数据驱动的细粒度中文属性对齐方法

发布时间：2018-11-11 15:58

【摘要】：为提高中文开源数据集间属性关系识别的准确率,提出一种数据驱动的细粒度对齐方法,综合利用属性的扩展、定义域等对属性间的同义、包含、相关等关系进行统一识别.方法首先利用统计理论确定属性的数据类型,并给出类型感知的属性相似度计算方法.在此基础上,将属性关系识别建模为多分类问题,抽取有效特征对不同关系进行描述并用于随机森林模型的构建.实验结果表明,该方法中属性数据类型判别的准确率达94.6%,最终对同义、包含、相关关系识别的F1值分别为71.3%,57.3%及59.9%.相比只关注同义属性的传统方法,细粒度属性对齐方法不仅提高了同义属性识别的准确性,而且可识别出相互包含和相关的属性,证明了其在中文开源数据集上的有效性.
[Abstract]:In order to improve the accuracy of attribute relationship recognition among Chinese open source data sets, a data-driven fine-grained alignment method is proposed to identify the synonyms, inclusions and correlations among attributes by using the extension of attributes and the domain of definition. Methods firstly, the data types of attributes are determined by statistical theory, and the method of calculating the similarity of attributes is given. On this basis, attribute relationship recognition is modeled as a multi-classification problem, and effective features are extracted to describe different relationships and to be used in the construction of stochastic forest model. The experimental results show that the accuracy of attribute data type discrimination in this method is 94.60.The F1 values of synonym, inclusion and correlation recognition are 71.37.3% and 59.9%, respectively. Compared with traditional methods which only focus on synonymous attributes fine-grained attribute alignment not only improves the accuracy of synonymous attribute recognition but also recognizes mutually contained and related attributes which proves its effectiveness in Chinese open source data sets.
【作者单位】：中科院空间信息处理与应用系统技术重点实验室;中国科学院电子学研究所;中国科学院大学;
【基金】：国家高技术研究发展计划(863计划)资助项目(2012AA011005)
【分类号】：TP391.1

【相似文献】