Fine-Grained Sentiment Analysis Of Restaurant Reviews Using Latent Dirichlet Allocation And Hybrid Ensembles

Abstract

Traditional sentiment analysis methods frequently reduce customer reviews to binary or ternary classifications, ignoring the complex viewpoints present in multi-aspect assessments, which are particularly common in the restaurant industry. This work introduces a novel hybrid framework for Aspect-Oriented Sentiment Analysis (AOSA) that combines a voting-based ensemble learning classifier for sentiment prediction with Latent Dirichlet Allocation (LDA) for unsupervised aspect extraction. In order to determine sentiment polarity, the suggested model aligns review sentences with latent topics that correspond to review aspects like food, service, ambiance, and pricing. With a high classification accuracy and little reliance on annotated data, ensemble classifiers made up of Logistic Regression, SVM, and XGBoost produce reliable and understandable results. The model outperforms state-of-the-art techniques, especially in aspect-level granularity, sentiment precision, and human interpretability, according to experimental evaluations on the Yelp and SemEval-2016 datasets. Customer experience analysis, business intelligence, and decision support in the hospitality industry are just a few of the many uses for this framework.