摘要
为新闻自动生成标题是一个极具挑战的任务。文章基于事件图,提出一种有效的无监督标题生成方法。给定一篇新闻文档,首先为其构造事件图以表示整个篇章,然后采用图排序方法以计算每个事件的显著性得分。随后为排序后的多个事件,抽取其在文中的依存片段作为候选标题,最后设计一个目标优化函数以搜索最终的标题。在英文和中文数据集上的实验结果表明,文章提出的方法能有效地学习显著性事件并能较好地生成标题。
Automatically generating news headline is a challenging task. This paper proposes an effective unsupervised method for this task based on event graph. Given a news report, firstly,a discourse event graph is constructed for and then graph ranking algorithms are used to compute the salient score for each event. Then,the dependency fragment in the text as the candidate title is extracted,and a target optimization func-tion is designed to search the final headline. Experimental results on English and Chinese datasets demon-strate that the proposed method can effectively learn the salient events based on the discourse event graph and generate headlines.
作者
孙锐
SUN Rui(School of Computer Science,Leshan Normal University ,Leshan Sichuan 61400, China)
出处
《乐山师范学院学报》
2017年第4期42-46,共5页
Journal of Leshan Normal University
关键词
事件抽取
互增强原则
标题生成
Event Extraction
Mutual Reinforcement Principle
Headline Generation
