Comparative Analysis of Word Embedding and Topic Modeling Techniques for Sentiment Analysis on Large-Scale Customer Reviews

Abstract

With the massive explosion of digital consumer feedback, analyzing large-scale unstructured text has become crucial for e-commerce intelligence, market trend analysis and brand management. However, extracting clear, interpretable themes from high-volume, noisy text datasets remains a significant challenge due to the semantic limitations of traditional models and the computational overhead or potential overfitting of deep architectures. To address these challenges, this research provides a comprehensive comparative study of eight distinct word embedding and topic modeling architectures as viable programmatic solutions. We evaluate frequency-based models and dense embeddings on a dataset of 250,000 Amazon Fine Food reviews. Using the $C_{v}$ coherence metric, we analyze model performance across clustering techniques such as K-Means, LDA and BERTopic. Our results show that TF-IDF with LDA achieved the highest numerical coherence $(C_{v}=0.525$). Following the identification of TF-IDF with LDA as the optimal model, we employ the VADER framework to quantify sentiment polarity across its clusters. Our findings demonstrate that while transformer models provide superior semantic stability and discover finer, more granular subthemes for large-scale NLP tasks, traditional frequency-based LDA remains highly effective for structured coherence in large-scale food review domains, offering an agile, computationally efficient solution for real-world enterprise applications.