Sentiment analysis of vaccine data using enhanced deep learning algorithms

Abstract

This paper investigates and experiments with an approach to improve sentiment analysis on vaccine datasets with deep learning. It evaluates random forest (RF), naïve Bayes (NB), and recurrent neural network (RNN) models across a variety of configurations, i.e., vector dimensions, pooling techniques, as well as evaluation methods, hierarchical SoftMax vs negative sampling. The results show that the model we proposed prevailed with an accuracy of 99.05% on a learning rate equal to 0.001, outperforming all other models based on metrics including precision, recall, and F1-score for benign/malignant cases. The results suggest that higher vector dimensions, average pooling, lowering the dropout rate, and employing hierarchical SoftMax for output significantly improve model performance. Hierarchical SoftMax performs better than negative sampling, whereas a lower dropout rate decreases overfitting and leads to improved generalization. Our results demonstrate the necessity to apply more sophisticated deep-learning tools around capturing nuances of public vaccine-related sentiment, which may be crucial for informing communication strategies and supporting decision-making in a real-world health emergency. The findings indicate that the performance of sentiment analysis with regard to COVID-19 vaccine deployment policy design and public monitoring could be enhanced by advanced deep learning algorithms.