一种基于Bootstrapping构建训练语料的方法.pdfVIP

ISSN 1 计算机研究与发展 1000．1239|CN1-1777／TP of Researchand Computer Journal Development 一种基于Bootstrapping构建训练语料的方法 尹继豪1 樊孝忠1 刘士宁2 于江德1’3 1(北京理工大学计算机科学技术学院北京100081) 2(青岛市国土资源和房屋管理局青岛266002) 3(安阳师范学院计算机科学系安阳455000) (yin__jhgg@sina．com) ConstructionBasedon Corpus Training BOOtstrapping Yin Yu Jiha01，Fan Shiningr2，andJiangdel，3 Xiaozhon91，Liu 1(School Scienceand Institute 100081) ofComputer Technology，BeOingofTechnology，Beijing LandResourcesandHouseAdrninistration 2(QingdaoCity Bureau，Qingdao266002) 3(Departmentof 455000) ComputerScience，AnyangTeachers’College，Anyang AbstractAn for constructionis isbasedon approachtrainingcorpus presented，whichbootstrapping smallsubsetisselected annotationsare aseedsetis algorithm．A randomly，thewrong corrected，and seedsettrainsoneclass—based model．andthenthismodelannotatestherest generated．This language smallsubsetsareselectedfromtherest isiterateduntilthe corpus．Other corpus．Thisprocessing quantity of is methodcanminimizethehumaneffortwhile the of trainingcorpusoptimization．This keepingquality theannotation forstatistical model． reasonably good language words model Key bootstrapping；namedentityrecognition；trainingcorpus；class—basedlanguage 部分语料，人工修正后生成种子集，用该种子集训练一个基于类的语言模型，然后使用该模型自动标注 剩余的语料；再从剩余语料中选取部分语料进行以上处理，如此循环直到训练语料标注质量理想．实验 结果表明，该方法在保证训练语料标注质量理想的情况下，能够大幅度地减少人工参与