北京大学学报（自然科学版） ›› 2026, Vol. 62 ›› Issue (5): 963-976.DOI: 10.13209/j.0479-8023.2026.008

上一篇    下一篇

基于图补全的多语言社交媒体事件检测

程效天1,2, 马龙轩1,2, 张凯1,2, 线岩团1,2, 余正涛1,2, 黄于欣1,2,†   

  1. 1. 昆明理工大学信息工程与自动化学院, 昆明 650500 2. 云南省人工智能重点实验室, 昆明理工大学, 昆明 650500
  • 收稿日期:2025-07-31 修回日期:2025-09-11 出版日期:2026-09-20 发布日期:2026-09-20
  • 基金资助:
    国家自然科学基金(U21B2027, 62266028, U23A2038)、云南省重大科技专项(202302AD080003, 202402AG050007,  202303AP140008)和云南省基础研究(202301AS070047, 202301AT070471)资助

Multi-language Social Media Event Detection Based on Graph Completion

CHENG Xiaotian1,2, MA Longxuan1,2, ZHANG Kai1,2, XIAN Yantuan1,2, YU Zhengtao1,2, HUANG Yuxin1,2,†   

  1. 1. Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500 2. Yunnan Key Laboratory of Artificial Intelligence, Kunming University of Science and Technology, Kunming 650500
  • Received:2025-07-31 Revised:2025-09-11 Online:2026-09-20 Published:2026-09-20

PDF

摘要：

针对多语言场景下, 语言差异导致嵌入相似度降低以及社交消息图中隐含的关联关系常被忽略这两大问题, 提出一种基于图补全的多语言社交媒体事件检测方法MGCGNN。首先基于预训练语言模型生成消息嵌入, 借助维基百科实体ID实现跨语言实体对齐, 构建初始多语言社交消息图; 然后采用随机游走策略, 捕捉节点间隐含的长距离语义关联, 补全消息图结构, 从而增强事件聚类效果。在基于Event2012数据集构建的三语混合数据集上的实验结果表明, 所提方法的性能显著超越基线模型, NMI值平均提升4%~5%, 证明该方法在复杂多语言场景下具备更高的精确性与有效性。

关键词: 多语言, 社交媒体事件检测, 属性对齐, 随机游走, 图补全

Abstract:

To address the two major issues of reduced embedding similarity due to language differences in multilingual scenarios and the frequent neglect of implicit relational connections in social message graphs, a multilingual social media event detection method called MGCGNN is proposed based on graph completion. First, message embeddings are generated using pre-trained language models, and cross-lingual entity alignment is achieved via Wikipedia entity IDs to construct an initial multilingual social message graph. Subsequently, a random walk strategy is employed to capture long-range semantic associations between nodes, completing the message graph structure and thereby enhancing event clustering effectiveness. Experiments on a trilingual hybrid dataset built upon the Event2012 dataset demonstrate that the proposed method significantly outperforms baseline models, with an average NMI improvement of 4%–5%, validating higher precision and effectiveness in complex multilingual scenarios.

Key words: multilingual, social media event detection, attribute alignment, random walk, graph completion

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司