SEMANTIC-SBERT: LEVERAGING SENTENCE TRANSFORMERS FOR GRANULAR PRODUCT REVIEW SENTIMENT ANALYSIS

Abstract

The rapid expansion of the global e-commerce ecosystem, projected to exceed USD 7 trillion by 2030, has significantly increased the influence of customer reviews on consumer decision-making, with a majority of users relying on online feedback before making purchases. However, extracting meaningful insights from large-scale review data remains challenging due to its volume, variability, and timesensitive nature. Manual evaluation is inefficient and inconsistent, while many existing approaches struggle to capture contextual semantics and often perform poorly on imbalanced datasets. To address these challenges, this study proposes an advanced framework based on Natural Language Processing (NLP) for automated sentiment classification of product reviews using annotated datasets. The workflow begins with structured preprocessing and Exploratory Data Analysis (EDA) to clean and analyze the data for meaningful patterns. Sentence Bidirectional Encoder Representations from Transformers (SBERT) is used to generate context-aware embeddings that effectively capture semantic relationships. To handle class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied to create a balanced training distribution, improving model robustness. In contrast to conventional models such as Random Forest Classifier (RFC), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGB), the proposed framework integrates Deep Neural Network (DNN)-based feature extraction with a Boosted Rules Classifier (BRC) to enhance both prediction accuracy and interpretability. The system classifies reviews into Negative, Neutral, and Positive sentiments, enabling a deeper understanding of customer opinions. Additionally, the trained model is deployed using Django, a Python-based web framework, allowing users to upload data, generate predictions in real time, and manage interactions through an intuitive interface. Experimental results demonstrate that the proposed system achieves higher accuracy, improved scalability, and reduced bias, making it a reliable solution for extracting actionable insights and supporting data-driven decision-making.