Scholarlink

Our study addresses the challenge of processing vast amounts of unstructured data by extracting and organizing key information via knowledge graphs. Focusing on disinformation detection, we analysed over 100,000 messages from nine Russian Telegram channels. We constructed knowledge graphs applying two triple extraction methods: Subject-Verb-Object (SVO) and Entity-Relation-Entity (ERE). We identified and studied disinformation cases by comparing Telegram data graphs with graphs of verified cases from the EUvsDisinfo database using Graph Kernels. Results revealed disinformation across all analysed channels, particularly regarding the downing of Flight MH17, claims about Western assistance, and narratives on Ukraine’s occupied regions. Our results demonstrated that the Shortest Path Graph Kernel and Subgraph Matching methods were the most effective for detecting disinformation as they accurately identified graphs containing substantial disinformation, highlighting the potential of knowledge graphs in large-scale media monitoring.

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

A Graph Network Approach to Disinformation Detection in Social Media

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Milita Songailaitė,
  • Justina Mandravickaitė,
  • Veronika Bryskina,
  • Maksym Bondar,
  • Tomas Krilavičius

摘要

Our study addresses the challenge of processing vast amounts of unstructured data by extracting and organizing key information via knowledge graphs. Focusing on disinformation detection, we analysed over 100,000 messages from nine Russian Telegram channels. We constructed knowledge graphs applying two triple extraction methods: Subject-Verb-Object (SVO) and Entity-Relation-Entity (ERE). We identified and studied disinformation cases by comparing Telegram data graphs with graphs of verified cases from the EUvsDisinfo database using Graph Kernels. Results revealed disinformation across all analysed channels, particularly regarding the downing of Flight MH17, claims about Western assistance, and narratives on Ukraine’s occupied regions. Our results demonstrated that the Shortest Path Graph Kernel and Subgraph Matching methods were the most effective for detecting disinformation as they accurately identified graphs containing substantial disinformation, highlighting the potential of knowledge graphs in large-scale media monitoring.