Sentiment-Polarity Aware Rating Prediction: Advanced Model Architecture for Review Analysis

Abstract

Introduction: Online review platforms have established themselves as key vectors for disseminating opinions on products and services, significantly influencing the consumer decision-making process. However, predicting ratings from free text remains a complex task, especially in the absence of explicit ratings. Methods: This study proposes an innovative hybrid framework, based on deep learning, that combines sentiment analysis with automatic rating estimation to effectively address the challenge of rating prediction from free text. The proposed model integrates Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs), and leverages three types of word embeddings; GloVe, FastText, and Word2Vec; to capture the semantic nuances of user reviews. The framework was evaluated on benchmark datasets (Amazon and Yelp) with a rigorous validation of performance metrics. Results: The results demonstrate a significant improvement over existing approaches, with average gains of 10.6% in recall, 10.7% in precision, and 10.5% in F1 score, as well as a 7% reduction in root mean square error. The robustness of the model is confirmed by its stable performance across different datasets and embeddings. Discussion: Experimental results show that our approach significantly improves the accuracy of rating prediction, outperforming benchmark methods. This performance reinforces the usefulness of deep models for the automatic analysis of textual reviews, contributing to better decision support for users of online review platforms. Conclusion: This research provides a methodological breakthrough for the automatic analysis of online reviews, with practical implications for consumer decision support. Prospective applications include the extension of the model to other languages and the integration of multimodal contexts.