计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (11): 265-270.DOI: 10.3778/j.issn.1002-8331.1902-0257

• 工程与应用 • 上一篇    下一篇

航空安全事故因果关系抽取方法的研究

王红,祝寒,林海舟   

  1. 中国民航大学 计算机科学与技术学院,天津 300300
  • 出版日期:2020-06-01 发布日期:2020-06-01

Research on Causality Extraction of Civil Aviation Accident

WANG Hong, ZHU Han, LIN Haizhou   

  1. School of Computer Science & Technology, Civil Aviation University of China, Tianjin 300300, China
  • Online:2020-06-01 Published:2020-06-01

摘要:

针对目前航空安全事故因果关系分析一般采用基于概率和统计的方法,缺乏对事故发生过程的详细分析这一问题,提出通过因果关系抽取挖掘事故的因果发展过程。针对世界航空安全事故调查报告构成的文本数据集,将航空安全事故因果关系分为显式因果关系和隐式因果关系,其中显式因果关系抽取采用模式匹配的方法,抽取准确率达到87.72%;隐式因果关系抽取则采用改进的基于自注意力机制的双向长短期记忆网络方法,该方法在公共数据集和航空安全数据集上的[F]值较基准方法分别提高近6%和10%。在有效实现单一航空安全事故因果关系对的识别与抽取的基础上,生成了每个事故的因果关系图,为深入分析航空安全事故发生过程和情景重现提供数据与方法支持。

关键词: 航空安全事故, 关系抽取, 显式因果关系, 隐式因果关系, 模式匹配, 自注意力机制

Abstract:

Previous research on causality analysis of civil aviation accidents is mainly based on probability and statistic methods, lacking of the analysis based on the whole process of accident occurrence. In order to excavate the process of aviation accidents, the causality extraction method is studied. The causality is divided into explicit causality and implicit causality, and the pattern matching method is applied to get explicit causality, which reaches 87.72% in precise rate. While the improved bidirectional long short-term memory networks with self-attention mechanism is adopted for implicit causality extraction. Compared to the baseline method, the F measure of this method on public dataset and aviation dataset is increased by nearly 6% and 10%, respectively. The proposed approach is applied to the text dataset of world aviation safety accident investigation reports. On the basis of the effectively extraction of causality, causality graph for each accident is generated, which provides the data and method to civil aviation accident’s comprehensive analysis and scene reappearance.

Key words: civil aviation accident, relation extraction, explicit causality, implicit causality, pattern matching, self-attention mechanism