| 江长斌,刘贝贝,金超男.基于大语言模型增强的高校网络舆情负面主题挖掘与演化分析[J].数字图书馆论坛,2026,22(6):26~38 |
| 基于大语言模型增强的高校网络舆情负面主题挖掘与演化分析 |
| Practical Experience and Insights from Health Information Services for Vulnerable Groups in American Public Libraries |
| 投稿时间：2026-04-10 |
| DOI：10.3772/j.issn.1673–2286.2026.06.003 |
| 中文关键词: 高校网络舆情；负面主题；大语言模型；主题增强；BERTopic；主题识别；演化分析 |
| 英文关键词: University Online Public Opinion; Negative Topics; Large Language Models; Topic Enhancement; BERTopic; Topic Identification; Evolution Analysis |
| 基金项目:本研究得到国家社会科学基金项目“大数据视域下‘隐形’政治舆情演化规律及治理路径研究”（编号：19BSH013）资助。 |
| 作者 | 单位 | | 江长斌 | 武汉理工大学管理学院 | | 刘贝贝 | 武汉理工大学管理学院 | | 金超男 | 武汉理工大学管理学院 |
|
| 摘要点击次数: 9 |
| 全文下载次数: 12 |
| 中文摘要: |
| 为应对高校网络舆情负面主题快速发酵给高校舆情风险识别与治理工作带来的挑战，需要对舆情负面主题形成和演化规律进行分析。本研究旨在利用大语言模型的语义泛化与逻辑推理能力提升主题提取、识别结果的可读性，同时增强主题演化分析的可解释性。首先，基于舆情生命周期理论，采用SnowNLP算法筛选高校网络舆情中的负面文本；其次，借助大语言模型增强的BERTopic主题模型挖掘高校网络舆情不同阶段的负面主题；最后，结合负面主题的相似度，利用大语言模型优化演化路径的可解释性，生成负面主题演化路径。实证结果表明，大语言模型优化了高校网络舆情传播三个阶段负面主题标签，共得到52个可读性高、逻辑性强的主题词，以及7条负面主题阶段式演化摘要，深入揭示了高校网络舆情负面主题由事实关注向理性反思转变、由核心议题向多元议题演化的特征，为高校识别舆情风险、开展舆情治理提供决策支持。 |
| 英文摘要: |
| The rapid escalation of negative topics in university online public opinion poses challenges to public opinion risk identification and governance, making it necessary to analyze the formation and evolution patterns of negative topics in university online public opinion. The study aims to leverage the semantic generalization and logical reasoning capabilities of large language models to improve the readability of topic extraction and identification results while enhancing the interpretability of topic evolution analysis. First, based on public opinion life cycle theory, the SnowNLP algorithm is employed to filter negative texts from university online public opinion. Second, a BERTopic model enhanced by large language models is employed to mine negative topics at different stages of university online public opinion. Finally, by incorporating the similarities among negative topics, large language models are utilized to improve the interpretability of evolution paths and generate evolution paths of negative topics. The empirical results indicate that large language models optimize the labels of negative topics across the three stages of university online public opinion dissemination, yielding 52 highly readable and logically coherent topic keywords and seven stage-based evolution summaries of negative topics. The results further reveal that negative topics in university online public opinion evolve from factual concerns toward rational reflection and from core issues toward more diverse issues, thereby providing decision-making support for universities in identifying public opinion risks and conducting public opinion governance. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
