基于BERT-DGCNN的中文事件抽取方法研究

Research on Chinese Event Extraction Method Based on BERT-DGCNN

本文构建了一个事件抽取pipeline模型,其旨在对新闻中的信息元进行有效的抽取。在管道抽取模式下,先对文本进行存在事件类型识别,而后再将事件类型与文本一并作为输入传入模型进行事件论元角色抽取,其中事件论元角色采用类似于BERT中SQuAD等阅读理解任务上的双指针输出。两个基本模型都是利用BERT预训练模型产生的词嵌入,使用DGCNN进行编码之后池化,再连接到dense层进行分类。实验结果表明,本模型可对新闻类内容进行高效抽取。

This paper constructs an event extraction model, which aims at effectively extracting information elements from the news. This model is based on the BERT pretraining model, which first identifies the existing event type of the text, and then inputs the event type and text into the model to extract the event role, in which the event role adopts double-pointer output similar to the reading comprehension task of a SQuAD in BERT. Experimental results show that this model can extract news content efficiently.