Abstract: Category distribution in a training set has a major effect on the quality of text categorization. Based on the study of data sets asymmetry for two types of text categorization, a feature weight adjustment method based on feedback information is proposed. In this method, in order to estimate the true influence of small categories in the training process, word weight is adjusted according to the number of correct classified and misclassified training samples. Experimental results show that the method is effective in resolving data sets asymmetry on the impact of text categorization, and the classification quality is observably improved.

Key words: data asymmetry, feedback information, weight adjustment, iterative

摘要： 训练集的分布对文本分类质量有重要影响。该文对两类文本分类中的数据集偏斜问题进行研究，提出一种基于反馈信息的特征权重调整方法，该方法综合考虑正确分类和错误分类的文本数来调整词的权重，以降低训练过程中对小类别的不公平待遇。实验结果表明，该方法有效地解决了数据集偏斜对文本分类的影响，分类质量得到提高。

关键词: 数据偏斜, 反馈信息, 权重调整, 迭代

CLC Number: 

• TP391