基于图神经网络和句法结构的因果事件检测

Event causality identification based on graph neural network and syntactic structure

事件因果关系识别(ECI)旨在识别非结构化文本中事件之间的因果关系。这是一项非常具有挑战性的任务,因为因果关系通常通过事件之间的隐性关联来表达。现有的方法通常是直接使用预训练语言模型对文本建模来捕获这种关联,而这种方法难以捕获文本中长间隔事件对的语义关系。随着文本中事件对间隔的增加,模型对于事件对的语义关系捕捉能力下降。通过对文本的语义结构建模来研究事件之间的隐式因果关联。利用基于图神经网络的事件聚合器集成事件对的结构信息,以及利用文本的句法结构信息对模型提取的事件对表征进行语义增强。在两个广泛使用的数据集上的实验结果表明,模型性能比基线方法有显著的提升。

Event Causality Identification(ECI)aims to identify causal relationships between events in unstructured text.This task is extremely challenging,as causal relationships are often expressed through implicit associations between events.Existing methods typically involve directly modeling text with pre-trained language models to capture this association,which struggles to grasp the semantic relationships of event pairs over long text spans.As the distance between event pairs in the text increases,the model’s ability to capture their semantic relationship weakens.By modeling the semantic structure of the text,implicit causal associations between events are studied.The use of graph neural network-based event aggregators integrates structural information of event pairs,and the syntactic information of the text is used to semantically enhance the representations of event pairs extracted by the model.Experimental results on two widely used datasets show that the model's performance significantly outperforms baseline methods.