Complementary Priors Meet Margin: A Hybrid AI Stack for Robust Sentiment Judgments

Abstract

This paper develops an AI ensemble that fuses class-conditional generative priors with strong discriminative margins to achieve robust binary sentiment decisions on long-form movie reviews. Class-specific language models (n-gram and recurrent) provide likelihood-ratio evidence, while NB-SVM on reweighted n-grams and paragraph-level embeddings supply invariant, high-margin features; a calibrated geometric fusion reconciles these signals without batching or domain-specific heuristics. Evaluated on the canonical IMDB split, the hybrid stack delivers state-of-the-art accuracy for its era and demonstrates measurable complementarity: weaker standalone generative paths consistently boost the discriminative core when ensembled under tuned weights. The study details training recipes, feature construction, and ablations that clarify when and why generative likelihoods add value to modern discriminative NLP, positioning the method as a practical blueprint for text-understanding agents in resource-constrained settings.