一种基于时空分析的事件抽取方法

新闻媒体及社交网站每天呈现大规模的时空文本数据,人们难以从中获取有价值的事件信息。针对前人方法依赖大量的标注数据,同时以孤立的方式考虑事件的时间要素和空间要素等的问题。该文提出一种基于时空分析的事件抽取方法,该方法首先引入数据立方体结构存储事件信息,用户可基于不同的时空粒度抽取出重要的事件;然后提出一种基于语义相似性的实时事件聚类算法,该聚类算法采用GloVe 模型训练词的语义关联性,使聚在同一事件类的事件元素具有强的语义关联性。在大量未标注的网络文本中,该方法取得了77.4%的F1值,表明了该方法能够实现时空分析下的事件抽取任务。

Approach of event extraction based on spatio-temporal analysis

News media and social networking sites present a great volume of spatio-temporal text data every day. It is difficult for people to get valuable event information. Aiming at the existing methods which depend on a large number of annotation data and consider the temporal elements and spatial elements of events in an isolated way, this paper proposes an event extraction approach based on spatio-temporal analysis. Firstly, a data cube structure was introduced to store event information. Users can extract important events based on different temporal and spatial granularity. Secondly, a real-time event clustering algorithm based on semantic similarity is proposed. In the clustering algorithm, a GloVe model was adopted to learn the vector space of words, so that the event elements which collected in the same event cluster have strong semantic relevance. According to a large number of unlabeled network text, the approach achieves a 77.4% value of F1 which indicates that it can achieve event extraction task based on spatio-temporal analysis.