BGCN:基于BERT和图卷积网络的触发词检测

BGCN:Trigger Detection Based on BERT and Graph Convolution Network

触发词检测是事件抽取的一项基本任务,该任务涉及对触发词进行识别和分类。目前,已有工作主要存在两方面的问题:1)用于触发词检测的神经网络模型只考虑了句子的顺序表示,且通过顺序建模的方法在捕捉长距离依赖关系时效率较低;2)基于表示的方法虽然解决了手动提取特征的问题,但用作初始训练特征的词向量对句子的表示程度有所欠缺,难以捕捉深层的双向表征。因此,文中提出了一种基于BERT模型和GCN网络的触发词检测模型BGCN,该模型通过引入BERT词向量来强化特征表示,并引入句法结构来捕捉长距离依赖,对事件触发词进行检测。实验结果表明,所提方法在ACE2005数据集上的表现优于其他现有的神经网络模型。

Trigger word detection is a basic task of event extraction,which involves the recognition and classification of trigger words.There are two main problems in the previous work:(1)the neural network model for trigger word detection only consi-ders the sequential representation of sentences,and the sequential modeling method is inefficient in capturing long-distance dependencies;(2)although the representation-based method overcomes the problem of manual feature extraction,the word vector used as the initial training feature lacks the degree of representation of the sentence,so it is difficult to capture the deep two-way representation.Therefore,we propose a trigger word detection model BGCN,based on BERT model and GCN network.This model strengthens the feature representation by introducing BERT word vector,and introduces syntactic structure to capture long-distance dependencies and detect event trigger words.Experimental results show that our method outperforms other existing neural network models on ACE2005 datasets.