宁慧涵, 眭海刚, 王金地, 胡烈云, 刘金硕, 刘俊怡. 顾及时空关系的事故灾难事理图谱构建方法研究[J]. 武汉大学学报 ( 信息科学版), 2024, 49(5): 831-843. DOI: 10.13203/j.whugis20230291
引用本文: 宁慧涵, 眭海刚, 王金地, 胡烈云, 刘金硕, 刘俊怡. 顾及时空关系的事故灾难事理图谱构建方法研究[J]. 武汉大学学报 ( 信息科学版), 2024, 49(5): 831-843. DOI: 10.13203/j.whugis20230291
NING Huihan, SUI Haigang, WANG Jindi, HU Lieyun, LIU Jinshuo, LIU Junyi. Construction of Disaster Event Evolutionary Graph Based on Spatiotemporal Relationship[J]. Geomatics and Information Science of Wuhan University, 2024, 49(5): 831-843. DOI: 10.13203/j.whugis20230291
Citation: NING Huihan, SUI Haigang, WANG Jindi, HU Lieyun, LIU Jinshuo, LIU Junyi. Construction of Disaster Event Evolutionary Graph Based on Spatiotemporal Relationship[J]. Geomatics and Information Science of Wuhan University, 2024, 49(5): 831-843. DOI: 10.13203/j.whugis20230291

顾及时空关系的事故灾难事理图谱构建方法研究

Construction of Disaster Event Evolutionary Graph Based on Spatiotemporal Relationship

  • 摘要: 事故灾难事理图谱可以全面表达事故发展过程、各子事件信息及多种事件关系,为事故灾难分析提供知识服务。针对事故灾难事理图谱构建中存在的时空关系中文语料匮乏、中文词汇边界模糊导致事件抽取不准确、隐式事件关系难以识别的问题,提出一种顾及时空关系的事故灾难事理图谱构建方法。该方法首先设计了基于命名规律性的词汇增强事件抽取模型,感知实体名称规律以确定事件信息边界和类型,然后采用一种融合注意力和门控空洞卷积的关系识别方法,获取多维度文本特征来挖掘潜在事件关系,并建立了含时空关系的中文事故灾难语料库(Chinese disaster corpus with spatiotemporal relationship,CDCSTR)。在CDCSTR和中文突发事件语料库上进行实验,结果表明,事件抽取模型的F1值分别达到88.59%和78.49%;与现有方法相比,关系识别模型在CDCSTR上的4个任务性能均有提升,尤其是空间关系识别优势明显,取得了3.08%以上的性能领先。以成乐高速追尾事故为例进行验证,构建的事理图谱能展示现实场景下的事故演化过程和时空变化特征,辅助事故应急工作。

     

    Abstract:
    Objectives Disaster event evolutionary graph can comprehensively express the development process of disaster, information of sub-events and multiple event relationships, providing knowledge services for disaster analysis. The problems in the construction of disaster event evolutionary graph include the lack of Chinese corpus containing spatiotemporal relationship, imprecise event extraction due to fuzzy boundaries of Chinese vocabulary, and difficulty in recognizing implied event relationships.
    Methods This paper proposes a method for constructing disaster event evolutionary graph based on spatiotemporal relationship. First, a vocabulary enhancement event extraction model based on naming regularity (VENR) is designed to discern entity name patterns, and determine event information boundaries and types. Second, a relation recognition model by fusing attention and gated-dilated convolution (AGDC) is proposed to obtain multi-dimensional text features, thereby uncovering potential event relationships. Additionally, a Chinese disaster corpus with spatiotemporal relationship (CDCSTR) is established.
    Results We experiment on CDCSTR and Chinese emergency corpus (CEC), the results show that the F1 scores of VENR are 88.59% and 78.49%,respectively on CDCSTR and CEC. Compared with the existing methods, AGDC improves the performance of four tasks on CDCSTR, especially excelling in spatial relationship recognition, achieving a performance lead of over 3.08%.
    Conclusions The Chengdu-Leshan expressway tailgating accident event evolutionary graph is built for verification, enabling the effective expression of event evolution and spatiotemporal features in real disaster scene, consequently supporting emergency responses to disaster.

     

/

返回文章
返回