Enhancing Sentiment Classification of E-commerce Reviews for Actionable Insights using LLMs

This paper explores the use of Large Language Models (LLMs) and Natural Language Processing (NLP) techniques to enhance sentiment classification in e-commerce product reviews. We develop a pipeline that fine-tunes pre-trained transformer models, such as Bidirectional Encoder Representations from Transformers (BERT) and RoBERTa, for sentiment analysis, achieving high accuracy in multi-category classification. Additionally, we introduce a root-cause analysis model that leverages Deepseek-based labeling to classify low-rated reviews into predefined issue categories. By sequentially fine-tuning on a large dataset before domain-specific adaptation, we aim to improve model robustness and generalization. To capture evolving customer concerns over time, we apply a Poisson thinning method for temporal trend analysis of complaint categories. Our approach demonstrates the potential of LLMs for actionable insights in e-commerce review analysis.