事件句的句法结构有助于语义理解。针对中文领域的事件检测任务, 本文设计了面向句义及句法的事件检测模型(BDD)以增强对事件句的理解能力。以基于来自变压器的双向编码器表示 (BERT)的动态词向量为信息源, 设计基于依存树的长短时记忆网络模型(D-T-LSTM)以融合学习句法结构及上下文语义, 并加入基于依存向量的注意力机制强化对不同句法结构的区分度, 在中文突发事件语料库(CEC)上的实验证明了本文模型的有效性, 精确率、召回率、F1值均靠前, 且 F1值比基准模型提升了 5.4%, 召回率提升了 0.4%。

The syntactic structure of event sentences contributes to semantic understanding. A novel event detection model called BERT(Bidirectional Encoder Representations from Transformers) +D (Dependency)-T(Tree)-LSTM(Long Short-Term Memory network)+D-Attention(BDD) is proposed, which aims to learn semantic and syntactic representation of sentences jointly to enhance the event-sentence understanding ability. Taking the word vector based on BERT as the information source, D-T-LSTM model is designed to integrate the learning of syntactic structure and sentence semantics. An attention mechanism based on the dependency vector is added to strengthen the distinction of different syntactic structures at the aim of event detection. Experiment results on the Chinese Emergency Corpus(CEC) prove the effectiveness of BDD. The precision, recall and F1 value of BDD are rather optimum, and the F1 value is 5.4% higher than that of the benchmark model, and the recall rate is 0.4% higher.