RoBERTa beats classic ML with 93% accuracy in sentiment analysis study

A new research paper, ‘From TF-IDF to Transformers: A Comparative and Ensemble Approach to Sentiment Classification’, systematically evaluates seven machine learning and deep learning models for binary sentiment analysis on the IMDb movie review dataset. The study, accepted at the International Conference on Intelligent Computing, Networks and Security (IC-ICNS 2026), covers traditional models like Naive Bayes, Logistic Regression, and SVM, alongside tree-based LightGBM, sequential LSTM, and transformer-based DistilBERT and RoBERTa. Using standard NLP preprocessing, the team measured accuracy, precision, recall, F1-score, and ROC-AUC to identify the best performer.

RoBERTa emerged as the top individual model with 93.02% accuracy, significantly outperforming classical approaches (SVM at ~89%) and lightweight transformers like DistilBERT (~91%). Notably, the authors implemented a soft voting ensemble that combined predictions from all models, which yielded even better classification metrics, demonstrating that ensemble strategies can squeeze additional performance from diverse architectures. The paper underscores that while transformer models excel at capturing contextual nuances, combining them with simpler models can create more robust systems for real-world opinion mining tasks. This work provides a practical benchmark for developers choosing between efficiency and accuracy in sentiment pipelines.