With the explosive growth of e-commerce and User-Generated Content (UGC), online reviews are critical for understanding consumer preferences and product issues. Yet their multi-dimensional, colloquial, domain-specific traits challenge traditional methods, while deep learning has high costs and low interpretability. This study aims to build a framework balancing insight depth and practicality—to capture dynamic customer priorities across periods and products, and offer actionable insights for pharmaceutical e-commerce platforms.
Using anti-infective drugs as an example and online reviews from JD Pharmacy, we integrated the Latent Dirichlet Allocation (LDA) model with multiple tools (perplexity, coherence, and pyLDAvis) to extract influencing factors with higher precision. We then constructed and refined a domain-adapted sentiment lexicon for accurate sentiment quantification. Finally, we combined a multi-attribute model with PROMETHEE-II to ensure a reliable ranking of factor importance across different dosage forms and periods.
The framework successfully extracted period-specific topics and significantly improved sentiment classification accuracy by 19.38 percentage points compared to traditional methods (p < 0.05). Consumer priorities exhibited significant two-dimensional dynamics (period × dosage form), and the priority rankings were proven statistically robust across all independent scenarios via Bootstrap resampling and non-parametric tests (p < 0.05). Based on these findings, targeted suggestions for improving satisfaction were proposed.
This study systematically constructs an integrated text mining framework embedded with a dual-dimensional contextual logic. It establishes a standardized, low-cost, interpretable, and decision-closed-loop paradigm for UGC analysis, achieves the accurate quantification of consumers' dynamic priorities across different contexts, and provides a reliable pathway for vertical domain insights under resource constraints.
