基于触发词语义选择的Twitter事件共指消解研究

Selective Expression Approach Based on Event Trigger for Event Coreference Resolution on Twitter

随着社交媒体的发展与普及,如何识别短文本中事件描述的共指关系已成为一个亟待解决的问题。在传统的事件共指消解研究中,需要从NLP工具和知识库中获得丰富的语义特征,这种方式不仅限制了领域的扩展性,而且还导致了误差传播。为了打破上述局限,提出了一种新颖的基于事件触发词来选择性表达句子语义的方法,以判断短文本中事件的共指关系。首先,利用双向长短记忆模型(Bi-LSTM)提取短文本的句子级语义特征和事件描述级语义特征;其次,通过在句子级特征上应用一个基于事件触发词的选择门来选择性表达句子级语义,以产生潜在语义特征;然后,设计了触发词重叠词数和时间间隔两个辅助特征;最后,通过融合以上特征形成一个分类器来预测共指关系。为评估上述方法,基于Twitter数据标注了一个新的数据集EventCoreOnTweets(ECT)。实验结果表明,与两个基准模型相比,提出的选择性表达模型显著提升了短文本共指消解的性能。

With the development and popularization of social media,how to recognize the coreference relation between two event mention in short texts is an urgent issue.In traditional researches about event coreference resolution,a rich set of linguistic features derived from pre-existing NLP tools and various knowledge bases is required,which restricts domain scalability and leads to the propagation of errors.To overcome these limitations,this paper proposed a novel selective expression approach based on event trigger to explore the coreference relationship on Twitter.Firstly,a bi-direction long short term memory(Bi-LSTM)is exploited to extract the features at sentence level and at mention level.Then,the latent features are generated by applying a gate on sentence level features to make it selectively express.Next,two auxiliary features named the overlapped words of trigger and time interval are designed.Finally,all these features are concatenated and fed into a simple classifier to predict the coreference relationship.In order to evaluate this method,this paper annotated a new dataset EventCoreOnTweet(ECT).The experimental results demonstrate that the selective expression approach significantly improves the performance of coreference resolution of short texts.