AI-Driven Ensemble for Enhanced Sentiment Polarity Detection in Movie Reviews

Abstract

This paper presents a weighted multi-model ensemble approach for discerning sentiment polarity in text documents, specifically consumer reviews. We address the binary classification problem of identifying positive versus negative sentiment by proposing a hybrid framework that integrates generative, discriminative, and deep embedding-based models. Our key contribution is a weighted voting mechanism that leverages cross-validation to assign model-specific weights, effectively harnessing the complementary strengths of its diverse constituents. This ensemble strategy is evaluated on widely recognized movie review datasets, where it demonstrates robust performance compared to standalone models. Our method achieves 93.1% accuracy on the IMDB dataset and 90.6% accuracy on the Rotten Tomatoes dataset. Our ensemble achieves improvements over individual models, with absolute accuracy gains of 1.6 percentage points on IMDB (93.1% vs.~91.5% for BERT) and 3.3 to 12.2 percentage points on Rotten Tomatoes. These results demonstrate the effectiveness of the proposed weighting strategy for sentiment classification.