基于图注意力网络的突发事件触发词抽取

Emergent triggers extraction based on graph attention network

针对目前突发事件触发词抽取方法存在由于分词引起的误差传递而导致触发词提取不准确的问题,提出基于图注意力网络的突发事件触发词抽取模型(ETEGAN)。ETEGAN首先使用Word2vec和BERT预训练语言模型对文本序列进行向量化,将获取到的向量表示与动态词向量相结合,使用双向门控循环单元BiGRU提取上下文特征,并利用图注意力网络GAT提取文本特征,调整重要特征的权重,突出重要词对事件触发词抽取的贡献。实验结果表明,本文模型有效地提高了突发事件触发词识别准确率。

To address the problem of inaccurate extraction of trigger words due to error transmission caused by splitting of words in current emergency trigger word extraction methods,we propose emergency triggers extraction based on graph attention network(ETEGAN).ETEGAN first uses Word2vec and BERT pre-trained language models to vectorize text sequences,combines the obtained vector representations with dynamic word vectors,extracts contextual features using BiGRU,a two-way gated cyclic unit,and extracts text features using GAT,adjusts the weights of important features,and highlights the contribution of important words to event-triggered word extraction.The experimental results show that the model in this paper effectively improves the accuracy of event-triggered word recognition.