Meta-classifier approach to reliable text classification(19)

A problem with automatic classifiers is that there is no way to know if a particular classification is just a guess or a certain answer. Reliable classification is the task of predicting whether a certain instance is correctly classified or not, i.e., a cl

3.2.CLASSIFICATIONALGORITHMS

Charactern-grams.Featuresareformedbyallsequencesofncharacters

inatext.Acommonvaluefornis3.Thecharactern-gramrepresentation

wasestablishedquitesomeyearsago[Shannon,1951].Anadvantageof

thisrepresentationisthatitcopesbetterwithspellingmistakes.

Besidesthefeatures,alsothefeatureweightshavetobedetermined.When

afeatureoccursinacertaindocument,ameasurehastodeterminethevalue

ofthisfeatureforthedocument.Someearlyna¨ veBayesapproachesusea

binary-valuedvectorrepresentation[RobertsonandSparck-Jones,1976].The

featurevalueequals1ifthefeatureoccursinthedocument,and0otherwise.

Informationinherentinthefrequenciesoffeaturesislostwhenonlytheabsence

orpresenceoffeaturesisconsidered.Thereforenewapproachesthattakeinto

accountfrequencyoffeaturesinadocumenthavebeendeveloped[Lewis,1998].

Themoststraightforwardadaptationistoconsidertheweightofafeatureequal

tothenumberofoccurrencesofthefeatureinthedocument.Thisweighting

functionisreferredtoastermfrequency(tf)[Lewis,1992].Thetffunction

encodestheintuitionthatthemoreoftenafeatureoccursinadocumentthe

moreitisrepresentativeofitscontent.Thesecondpopularweightingfunction

isthetfidfweightingfunction.Itistheproductofthetermfrequencyandthe

inversedocumentfrequency(idf).Thedocumentfrequency(df)ofafeature

isthenumberofdocumentsinthedatasetinwhichthefeatureoccursatleast

once.Theinversedocumentfrequencycanbecalculatedfromthedocument

frequencyasfollows:

idf(fi)=log|N|

i,where

|N|isthetotalnumberofdocumentsinthedataset,anddf(fi)isthedocument

frequencyoffeature(fi).Theidffunctionencodestheintuitionthatthemore

documentsafeatureoccursin,thelessdiscriminatingitis[Sebastiani,2002,

Smirnovetal.,2003b].Variationsonthetfidfweightingfunctioncanbecre-

atedthroughapplyingdi erentlogarithms,normalizations,andothercorrection

factors.

Inthisresearchweusethefollowingtextrepresentation.Toconvertatext

documentintoaninstancesuitableforclassi cationthevectorofallwords

thatoccurinthedatasetisthefeaturespace.PreviousCBSresearchshows

thatwordsworkwellasfeaturesfortheProfession1dataset,thereforeinthis

researchweusewordsasfeatures[Smirnovetal.,2003a].Thefeatureweightsof

theinstancesarecomputedbyatforatfidfweightingfunction.Weonlyuse

theweightingfunctionsthattakeintoaccountfrequenciesoffeaturesinorder

nottoloseinformationinherentinthefrequenciesoffeatures.

3.2Classi cationAlgorithms

Whenthetextofadocumentistransformedintoacompactrepresentation,in

principleanymachine-learningalgorithmcanbeapplied.Somespecialtextclas-

si cationalgorithmsthatincorporatefeaturede nitionsandweightingfunctions

havebeenproposed,forexampletheRocchioclassi er[Rocchio,1971].Theclas-

si cationalgorithmsthatareexaminedinthisresearch,i.e.,na¨ veBayesand

nearestneighbour,canuseeitherthetforthetfidfweightingfunction.

13