融入依存句法信息的事件时序关系识别

Event Temporal Relation Identification with Dependency Information

事件时序关系识别有助于读者理清文章脉络,把握全局发展趋势,是重要的自然语言理解任务之一。现有的事件时序关系识别方法专注于提取事件触发词前后的局部信息,然而事件句中的事件信息分布较为分散,导致模型在编码过程中丢失部分事件信息。针对上述问题,针对文本特征提出一种双路依存注意力机制来聚合事件句信息,通过单词的父子节点信息构建出双路依存矩阵,将句法信息融入到词嵌入中。将该机制与双向长短期记忆网络(bidirectional long short term memory,Bi-LSTM)结合,可以使事件时序关系模型的性能得到显著提高。该文在越南语数据集与英语数据集上进行对比实验,结果表明所提方法优于主流的神经网络方法。

Event temporal relation identification is an important natural language understanding task.This task helps readers to analyze the content of the article and clarify the development of events.Most of existing event temporal relation identification methods focus on extracting the semantic information around the trigger,but the information in the event sentence is relatively scattered,which leads the model to lose part of the event information.To address the above problems,a bidirectional dependency attention mechanism is proposed to aggregate event information.A bidirectional dependency matrix is constructed based on the parent-child node information of words,and syntactic information is integrated into word embeddings.Combining this mechanism with Bi-LSTM(bidirectional long short term memory)network,this model can significantly improve the performance of event temporal relation identification.It conducts comparative experiments on Vietnamese datasets and English datasets,the results show that this method is superior to the mainstream neural network methods.