ASPECT-BASED SENTIMENT ANALYSIS USING N-GRAMS, THRESHOLD ADJUSTMENT, AND 3-D SENTIVALUES WITH A NAIVE BAYES ENSEMBLE

Abstract

Aspect-Based Sentiment Analysis (ABSA) supports fine-grained understanding of user opinions by identifying sentiment toward specific aspects. Nevertheless, many existing approaches either rely on rigid feature representations or adopt deep learning models that introduce high computational cost and limited interpretability, reducing their suitability for scalable soft computing systems. This study proposes a hybrid intelligence framework for ABSA that combines TF-IDF n-gram representations with three-dimensional lexicon-based sentiment values and a threshold-adjusted Naïve Bayes ensemble. Contextual information is captured using unigram, bigram, and trigram features, while semantic polarity, objectivity, and subjectivity scores derived from SentiWordNet provide complementary sentiment knowledge. A weighted fusion of Multinomial and Gaussian Naïve Bayes classifiers is employed, alongside adaptive threshold calibration to improve minority-class detection. Experiments on a large restaurant review dataset demonstrate that the proposed approach achieves an overall accuracy of 0.92 with strong macro-averaged and weighted F1-scores, outperforming multiple baseline and hybrid methods. Statistical significance is confirmed using the Wilcoxon signed-rank test. Computational complexity analysis shows linear scalability with respect to corpus size and document length. The results indicate that the proposed hybrid framework delivers an effective balance between accuracy, interpretability, and computational efficiency, making it suitable for scalable soft computing systems and resource-constrained sentiment analysis applications.