基于大语言模型的电商用户评论结构化分析方法研究

Research on Structured Analysis Method of E-Commerce User Reviews Based on Large Language Model

摘要: 电商评论数据挖掘对于提升运营效率和用户满意度具有实际应用价值。针对海量电商评论口语化、碎片化的特点，以及传统关键词匹配方法语义理解不足、主题模型结果晦涩难懂的缺陷，本文提出预训练语言模型PLM与大语言模型LLM深度协同的结构化分析方法，构建了从非结构化文本到结构化决策支持的自动化流程。该方法经BERT语义提取、主成分分析PCA降维、K-Means聚类得到各主题代表性样本，将这些样本作为少量示例融入提示词工程输入DeepSeek模型，让LLM适配电商评论表述特点，提升问题提取与结构化输出准确性。基于Amazon商品评论数据集的实验验证表明，该协同方法在聚类质量、问题提取精度及结构化输出效果上均优于传统基线方法，其输出成果可直接为电商运营决策提供数据支撑，具备良好的实用价值。

Abstract: E-commerce review data mining has practical application value for improving operational efficiency and user satisfaction. Addressing the characteristics of massive e-commerce reviews being colloquial and fragmented, and the shortcomings of traditional keyword matching methods such as insufficient semantic understanding and obscure topic model results, this paper proposes a structured analysis method that deeply collaborates with a pre-trained language model (PLM) and a large language model (LLM), constructing an automated process from unstructured text to structured decision support. This method uses BERT semantic extraction, principal component analysis (PCA) dimensionality reduction, and K-Means clustering to obtain representative samples for each topic. These samples are then used as a small set of examples and input into the DeepSeek model via prompt word engineering, helping the LLM accurately adapt to the characteristics of e-commerce review expressions and improving the accuracy of question extraction and structured output. Experimental validation based on the Amazon product review dataset shows that this collaborative method outperforms traditional baseline methods in clustering quality, question extraction accuracy, and structured output performance. Its output can directly provide data support for e-commerce operational decisions and have significant practical value.