning-based Cross-modal Alignment model Model and Hybrid deep learning model. These models have significantly improved the classification accuracy in the joint sentiment analysis of emojis and texts by integrating deep learning architectures. Compared with traditional single-modal machine learning or deep learning models, the new multi-modal architecture shows stronger adaptability in capturing the implicit emotions of emojis and handling the complementarity of modal information. Furthermore, this paper delves deeply into the challenges and limitations faced by emoji multimodal sentiment analysis, including issues such as model interpretability, data annotation bias, and cross-cultural semantic differences. In response to these bottlenecks, the research further proposes future directions, e.g. generating balanced sentiment text data by CycleGAN, enhancing the interpretability of the model by using Shapley value method. Through technical sorting and bottleneck analysis, this review provides a path reference for the model optimization and scene implementation of emoji multimodal sentiment analysis.