Enhanced Sentiment Analysis on Online Amazon Reviews Using RoBERTa with PSO-Based Hyperparameter Tuning

Abstract

Sentiment analysis stands vital for businesses that want to deliver superior customer satisfaction by optimizing their strategies. Rising dependence on digital platforms along with e-commerce leads consumers to use online reviews for their purchase choices. This work focuses on the customer product reviews and incorporates classical ML approaches, as well as DL models using the transformer coupling with optimization methods. The study creates a better sentiment analysis algorithm that uses RoBERTa to read product reviews on Amazon in Kaggle which has 34,000 records with rating information and user comments. The strategy to carry out this project necessitates an ample data preparation cycle as well as pre-lemmatization prior to the TF-IDF features operations. The dataset is organized into a training and testing partition approach and RoBERTa performs classification since it is a powerful self-attention mechanism and is pretrained using masked language modeling. Hyperparameter tuning that is implemented with Particle Swarm Optimization (PSO) is an advantage of the model optimization. Both confusion matrix and epoch-based performance plots display 85.31% accuracy and 92.26% precision and 85.31% recall and F-1 values. The results show that whereas RF, LR, Naive Bayes, and DT all perform poorly, RoBERTa beats the other fundamental models in terms of sentiment classification accuracy and F1 Score value. The evaluation finds that RoBERTa excels at generalizing and predicting sentiment which positions it as an effective model for sentiment analysis applications in practical use.