State-of-the-Art Machine Learning Techniques in Sentiment Analysis for Social Media

Abstract

This review examines the evolution of machine learning for social media sentiment analysis, highlighting its transition from traditional methods, such as SVM and Naïve Bayes, to modern deep learning and transformer-based models. This evolution has been characterized by the use of contextual embeddings, hybrid neural networks, and the analysis of non-textual cues, such as emojis, which together have enhanced classification accuracy in complex social media settings. The most advanced hybrid architectures, fine-tuned with algorithmic search, have reached top-tier performance with accuracy and F1-scores approaching 97% on large Twitter datasets. Other successful models, which combine text and images with attention mechanisms or use transfer learning and automated feature selection, achieve F1-scores from 0.57 to 0.83 on difficult multilingual benchmarks. This progress demonstrates that successfully analyzing global social media data depends on integrating multiple data types, using sophisticated preprocessing, and creating adaptive models.