基于类别层次结构的多层文本分类样本扩展策略  

Expanding Training Dataset with Class Hierarchy in Hierarchical Text Categorization

作　　者：李保利[1] 

机构地区：[1]河南工业大学计算机科学系,郑州450001

出　　处：《北京大学学报（自然科学版）》2015年第2期357-366,共10页Acta Scientiarum Naturalium Universitatis Pekinensis

基　　金：河南省基础与前沿技术研究项目(112300410007);河南工业大学高层次人才基金(2012BS027)资助

摘　　要：针对大规模多层文本分类训练样本获取代价高、类别分布不均衡等问题,提出并比较几种基于类别层次结构的大规模多层文本分类样本扩展策略,即利用类别层次体系中蕴含的类别名称、描述以及类别间的层次结构关系,从内涵和外延两方面入手构造或扩展类别训练样本。在首次大规模中文新闻信息多层分类评测数据集上,基于外延的局部样本扩展策略取得较好的性能。参测系统在第一级类别和第二级类别上宏平均F1分别为0.8413和0.7139,在10个参赛系统中位列第二。As the number of classes is quite large in a hierarchical text categorization problem, it usually costs much to obtain a training dataset of reasonable size and sample distribution. Several strategies are proposed and compared to generate new training samples from the class hierarchy in a hierarchical text classification problem. These solutions try to make full use of the class hierarchy （including class names, their descriptions if any, and relationships between them）, and derive new pseudo training samples based on connotations and extensions of classes. Experiments on the dataset of the first large scale Chinese News Categorization at NLPCC 2014 show that the localized expanding strategy based on class extensions performs better. The proposed official system achieved MacroF1 0.8413 and 0.7139 at level 1 and level 2 respectively, which ranked the proposed system the second place among the 10 participating systems.

关 键 词：多层文本分类 大规模中文新闻分类 中文新闻信息分类 类别层次体系 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]