ISSN 1000-1239 CN 11-1777/TP

• 论文 • 上一篇    下一篇


王宇1,2 谭松波1 廖祥文1,2 曾依灵1,2   

  1. 1(中国科学院计算技术研究所 北京 100190) 2(中国科学院研究生院 北京 100190) (
  • 出版日期: 2010-09-15

Extended Domain Model Based Named Attribute Extraction

Wang Yu1,2, Tan Songbo1, Liao Xiangwen1,2, and Zeng Yiling1,2   

  1. 1(Institute of Computing Technology, Chinese Academy of Sciences, Bejing 100190) 2(Graduate School of Chinese Academy of Sciences, Bejing 100190)
  • Online: 2010-09-15

摘要: 网页信息抽取是互联网挖掘的重要课题.为了自动化抽取过程,最新的研究利用特定领域的特征,通过机器学习方法对信息抽取过程进行统一建模.但是,对领域特征的依赖使得这类方法难以推广到其他领域中去.因此,对信息抽取问题进行了分析,从中分离出一个可以完全自动化的信息抽取子任务,即有名属性抽取任务.在多个领域的数据集上进行的统计表明,这个子任务覆盖了60%以上的待抽取属性,因此它在整个信息抽取中占有重要地位.并给出了一种基于扩展领域模型的有名属性抽取方法,实验结果表明,这种方法的准确率接近或大于80%,召回率大于90%.

关键词: 信息抽取, 属性抽取, 有名属性, 扩展领域模型, 网页视觉分析

Abstract: Web information extraction is an important task of Web mining. Various applications could benefit from the advancement in this area. These applications include semantic Web, vertical search, sentiment analysis, etc. Current techniques require lots of human interaction which preclude the universal application of Web information extraction. To automate the extraction process, recent research works identify specific features of special domains and extract information by machine learning techniques. However, because of the dependence on specific features, it is very difficult to extend such methods to other domains. In this paper, the Web information extraction problem is analyzed and a subtask is proposed. This new subtask is called named attribute extraction task. Statistics results from multiple datasets prove that named attribute extraction task covers more than 60% attributes in these domains, which show the importance of this subtask. Named attributes are attributes of objects which are encoded in the name-value pair form. That is, the names and values of attributes are settled nearby in the Web pages. Therefore, once the names of attributes are located, the values can be extracted automatically. In this paper, an extended domain model is proposed to summarize attribute names of a domain. And an information extraction method based on this model is developed. Experiments show that the method can extract named attributes at the precision 80%, and at the recall higher than 90%.

Key words: information extraction, attribute extraction, named attribute, extended domain model, visual Web page analysis