1、文本表示训练过程分类过程训练文本统计统计量特征表示学习分类器新文本特征表示类别kkkjiijijDFNTFDFNTFa2)/log(*)/log(*kkkjiijijDFNTFDFNTFa2)/log(*)0.1log()/log(*)0.1log()log(log11*)0.1log(1NjiijiijijijDFTFDFTFNTFa考虑该特征后的熵)|(log)|()()|(log)|()()(log)()Entropy(Expected)(EntropyGain(t)111tcPtcPtPtcPtcPtPcPcPSSiMiiiMiiiMiit(Kullback-Leibler diver
2、gence)iiitcPtcPtEntropy)|(log)|()(iiiicPtcPtcPtCE)()|(log)|()(ABCDttcc)()()()(),(22DCBADBCACBADNct),()()(212imiiAVGctcPt),(max)(212imiMAXctt)(log)()|(log)()()(log),(BACANAtPctPcPtPctPctImiiiAVGctIcPtI1),()()(),()(max)(1iimiMAXctIcPtI)|()|(log),(jjjjjctPctPtctcctRSJ的概率中出现非的概率中出现类文档个数的为出现jjjjctrctPctP
3、rctTSV,)|()|(log*),()|(log)|(1log()|(1log()|(logjjjjctPctPctPctP是相关的两篇文档yxxtytP,),|(YangYi-mingCCiijCCiijjcjcnnxnxww类C中心向量的权重训练样本中正例个数文档向量的权重22cxw)(ijcjijcjiicxwxwdCSV)()|()()()|()|(jjiijjiijcPcdPdPcPcdPdcP)()|()()()|()|(jjiijjiijcPcdPdPcPcdPdcP,独立性假设rkjikjicwPcdP1)|()|(参数计算Bayes公式1)(|)(1)()()(kkjkk
4、jjjcNccNcNcNccP总文档个数的文档个数kkjijjjijiNNccwcwP不同词个数的次数类所有文档中出现的词在类别文档中出现的次数在1)|(新文本k=1,A类k=4,B类k=10,B类带权重计算,计算权重和最大的类。k常取3或者5。wheat&form WHEATwheat&commodity WHEATbushels&export WHEATwheat&agriculture WHEATwheat&tonnes WHEATwheat&winter&soft WHEAT(粗糙集)RoughSet 逻辑表达式(AQ11算法)学习到如下规则22cxw)(ijcjijcjiicxwxw
5、dCSVOnline Learningijitcjtxywwcj)xw(2ic)()1(类c向量的第j个分量xi的第j个分量Learning RateTarget Value(0 or 1).c1c2cnInput LayerHidden LayerOutput LayerSupport VectorOptimal Separating HyperplanenRegression based on Least Squares Fit(1991)nNearest Neighbor Classification(1992)*nBayesian Probabilistic Models(1992)*
6、nSymbolic Rule Induction(1994)nDecision Tree(1994)*nNeural Networks(1995)nRocchio approach(traditional IR,1996)*nSupport Vector Machines(1997)nBoosting or Bagging(1997)*nHierarchical Language Modeling(1998)nFirst-Order-Logic Rule Induction(1999)nMaximum Entropy(1999)nHidden Markov Models(1999)nError
7、-Correcting Output Coding(1999)n.nPapersK.Aas and L.Eikvil.Text categorisation:A survey.Technical report,Norwegian Computing Center,June 1999 http:/ Xiaomeng Su,“Text categorization”,Lesson PresentationYiming Yang and Xin Liu.1999.A re-examination of text categorization methods.22ndAnnual Internatio
8、nal SIGIR http:/www.cs.cmu.edu/yiming/publications.htmlA Survey on Text Categorization,NLP Lab,Korean U.庞剑峰,基于向量空间模型的自反馈的文本分类系统的研究与实现,中科院计算所硕士论文,2001黄萱菁等,独立于语种的文本分类方法,中文信息学报,2000年第6期nSoftware:Rainbow http:/www-2.cs.cmu.edu/mccallum/bow/BoosTexter http:/ http:/ilk.kub.nl/software.html#timbl C4.5 http:/www.cs.uregina.ca/dbd/cs831/notes/ml/dtrees/c4.5/tutorial.html