多模型融合的跨境电商用户评论细粒度情感归因方法研究——以TF-IDF、LDA与TextRank-VADER为例

A Multi-Model Fusion Approach to Fine-Grained Sentiment Attribution in Cross-Border E-Commerce User Reviews—Taking TF-IDF, LDA, and TextRank-VADER as Examples

摘要: 随着跨境电商用户评论数据的规模急剧增长，如何从海量非结构化文本中实现高效、精准的情感归因，已成为文本挖掘领域的重要研究议题。现有研究多依赖单一模型进行关键词提取、主题发现或情感判断，难以形成从表层特征到深层归因的完整分析闭环，且缺乏多方法交叉验证机制，分析结论的可靠性受限。本文构建了一种基于TF-IDF-LDA与TextRank-VADER的多模型融合分析框架，通过“高频特征提取–隐含主题识别–细粒度情感归因”的递进式分析链条，实现从离散词汇到语义主题，再到情感驱动因素的深度挖掘。以Temu平台北美市场浴巾品类10,813条用户评论为实证场景，验证了该框架在跨境电商用户评论分析中的有效性、互补性与稳健性。研究结果表明：TF-IDF精准定位核心评价维度，LDA揭示用户需求的层次结构与优先级差异，TextRank-VADER将情感分析从整体极性判断推进至细粒度归因层面；三种方法独立运行却高度一致，形成完整的验证闭环。该框架具备良好的场景迁移性与模块可扩展性，为跨境电商用户评论的智能化分析提供了可操作的方法论范式。

Abstract: With the explosive growth of user review data in cross-border e-commerce, achieving efficient and accurate sentiment attribution from massive unstructured texts has become a critical research issue in text mining. Existing studies predominantly rely on single models for keyword extraction, topic discovery, or sentiment classification, making it difficult to form a complete analytical closed loop from surface features to deep attribution, and the absence of cross-validation mechanisms across multiple methods limits the reliability of analytical conclusions. This paper constructs a multi-model fusion analysis framework based on TF-IDF-LDA and TextRank-VADER, implementing a progressive analytical chain of “high-frequency feature extraction-latent topic identification-fine-grained sentiment attribution” to achieve deep mining from discrete vocabulary to semantic themes and further to emotional driving factors. Taking 10,813 user reviews of towel products on the Temu platform in the North American market as an empirical scenario, the effectiveness, complementarity, and robustness of the framework in cross-border e-commerce user review analysis are validated. The results indicate that: TF-IDF precisely locates core evaluation dimensions, LDA reveals the hierarchical structure and priority differences of user needs, and TextRank-VADER advances sentiment analysis from overall polarity judgment to fine-grained attribution; the three methods operate independently yet yield highly consistent conclusions, forming a complete verification closed loop. The framework demonstrates strong scenario transferability and modular scalability, providing an actionable methodological paradigm for intelligent analysis of cross-border e-commerce user reviews.