软件导刊2026，Vol.25Issue(8)：49-59,11.DOI:10.11907/rjdk.261274
基于人工智能的社交网络虚假信息检测与内容治理研究
Research on Social Network False Information Detection and Content Governance Based on Artificial Intelligence
摘要
Abstract
In a social network environment,misinformation often spreads rapidly by exploiting the high dissemination and strong interactivity of platforms.Once it enters trending topic scenarios,it tends to amplify public misjudgment and increase platform governance costs.To ad-dress this practical issue,this paper investigates the connection between misinformation detection and content governance.In terms of experi-mental design,the LIAR dataset is selected to construct a binary classification task,with unified preprocessing applied to the training,valida-tion,and test sets.For model configuration,TF-IDF+Logistic Regression,TF-IDF+Support Vector Machine,DistilBERT,and Distil-BERT+LoRA(Parameter-Efficient Fine-Tuning)are evaluated for comparison.Regarding feature expansion,the built-in speaker historical records from the LIAR dataset(counts of barely-true,false,half-true,mostly-true,and pants-fire)are introduced to construct a speaker reputation score,which is fused with text features to form a metadata augmentation scheme.For result evaluation,the recognition performance of different methods is analyzed using Accuracy,Precision,Recall,and F1 scores,alongside confusion matrices,training loss curves,and validation set F1 curves.Experimental results show that DistilBERT achieves an Accuracy of 0.6314 and an F1 score of 0.6394 on the test set.After introducing the speaker reputation metadata,the Accuracy of TF-IDF+Meta+LR improves to 0.7466 and the F1 score to 0.6940,achiev-ing a significant enhancement over pure-text methods.Building upon this,quantitative thresholds(low-risk τ_L=0.40,high-risk τ_H=0.70)are established based on the model's prediction confidence P(fake)to conduct a three-level risk stratification experiment on 1,267 test samples.The research findings indicate that the metadata augmentation scheme incorporating speaker reputation features effectively improves detection performance,and the confidence-based tiered moderation mechanism provides a quantifiable empirical basis for platform content governance.关键词
社交网络/虚假信息检测/内容治理/LIAR数据集/DistilBERT/预训练语言模型Key words
social networks/false information detection/content governance/LIAR dataset/DistilBERT/pretrained language model分类
信息技术与安全科学引用本文复制引用
罗素梦,胡苓芝..基于人工智能的社交网络虚假信息检测与内容治理研究[J].软件导刊,2026,25(8):49-59,11.基金项目
广州商学院校级质量工程项目(2026ZLGC69) （2026ZLGC69）
广州商学院校级科研项目(2024XJYB018) （2024XJYB018）
