使用层次结构改善平面文本分类器的性能

与以往的层次化分类不同,本文使用了一种本质为图的层次结构,利用这种层次结构解决平面分类问题,从而提高平面分类的查准率和查全率.在普通的类别层次结构中,同一父类的兄弟类别之间的混淆关系是对称的,但事实上类别之间的混淆关系不是对称的.本文从分类器的混淆矩阵入手,引入了混淆类别的概念.利用混淆类别构造的类别层次结构,从查准率和查全率的角度来考虑类别之间的关系,表达出了混淆关系的非对称性.实验结果显示,使用类别的混淆类别构建类别层次结构的方法,无论从宏观上还是微观上都可以提高分类的准确率.

USING HIERARCHICAL STRUCTURE TO IMPROVE PERFORMANCE OF FLAT DOCUMENT CLASSIFIER

Different from earlier research on hierarchical classification,in this paper a hierarchical structure,which is a graph in nature,is presented to improve the precision and recall of flat classification.In the general hierarchical structure,confusion relationship among brother classes that have the same parent class is symmetrical.But in fact that's not the case.In this paper,the concept of confusion class is introduced from the viewpoint of confusion matrix of classifier,and a hierarchical structure is built using confusion classes.It describes the asymmetry of confusion relationship from the angle of precision and recall of classifier.Experiment results also show that our method can improve the percision of classification when using confusion categories to build the hierarchical structure of categories.