摘要
事件因果关系抽取(Event Causality Extraction,ECE)是从文本中抽取出表示因果事件对的事件类型、事件要素及事件间的关系.之前的工作都在含有触发词的文本上进行,并且事件抽取和关系识别也都依靠触发词等事件主体.然而,现实中有许多文本没有触发词,因此该文的抽取任务则是在无触发词标注的文本上进行.该任务的难点在于不仅要抽取多个独立事件,还要判断相互间的因果关系,并且存在事件主体缺失、多事件对及事件类型重叠的问题.该文提出一种分阶段的联合抽取模型,在第一阶段,利用层叠结构模型识别出文本中的事件类型与因果关系;在第二阶段,利用"双定位"和阅读理解机制获得嵌入事件类型信息的句子表示,并通过多层二元标志解码器预测各事件要素的首尾位置.为缓解误差传播问题,该文将两阶段模型通过共享编码层的方式联合训练.实验表明,该文提出的方法可以在完全无规则的情况下有效抽取出无触发词文本中的因果事件对.
Abstract
Event Causality Extraction(ECE)extracts the event types,event arguments and causal relationships be-tween events from the text.In contrast to the previous works relying on triggers in texts,this paper investigate the task on texts without triggers and proposes a two-stage joint extraction model.In the first stage,the paper identifies the event types and causal relationships by the cascade model.In the second stage,the paper obtains sentence repre-sentations embedded with event type information by using dual localization and machine reading comprehension mechanism,and predicts the first and last positions of each event element by using a multi-layer binary tagging de-coder.To alleviate the error propagation,the two-stage model is jointly trained by shared encoding layers.It is shown that the method proposed in this paper can effectively extract causal event pairs from the target-free text in a completely rule-free situation.
维普
万方数据