Deep Learning-Driven Text Sentiment Analysis: Research Progress, Challenges, and Future Trends in the Past Five Years

Text sentiment analysis has developed from a single text classification task into a complex multimodal fusion system that can achieve cross-scenario sentiment understanding with the help of deep learning technology. In this article, we will review the cutting-edge studies from 2020 to 2025 on domain adaptation, low-resource learning, complex semantic modeling, multimodal fusion, and ethical fairness based on review papers in leading journals (e. g., IEEE TPAMI, Nature) and at top conferences (e. g., ACL, NeurIPS). It has been found that the combination of domain knowledge graphs reduces the performance drop caused by cross-domains by 60%, causal reasoning increases the accuracy of mixed sentiment disentanglement by 12.4%, and multimodal dynamic alignment improves precision by 8% in conflicting scenarios. This article shows that technological development has a transversal and fused trend of "knowledge promotion-causal modelling-modal collaboration" and lightweight mechanism as well as fair mechanism are important in practical engineering implementation.