摘要： 针对现有事件因果关系抽取方法关系边界识别能力弱和文本语义表征不足的问题，提出一种基于双层CNN-BiGRU-CRF深度学习模型的事件因果关系抽取方法。将因果关系抽取任务转换为两次序列标注任务分别由两层CNN-BiGRU-CRF模型完成，上层模型用于识别事件因果关系语义角色词，其标注结果作为特征输入下层模型划分因果关系边界。在每层模型中，采用突发事件样本数据对BERT模型进行微调，形成文本表示模型以获取语义特征向量矩阵，利用卷积神经网络和双向门控循环单元分别提取局部和全局深层特征，并将上述特征在每个时间序列步进行线性加权融合以增强语义表征能力，最终基于残差思想将高区分度特征输入CRF模型解码完成序列标注任务。在中文突发事件语料集上的实验结果表明，与BiLSTM-Att-规则特征、GAN-BiGRU-CRF等因果关系抽取方法相比，该方法的事件因果关系抽取效果更好，F值达到91.81%，能有效实现事件因果关系的准确抽取。

关键词: 因果关系抽取, 深度学习, 卷积神经网络, 特征融合, 突发事件

Abstract: The existing event causality extraction methods have poor performance in relationship boundary recognition and are limited by insufficient semantic representation of texts. To solve the problem, this paper proposes a method of event causality extraction based on two-layer CNN-BiGRU-CRF deep learning model. The task of causality extraction is transformed into two sequential annotation tasks, which are respectively completed by two-layer CNN-BiGRU-CRF model. The upper model is used to identify the semantic role words of event causality, and the annotation results are input as features into the lower model to divide the causality boundary. In both models, the emergency sample data is used to fine-tune the BERT model to form a text representation model to obtain the semantic feature vector matrix. Convolutional Neural Network(CNN) and Bi-directional Gated Recurrent Unit(BiGRU) are used to extract local and global deep-layer features respectively, and the above features are linearly weighted and fused in each time series step to enhance the semantic representation. Finally, based on the idea of residuals, high resolution features are input into the CRF model for decoding and to complete the task of sequence annotation. Experimental results on Chinese emergency corpus show that compared with the causality extraction methods such as BiLSTM-Att-rule feature and GAN-BiGRU-CRF, the proposed method has better causality extraction performance with F value reaching 91.81%. It can effectively achieve the accurate extraction of event causality.

Key words: causality extraction, deep learning, Convolutional Neural Network(CNN), feature integration, emergency event

中图分类号: 

• TP391