Enhanced Sentiment Analysis of Online Reviews Using Fine-Tuned RoBERTa with Particle Swarm Optimization-Based Hyperparameter Tuning

Abstract

Sentiment analysis of online reviews has become a critical component of intelligent decision-making systems in e-commerce, hospitality, fintech, healthcare, and public governance. The exponential growth of user-generated content across digital platforms demands high-precision, scalable, and context-aware models capable of capturing nuanced linguistic expressions, sarcasm, implicit polarity shifts, and domain-specific sentiment patterns. Transformer-based architectures, particularly RoBERTa, have demonstrated superior contextual representation learning capabilities compared to conventional deep learning approaches. However, the performance of fine-tuned transformer models is significantly influenced by hyperparameter configurations, and manual or grid-based tuning methods often fail to explore the solution space efficiently.