基于篇章级信息特征增强的中文突发事件检测

Chinese emergency event detection based on chapter-level information feature enhancement

事件检测旨在检测句子中的触发词并将其分类为特定的事件类型。针对目前大多数的中文事件检测方法存在篇章内句子间相互依赖信息利用不充分的问题,提出了基于篇章级信息特征增强的中文突发事件检测方法。首先,通过BERT(Bidirectional Encoder Representations from Transformers)预训练语言模型编码字向量,使用平均池化将字向量转换成句子表示信息;其次,利用双向门控循环神经网络(Bidirectional gated recurrent neural network,Bi-GRU)学习句子内和篇章内的上下文隐层信息,接着通过注意力机制(Attention Mechanism)分别得到句子级信息和由若干句子间的相互依赖信息生成的篇章级信息;最后,将篇章级信息融合到句子级信息上后与字向量拼接,再使用条件随机场完成对句子中触发词的识别和标注。实验结果表明,该方法有效的提升了中文突发事件检测效果,F1值达到79.95%。

Event detection aims at detecting trigger words in sentences and classifying them into specific event types.For most current Chinese event detection methods suffering from insufficient utilization of interdependent information among sen-tences within a chapter,a Chinese burst event detection method based on chapter-level information feature enhancement is pro-posed.Firstly,the word vector is encoded by a BERT pre-trained language model,and then the word vector is converted into sen-tence representation information using averaging pooling;Secondly,a bidirectional gated recurrent neural network is used to learn the contextual implicit information within the sentence and within the chapter;Thirdly,the sentence-level information and chapter-level information dependent on several sentences is achieved by the attention mechanism;Finally by the chapter-level in-formation is fused to the sentence-level information and spliced with the word vector,and then the conditional random field is used to identify and label the trigger words in the sentence.The experimental results show that the method effectively improves the detection of Chinese bursts with an F1 value of 79.95%.