基于核心句的端到端事件共指消解

End-to-End Event Coreference Resolution Based on Core Sentence

大多数先前的事件共指消解模型都属于成对相似度模型,通过编码两个事件提及的表示并计算相似度来判断是否共指。但是,当两个事件提及在文档内出现的位置接近时,编码其中一个事件提及的上下文表示会引入另一事件的信息,从而降低模型的性能。针对此问题,提出了一种基于核心句的端到端事件共指消解模型(End-to-end Event Coreference Resolution Based on Core Sentence,ECR-CS),该模型自动抽取事件信息并按照预先设置好的模板为每个事件提及构造核心句,利用核心句的表示代替事件提及的表示。由于核心句中只包含单个事件的信息,因此所提模型可以在编码事件表示时消除其他事件信息的干扰。此外,受到事件信息抽取工具的性能限制,构造的核心句可能会丢失事件的部分重要信息,提出利用事件在文档中的上下文表示来进行出弥补。所提模型引入了一种门控机制,将上下文嵌入向量分解为分别与核心句嵌入向量平行和正交的两个分量,平行分量可以认为是与核心句信息维度相同的信息,正交分量则是核心句中不包含的新信息。通过上下文信息和核心句信息的相关度,控制正交分量中被用来补充核心句中缺失的重要信息的新信息的量。在ACE2005数据集上进行实验,结果表明,相比最先进的模型,ECR-CS的CoNLL和AVG分数分别提升了1.76和1.04。

Most previous event coreference resolution models belong to pairwise similarity models,which judge whether the two events are coreferences by calculating the similarity between them.However,when two event mentions appear close to each other in the document,encoding one event contextual representation will introduce information from the other event,which degrades the performance of the model.To solve the problem,an end-to-end event coreference resolution method based on core sentence(ECR-CS)is proposed.The model automatically extracts event information and constructs a core sentence for each event mention according to the preset template,and uses the core sentence representation instead of the event representation.Since the core sentence contains only the information of a single event,the model can eliminate the interference of other event information when encoding the event representation.In addition,limited by the performance of event extraction,the core sentence may lose some important information of the event.The contextual representation of the event in the document is used to make up for this problem.To supplement the missing important information in the core sentence with the contextual information,a gated mechanism is introduced to filter the noise in the contextual representation.Experiments on dataset ACE2005 show that the CoNLL and AVG scores of ECR-CS improves by 1.76 and 1.04,respectively,compared with the state-of-the-art baseline model.