新闻文本挖掘下“暴雨-山洪-地质灾害”的情景演化分析  

Scenario evolution analysis of“rainstorm-flash flood-geological disasters”based on news text mining

作　　者：郭昕曜 星宇铮 王远声 吕伟[5] GUO Xinyao;XING Yuzheng;WANG Yuansheng;L Wei(School of Civil Aviation,Zhengzhou University of Aeronautics,Zhengzhou 450046,China;Safety Science and Engineering Postdoctoral Research Station,Henan Polytechnic University,Jiaozuo 454003,Henan,China;Henan International Joint Laboratory of Man Machine Environment and Emergency Management,Anyang Institute of Technology,Anyang 455099,Henan,China;School of Environmental and Municipal Engineering,North China University of Water Resources and Electric Power,Zhengzhou 450046,China;School of Safety Science and Emergency Management,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]郑州航空工业管理学院民航学院,郑州450046 [2]河南理工大学安全科学与工程博士后流动站,河南焦作454003 [3]安阳工学院河南省人机环境与应急管理国际联合实验室,河南安阳455099 [4]华北水利水电大学环境与市政工程学院,郑州450046 [5]武汉理工大学安全科学与应急管理学院,武汉430070

出　　处：《安全与环境学报》2025年第4期1466-1476,共11页Journal of Safety and Environment

基　　金：国家自然科学基金项目(72304253,52072286);教育部人文社会科学研究项目(22YJC630027);河南省高等学校青年骨干教师培养计划项目(2024GGJS098);河南省人机环境与应急管理国际联合实验室开放项目(KFKT-003);河南省重点研发专项(221111321000)。

摘　　要：为分析“暴雨-山洪-地质灾害”的灾变特点,利用自动化抓取技术提取了2010—2022年国内权威新闻媒体对长江中上游区域的暴雨、山洪、滑坡、泥石流等灾害的报道数据。基于自然语言处理(Natural Language Processing, NLP)技术和机器学习方法,对新闻文本进行了预处理与数据清洗,实现了灾害信息的自动分类。进而,采用贝叶斯网络模型构建了灾害链的拓扑结构,推演了灾害演化过程中的各节点概率,揭示了“暴雨-山洪-地质灾害”链的情景演化规律。最后,以四川省凉山州冕宁县2020年的灾害事件为例,预测了“暴雨-山洪-地质灾害”网络中各情景节点概率,验证了贝叶斯网络模型的可靠性。结果表明,构建的“暴雨-山洪-地质灾害”的贝叶斯网络模型在山洪、泥石流、滑坡、人员伤亡、房屋倒塌等目标变量预测中,预测结果与实际数据基本一致,各目标变量的Brier检验平均结果为0.115。研究结论为“暴雨-山洪-地质灾害”的预测和情景演化分析提供了方法支撑。To analyze the characteristics of the“rainstorm-flash flood-geological disaster”chain,automated extraction techniques were employed to collect data from reports on rainstorms,flash floods,landslides,and debris flows in the middle and upper reaches of the Yangtze River region by authoritative Chinese news media from 2010 to 2022.Combined with Natural Language Processing(NLP)techniques and machine learning methods,the news texts were preprocessed and cleaned to establish a trigger word set for disaster causes,a keyword set for disaster impacts,and a disaster category dictionary.A disaster type identification and automatic classification method was designed,and the accuracy of the classification method was validated using multinomial logistic regression,achieving the transformation of news texts from unstructured to structured data.Subsequently,a Bayesian network model was utilized to construct the topological structure of the“rainstorm-flash flood-geological disaster”chain.Using the Expectation Maximization(EM)algorithm,the conditional probability distributions between nodes were estimated,and the probabilities of various nodes in the disaster evolution process were deduced,revealing the chain-like scenario evolution patterns of the“rainstorm-flash flood-geological disaster”chain.Using the 2020“rainstorm-flash flood-geological disaster”event in Mianning County,Liangshan Prefecture,Sichuan Province,as a case study,the probabilities of scenario nodes in the disaster network were predicted,and the reliability of the model was validated using Brier score testing.Results show that the multinomial logistic regression method achieves an average precision of 84%,a recall of 82%,and an F 1-score of 83%for classifying four disaster types:rainstorm,flash flood,debris flow and landslide.For the Bayesian network model,the prediction for target variables such as flash floods,debris flows,landslides,casualties,and house collapses is consistent with actual data.The Brier scores for the flash flood,debris flow,lands

关 键 词：公共安全 暴雨-山洪-地质灾害 情景演化 文本挖掘 贝叶斯网络 

分 类 号：X915.5[环境科学与工程—安全科学]