基于特征隶属度的文本分类相似性度量方法  

Similarity Measure for Text Classification Based on Feature Subjection Degree

作　　者：池云仙 赵书良[2] 罗燕[2] 赵骏鹏 高琳[2] 李超[2] 

机构地区：[1]河北师范大学资源与环境科学学院,石家庄050024 [2]河北师范大学数学与信息科学学院,石家庄050024

出　　处：《计算机科学》2017年第11期289-296,共8页Computer Science

基　　金：国家自然科学资金项目(71271067);国家社科基金重大项目(13&ZD091);河北省高等学校科学技术研究项目(QN2014196);河北师范大学硕士基金(xj2015003)资助

摘　　要：基于相似性进行文本分类是当前流行的文本处理方法。基于特征隶属度的文本分类相似性度量方法旨在利用特征与文档间的隶属关系度量文档相似性,从而实现文本分类。该方法基于特征与文档的隶属关系,对特征进行全隶属、偏隶属和无隶属词集划分,并基于3种隶属词集定义隶属度函数。全隶属词集隶属于两篇文档,隶属度随权差增大而降低;偏隶属词集仅隶属于其中某一篇文档,隶属度为一个定值;无隶属词集与两篇文档无隶属关系,隶属度为零。在度量相似性时,偏隶属关系高于全隶属关系。由于同类文档词集相近,异类文档词集差异明显,因此,基于特征与文档的隶属度进行相似性度量,可清晰界定词集与类别的隶属关系,提升分类精度。最后,采用数据集20-Newgroups和Reuters-21578对分类有效性进行验证,结果表明基于特征隶属度的相似性度量方法的性能优于目前流行的相似性度量方法。It is a fashionable method to do text classification based on similarity.Algorithm similarity measure for text classification based on feature subjection degree（SMTCFSD）aims at measuring similarity of documents through subjection relationship between feature words and documents.Algorithm SMTCFSD divides words into total subjection word sets,partial subjection word sets and none subjection word sets based on the subjection relationship,and defines subjection function based on three subjection word sets.Total subjection word sets subject to two documents,and subjection degree will decrease when the differences between two total subjection words increase.The words that only belong to one of the two documents are subsumed into partial subjection word sets,in which subjection degree is a definite value.Subjection degree of none subjection word sets is zero,because the words subject to neither of two documents.Total subjection relationship is more important than partial subjection relationship for similarity measure.Due to word sets of documents in the same category is similar to each other,while the ones in different categories have great distinction,classification accuracy will be promoted obviously based on different values of feature words,which are decided by subjection degree.Algorithm SMTCFSD is superior to the widely used similarity measure methods through experimental results on data sets from Reuters-21578 and 20-Newsgroups.

关 键 词：数据挖掘 文本分类 相似性度量 隶属度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]