摘要
目前火灾事故报告都是以文本形式存在,有关火灾事故的经验知识则蕴含在非结构化的文本中,应急主体要获得经验知识,保证应急决策中的方案更精确,依靠人工对文本进行抽取,将耗费较高的成本.本文提出了基于分句模型的燃烧物关键词的抽取方法,抽取出燃烧物描述的句子,并采用双向长短期记忆(BILSTM,BI-DIRECTIONAL LONG SHORT TERM MEMORY)神经网络,结合注意力(ATTENTION)机制构建句子级的燃烧物抽取模型,对句子中描述的燃烧物进行处理,最后得到燃烧物的向量.通过实验研究了分句与全文两种状态下,转换成词向量后分别输入LSTM、LSTM+Attention、BiLSTM+Attention和支持向量机(support vector machines,SVM)等四个模型进行燃烧物抽取,对抽取结果进行预测准确率、F1值宏平均、Precision值宏平均和Recall值宏平均四个指标对比,证实了分句模型的优越性和基于BILSTM结合ATTENTION机制框架模型的正确性.
At present,fire accident reports are in the form of text,and the experience knowledge about fire accidents is contained in unstructured text.The emergency subject should obtain experience knowledge to ensure the scheme in emergency decision-making more accurate.It will cost more to extract the text manually.In view of this,this paper puts forward a method of extracting keywords of burners based on Clause model,extracts sentences describing burners,constructs sentence-level burner extraction model using BiLSTM(Bi-directional long short-term memory)neural network,combines Attention mechanism,processes the burners described in the sentence,and finally obtains the vectors of burners.Four models,LSTM,LSTM+Attention,BiLSTM+Attention and SVM,were input to extract combustion substances in two states:clause and text.The prediction accuracy,F1 value macro-average,Precision value macro-average and Recall value macro-average were compared.The superiority of the sentence model and the correctness of the BiLSTM-based Attention mechanism framework model are verified.
作者
季峰
JI Feng(School of Xinglin,Nantong University,Nantong 226000,China;School of Engineering,Mokwon University of Korea,Daejeon 302828,Korea)
出处
《青海师范大学学报（自然科学版）》
2020年第3期14-21,共8页
Journal of Qinghai Normal University(Natural Science Edition)
基金
国家自然科学基金项目(51305212)
南通市市级科技计划项目(JCZ18021)
南通大学杏林学院自然科学项目(2018K124)
南通大学杏林学院创新项目(201913993035xl)。
关键词
火灾文本
燃烧物
信息抽取
注意力机制
神经网络
fire text
combustion
information extraction
attention mechanism
neural network
