差错网络的文本分类反馈校正方法

中文新闻信息分类的类别数量大,难以一次性获取均衡的分类性能。针对这一问题,提出了一种基于差错网络的文本分类反馈校正方法。首先对文本进行一次分类，^根据分类结果生成有向差错网络,得到标注类别与真实类别的候选映射规则。然后计算差错网络的类别关联度参数,再对候选映射规则进行筛选,得到标注类别与真实类别的映射规则。最后根据映射规则进行二次分类，实现分类反馈校正。实验表明，差错网络清晰地反映了类 别的相关度。通过映射关系进行反馈校正，比普通文本分类的F值提高了6.2%。在NLP&CC2014评测中，基于差错网络的方法平均正确率达到73% ,证明了该方法的有效性。

Feedback correction method of text classification based on error network

In study of Chinese news texts classification, it''s hard to achieve a balanced performance because the number of categories. In order to solve the problem, this paper puts forward a feedback correction method of text classification based on error network. First, we performed the first level classification, generated the error network and extracted the candidate mapping rules set between annotation and real categories. Then we calculated the references of type relationship, extracted the mapping rules set between annotation and real categories by screening the candidate mapping rules. At last, we executed the second classification based on mapping rules set and implemented the feedback correction of classification. Experiments showed that the error network reflected the relationship between annotation categories and real categories clearly. It also showed that the method based on feedback correction improved the F measure about 6. 3% than traditional method. The average precision achieved at 73% in NLP&CC 2014 by implementing the method, which proved the effectiveness of the method.