【作者】 刘晓；

【导师】 郭海湘；

【作者基本信息】 中国地质大学 ， 管理科学与工程， 2021， 博士

【摘要】 气候变化加剧、城市化进程加快,以及人类社会活动的影响,灾害的联动发生日益表现为一种常态。而城市化进程的不断加快,使得灾害系统变得更加复杂。传统的灾害管理和防灾减灾模式面临严峻挑战。因此,基于我国各类灾害的时空分布规律,分析灾害的多级联动模式,预测灾害多级联动发生的概率,已成为防灾减灾中提供应急决策支持的有效方法之一。而依赖国家官方数据获取灾害信息的方式存在数据收集困难、数据滞后等问题。随着互联网和信息技术的迅速发展,网络信息传播更加便捷和高效。网络文本逐渐成为大数据背景下一种重要信息资源,为灾害领域相关研究提供了前所未有的机遇。网络新闻、百度百科等网络文本数据中隐藏着很多有价值的灾害信息,如何基于文本挖掘的方法识别和提取出这些潜在的有价值的灾害信息,并基于此构建灾害多级联动分析模型是需要解决的关键科学问题。本文以国家自然科学基金项目“灾害多级联动模式下城市群综合承灾能力的评价与仿真研究”为依托,以网络文本数据为研究视角,围绕灾害关联关系分析与挖掘这一核心主题,综合运用文本挖掘、机器学习、自然语言处理、贝叶斯网络、复杂网络等方法和技术,在分析我国自然灾害时空特征的基础上,以暴雨灾害为灾害多级联动的研究对象,进一步探索其领域特征和文本识别方法,抽取事件因果关系,构建贝叶斯网络模型,实现对其次生灾害节点的推理与预测,为防灾减灾和断链减灾提供科学的决策支持。主要的研究工作和结论如下:(1)基于网络新闻数据的我国自然灾害时空特征分析。采用信息抽取方法对中国新闻网2008-2017年的灾害报道进行了信息抽取和分析,得到我国主要自然灾害类型,并通过与官方信息的对比验证了网络新闻数据用于灾害研究的合理性和有效性,结果表明我国气象灾害主要分布在每年的4月到9月,高峰期在7月和8月;灾害主要发生在云南、四川、贵州、湖南等地,而这些地区的主要灾害类型为暴雨、洪涝和地质灾害;暴雨和洪涝灾害很有可能存在空间关联性。(2)灾害多级联动新闻文本识别研究。网络新闻涉及到方方面面的信息,需要进一步识别和筛选灾害多级联动相关的文本数据。针对领域敏感性,采用基于标签传播的领域词典构建方法,融合领域主题特征和Word2vec词向量,结合基于集成思想的XGBoost方法,研究灾害多级联动新闻文本的识别问题,实验结果表明本研究提出的方法在准确率和召回率方面相比单独使用Word2vec有了一定的提高。(3)灾害多级联动事理图谱构建与分析。基于单一案例手工构建的灾害链模型不仅费时费力,还容易产生冗余和遗漏。研究从灾害新闻文本中抽取灾害事件因果关系,获取灾害因果知识和经验,并将灾害事件的多级联动模式刻画成一个有向的事理知识库,进而引入复杂网络方法,对灾害多级联动事理知识图谱中的关键节点进行分析,为防灾减灾决策提供科学的参考依据。(4)基于贝叶斯网络的灾害多级联动分析与预测。基于灾害多级联动事理知识图谱,以相关文献中的灾害节点影响因素作为补充,构建一个比较全面的、客观的灾害多级联动贝叶斯网络模型。以暴雨灾害为例,通过构建暴雨-地质和暴雨-洪涝的贝叶斯网络模型,并基于历史案例数据进行参数学习,预测多级联动事件发生的概率和后果的严重程度,据此提出暴雨灾害的断链减灾措施。本文的主要创新点体现在以下三个方面:(1)针对依赖专家知识构建贝叶斯网络的局限性,提出了一种基于事理图谱的贝叶斯网络建模方法,推理预测灾害多级联动。(2)针对特征词领域敏感问题,提出了一种融合领域主题特征和全局文本特征的灾害多级联动文本识别方法。(3)针对当前因果关系抽取方法未考虑文本中的灾害多级联动问题,提出了一种基于因果提示词扩展词典构建的多层灾害因果链抽取方法。更多还原

【Abstract】 With the intensification of global climate change and the impact of human social activities,the linkage of disasters is increasingly becoming a norm.The accelerating urbanization process makes the disaster system more complex.The traditional disaster management and the model of disaster prevention and mitigation have facing severe challenges.Therefore,based on the spatial and temporal distribution of various disasters in China,analyzing the multi-level linkage mode of disasters and predicting the probability of multi-level linkage of disasters have becoming one of the effective methods to provide emergency decision support in disaster prevention and mitigation.There are some problems in the way of obtaining disaster information based on national official data,such as data collection difficulties and data lag.With the rapid development of Internet and information technology,network information dissemination is more convenient and efficient.Network text has become an important information resource under the background of big data,which provides unprecedented opportunities for related research in the field of disaster.There are many valuable disaster information hidden in network text data such as network news and Baidu encyclopedia.How to identify and extract these potential valuable disaster information based on text mining method,and build a multi-level disaster linkage analysis model based on this is a key scientific problem to be solved.Based on the National Natural Science Foundation of China(NSFC)project “Evaluation and Simulation of Comprehensive Disaster-bearing Capacity of Urban Agglomeration under Multi-level Linkage Mode of Disasters”,from the perspective of network text data,this thesis focuses on the core theme of disaster association analysis and mining,and comprehensively uses methods and technologies such as text mining,machine learning,natural language processing,complex network and Bayesian network method.On the basis of analyzing the spatial and temporal characteristics of natural disasters in China,the rainstorm disaster is determined as the research object of multi-level linkage of disasters,and further explores its field characteristics and text recognition method,extracts event causal relationship,constructs Bayesian network model,realizes the reasoning and prediction of secondary disaster nodes,and provides decision support for disaster prevention and mitigation and chain breaking disaster reduction.The main research work and conclusions are as follows:Firstly,spatial and temporal characteristics of natural disasters in China based on network news data.The method of information extraction is used to extract and analyze the disaster reports of China News Network from 2008 to 2017,and the main types of natural disasters in China are obtained.The rationality and effectiveness of network news data for disaster research are verified by comparing with official information.The results show that the meteorological disasters in China are mainly distributed from April to September,and the peak periods are July and August.Disasters mainly occur in Yunnan,Sichuan,Guizhou and Hunan,and the main types of disasters in these regions are rainstorm,flood and geological disasters.Rainstorm and flood disasters are likely to have spatial correlation.Secondly,cascading disaster news text recognition research.Network news involves all aspects of information,need to further identify and filter cascading disaster related text data.In view of the domain sensitivity,this thesis adopts the domain dictionary construction method based on label propagation,integrates the domain topic feature and Word2 vec word vector,and combines the XGBoost method based on integration idea to study the identification of cascading disaster news text.Experimental results show that the proposed approach works well for the accuracy and recall of text classification compared to separate using Word2 vec.Thirdly,the construction and analysis of cascading disaster map.The disaster chain model constructed manually based on a single case is not only time-consuming and laborious,but also prone to redundancy and omission.The causal relationship of disaster events are extracted from disaster news texts.And the knowledge and experience of disaster causal relationship are obtained.The multi-level linkage mode of disaster events is characterized as a directed knowledge base,and then the complex network method is introduced to analyze the key nodes in the multi-level linkage knowledge map of disaster events,which provides scientific reference for disaster prevention and mitigation decisions.Finally,The analysis and prediction of multi-level linkage disaster based on Bayesian network.Based on the knowledge map of cascading disaster,a comprehensive and objective cascading disaster Bayesian network model is constructed with the influence factors of disaster nodes in relevant literature as supplement.Taking the rainstorm disaster as an example,the Bayesian network model of rainstorm-geology and rainstorm-flood is constructed,and the parameter learning is carried out based on historical case data to predict the probability of multi-level linkage events and the severity of the consequences.Based on this,the chain breaking mitigation measures of rainstorm disasters are proposed.The main innovations of this thesis are reflected in the following three aspects:(1)Aiming at the limitation of relying on expert knowledge,a Bayesian network modeling method based on event evolution graph is proposed to infer and predict cascading disaster.(2)Aiming at the problem of domain sensitivity,a cascading disaster text recognition method based on domain theme features is proposed.(3)In view of the fact that the current causality extraction method fails to consider cascading disasters in texts,a multilevel causality chain extraction method based on the extended causal word dictionary is proposed was proposed.更多还原