Scholarlink

<p>Current rumor detection studies rarely model the semantic relevance between source posts and their subsequent reposts, which can lead to noisy inputs, higher computational costs, and delayed early detection. To address this gap, we propose BRAKE, a relevance-aware and malice-sensitive early rumor detection framework that captures entity-level semantic alignment and quarrel-oriented malicious features. Evaluated on our multilingual controversial rumor dataset and two widely used benchmarks (Twitter15, Weibo), BRAKE achieves up to 23.13% higher accuracy than competitive baselines such as CICAN, DDGCN, ClaHi-GAT, BiGCN, RvNN, dEFEND, EHGCN, and PPC, demonstrating substantial gains in both early- and late-stage detection. This paper provides strong evidence that reposts following source posts usually become fierce quarrels, increasing the hostility level in discussion environments and swaying public opinions, thus clearly signaling an upcoming misinformation tide. A novel framework called BRAKE is proposed to efficiently detect sequentially propagating rumors via bidirectional encoder representations from transformers-enhanced concatenation in a knowledge forest structure (FK-BERT) and malice-aware adaptive kernel convolution from the content correlation perspective. The experimental results of a real-time application demonstrate that BRAKE outperforms other competitive state-of-the-art methods in rumor detection; Its accuracy is maximally 23.13% higher than those of the other methods. A correlation analysis also reveals that rumor content features are correlated more with malicious repost features than with the content features of the original news article itself. Rumor-related samples of all categories that are likely or unlikely to lead to quarrels are also presented with detailed descriptions. This framework is applicable across multiple social media platforms, supports both English and Chinese language content, and can be adapted to different rumor detection stages, from early propagation to full-scale diffusion.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quarrels occurred matters: relevant-aware rumor detection with FK-BERT-enhanced Maak convolution

  • 期刊论文

  • 发表日期：

    2025年11月5日

查看全文
  • Hongchen Wu,
  • Xiaochang Fang,
  • Hongzhu Yu,
  • Bing Yu,
  • Meng Li,
  • Hongxuan Li,
  • Zhaorong Jing,
  • Chenfei Sun,
  • Shuang Gao,
  • Huaxiang Zhang

摘要

Current rumor detection studies rarely model the semantic relevance between source posts and their subsequent reposts, which can lead to noisy inputs, higher computational costs, and delayed early detection. To address this gap, we propose BRAKE, a relevance-aware and malice-sensitive early rumor detection framework that captures entity-level semantic alignment and quarrel-oriented malicious features. Evaluated on our multilingual controversial rumor dataset and two widely used benchmarks (Twitter15, Weibo), BRAKE achieves up to 23.13% higher accuracy than competitive baselines such as CICAN, DDGCN, ClaHi-GAT, BiGCN, RvNN, dEFEND, EHGCN, and PPC, demonstrating substantial gains in both early- and late-stage detection. This paper provides strong evidence that reposts following source posts usually become fierce quarrels, increasing the hostility level in discussion environments and swaying public opinions, thus clearly signaling an upcoming misinformation tide. A novel framework called BRAKE is proposed to efficiently detect sequentially propagating rumors via bidirectional encoder representations from transformers-enhanced concatenation in a knowledge forest structure (FK-BERT) and malice-aware adaptive kernel convolution from the content correlation perspective. The experimental results of a real-time application demonstrate that BRAKE outperforms other competitive state-of-the-art methods in rumor detection; Its accuracy is maximally 23.13% higher than those of the other methods. A correlation analysis also reveals that rumor content features are correlated more with malicious repost features than with the content features of the original news article itself. Rumor-related samples of all categories that are likely or unlikely to lead to quarrels are also presented with detailed descriptions. This framework is applicable across multiple social media platforms, supports both English and Chinese language content, and can be adapted to different rumor detection stages, from early propagation to full-scale diffusion.