基于深度学习的交通事故文本因果关系抽取

Causality extraction from traffic accidents texts based on deep learning

针对交通事故文本因果关系抽取过程中因果事件边界难以识别及连锁因果关系难以抽取的问题,将抽取问题转化为序列标注问题,提出了相对逗号位置特征及基于该特征与字词向量混合的多头注意力卷积双向长短时记忆网络的因果关系抽取方法。首先将字词分别编码后与相对逗号位置特征拼接,其次通过卷积神经网络(convolutional neural network,CNN)、双向长短时记忆网络(bidirectional long and short-term memory networks,Bi-LSTM)及多头注意力机制(multihead self-attention,MHSA)提取深层次的语义信息及长距离特征信息,最后采用条件随机场(conditional random field,CRF)分类器进行分类,得到最终的输出结果。在我们创建的交通事故文本数据集上将本模型与主流模型进行比较,结果表明:本模型抽取结果的召回率与F_(1)值分别提高了5.75%和2.54%,可以更有效地抽取交通事故文本中的因果关系。较完整地抽取因果关系有利于人们分析交通事故的成因,从而为如何有效地预防和避免交通事故的再次发生提供参考。

In response to the problem that the event boundary and chain causalities are difficult to identify in the process of causality extraction from traffic accident texts,the causality extraction was transformed into a sequence labeling task,in which a method was proposed to extract the relative comma position feature.This method encoded the words and chars,then combined them with the relative comma position features,introduced convolutional neural networks(CNN),bidirectional long and short-term memory networks(Bi-LSTM) and multihead self-attention(MHSA) to extract deep features and long-distance features,and finally used a conditional random field(CRF) classifier for classification to obtain the ultimate outputs.By comparing the model with mainstream models on the traffic accident text dataset,the results show that the model has a marked increase in the recall rate and F_(1) value,by 5.57% and 2.54% respectively,capable of more effectively extracting the causal relationship in traffic accident texts.And it is more conducive to the analysis into the causes of traffic accidents,so it can provide reference for effectively preventing and avoiding traffic accidents.