高速公路突发事件实体识别及事件分类联合模型研究
2021-06-08范晓武葛嘉恒
范晓武 葛嘉恒



摘 要: 针对高速公路突发事件实体识别和事件分类任务中文本表征时存在的一词多义问题,提出使用层次多头注意力网络HMAN来学习文本字向量的高层次特征表示,结合经典的BiLSTM-CRF模型,构建一个称为HMAN-BiLSTM-CRF的多任务联合学习模型。模型共享文本特征表示模块,使用CRF对共享表征进行解码获得最优实体标注序列,而全连接层则根据输入的文本特征预测事件类别。在FEIC数据集上的实验结果显示,本文所提出的HMAN-BiLSTM-CRF在突发事件实体识别和分类两项任务中都优于其他对比模型。
关键词: 实体识别; 事件分类; 层次多头注意力网络; HMAN-BiLSTM-CRF模型
中图分类号:TP391.1 文献标识码:A 文章编号:1006-8228(2021)01-11-05
Research on the joint model of entity recognition and event
classification of freeway emergency
Fan Xiaowu, Ge Jiaheng
(Zhejiang Comprehensive Transportation Big Data Center Co., Ltd., Hangzhou, Zhejiang 310018, China)
Abstract: Aiming at the polysemy problem in text representation in freeway emergency entity recognition and event classification tasks, this paper proposes to use a hierarchical multi-head self-attention network to learn high-level feature representations of text word vectors, and combines with the classic BiLSTM-CRF Model to construct a multi-task joint learning model called HMAN-BiLSTM-CRF. The model shares the text feature representation module, and uses CRF to decode the shared representation to obtain the optimal entity annotation sequence. Meanwhile, the fully connected layer predicts the event category according to the input text feature. The experimental results on the FEIC data set show that the HMAN-BiLSTM-CRF proposed in this paper is superior to other comparison models in the two tasks of emergency entity recognition and classification.
Key words: entity recognition; event classification; hierarchical multi-head self-attention network; HMAN-BiLSTM-CRF model
0 引言
隨着我国高速公路建设规模的不断增长与道路交通量的快速增加,交通事故、恶劣天气、道路拥堵,以及危化品泄露等高速公路突发事件日益增长,严重影响高速公路的通行能力和运营效率。当高速公路突发事件发生后,交通应急指挥部门应根据报警信息快速定位事故点,调配应急救援物资并制定最佳救援路径,使高速公路能够迅速恢复平稳通行。在整个应急救援实施的过程中,精确确定事发点并分析出事件类别是应急救援能够正确、顺利开展的关键。然而,突发事件报警信息大多以语义来表述事发地理位置和事件情况,如何识别出突发事件位置等实体信息并对事件进行分类是亟待解决的问题,两者本质上是自然语言处理领域的经典任务:命名实体识别和文本分类。……
