基于结果修剪法解决文本分类中属性重叠问题

在文本分类中,当两个多属性类别发生属性重叠时,采用传统的文本分类算法m acro F1值仅为45%左右.为了提高文本分类算法的m acro F1值,提出了基于结果修剪的方法.在该方法中,分类器由多个子分类器组成.每个子分类器对应于类别中的一个属性;在每一个阶段中,每一个子分类器将不属于该属性的文本剔出.当所有子分类器运行结束后,留下的文本即属于该分类的文本.实验数据表明,基于结果修剪的文本分类方法在解决属性重叠问题时能够将m acro F1值提高到65%左右.

Result-prune based solution for overlapped attributes confusion in multi-attribute text categorization

In practical text categorization task,some categories contain more than one attribute.When the attributes are overlapped from any two such categories;the traditional text categorization method results in an unacceptable accuracy.The macro F_1 value is only 45%.In order to improve the macro F_1 value,the result-prune method is proposed.By this method,the classifier is divided into several sub-classifiers.Each sub-classifier matches an attribute.In each phase,the sub-classifier discards the text which doesn't belong to this sub-classifier.When all the texts are filtered by these sub-classifiers,the correct texts are left.The experiment shows that this method can increase the macro F_1 value of such text categorization task to 65%.