面向复杂网络舆情知识发现的事理图谱方法优化

Optimization of Event Evolutionary Graph Method for Complex Online Public Opinion Knowledge Discovery

[研究目的]优化事理图谱构建方法,提升事理图谱工具在非结构化网络舆情数据中的知识发现能力,能更好挖掘复杂网络舆情事件内部的因果关系和演化路径。[研究方法]研究采用RoBERTa预训练模型进行序列标注以取代传统模式匹配方法,引入Word2Vec词向量和BERTopic主题模型取代传统机器学习聚类算法,对知乎平台“硅谷银行破产”网络舆情进行实证分析。[研究结论]结果表明,融合深度学习与序列标注的因果关系抽取在114901个上下文中识别到68613条原始因果事件对,相较模式匹配方法高出46.47%;基于词向量与主题聚类模型的事件泛化将2148个代表事件划归为14个主题,在文本特征和语义特征层面的泛化效果优于K-means算法。该文依据优化方法构建的网络舆情事理图谱围绕核心主题呈现“循环型”“紧密型”“长链型”的演化路径特征,构建流程和分析过程可为网络舆情治理提供工具及决策支持。

[Research purpose]Optimizing the event evolutionary graph construction method can enhance the knowledge discovery ability of the event evolutionary graph tool in unstructured online public opinion data,and it can better explore the causal relationship and evolutionary path within complex online public opinion events.[Research method]The study adopts RoBERTa pre-training model for sequence labeling to replace the traditional pattern matching method,introduces Word2Vec and BERTopic to replace the traditional machine learning clustering algorithm,and empirically analyzes the online public opinion of"Silicon Valley Bankruptcy"on Zhihu.[Research conclusion]The results show that causal extraction integrating deep learning and sequence labeling identifies 68613 original causality in 114901 contexts,which is 46.47%higher than the pattern matching method;event generalization based on word vector and topic clustering model classifies 2148 representative events into 14 topics,which outperforms the generalization effect at the level of textual and semantic features than the K-means algorithm.The research constructs an online public opinion event evolutionary graph based on the optimization method,which presents the characteristics of"cyclic""tight"and"long-chain"evolutionary paths around the core topics,and the construction process and analysis process can provide tools and decision support for online public opinion management.