大規(guī)模信息過(guò)濾技術(shù)研究及其在Web問(wèn)答系統(tǒng)中的應(yīng)用.pdf

1、中國(guó)科學(xué)院研究生院博士學(xué)位論文大規(guī)模信息過(guò)濾技術(shù)研究及其在Web問(wèn)答系統(tǒng)中的應(yīng)用姓名：許洪波申請(qǐng)學(xué)位級(jí)別：博士專(zhuān)業(yè)：計(jì)算機(jī)軟件與理論指導(dǎo)教師：白碩20030701人規(guī)模信息過(guò)濾技術(shù)研究及其經(jīng)Web問(wèn)答系統(tǒng)中的應(yīng)用：AbstractResearchonTechniquesofLarge—scaleInformationFilteringanditsApplicationinWebQuestionAnsweringSystemXuHongb

2、o(ComputerApplicationTechnology)DirectedByProfessorBaiShuoThisdissertationaddressesthetechniquesofInformationFilteringinknowledgediscoveryanddatamining，makesathoroughanalysisonsomekeyproblemsinInformationFiltering，especi

3、allyinAdaptiveFiltering，andproposedalarge—scaleandhighcapableintegratedinformationfilteringmethodThismethodproducedthebestresultsamongcurrentinformationfilteringsystemsUserrequirementanduserinterestprofilearethebasisofIn

4、formationFilteringThedissertationintroducesdifferentuserrequirementexpandingmethodsusingtraditionalthesaurus—basedandpseudo—relevancefeedbacktechniques，thenproposedatwolayerprofileinitializationmethodbasingonanunbiasedse

5、lectingtechniqueforpseudorelevantdocumentsTosolvetheproblemthat“smalltopic”ishardtofiltering，thedissertationproposeda‘‘smalltopic”automaticidentificationandoptimizationmethodThismethodeffectivelyimprovesthefilteringperfo

6、rmanceon“smalltopic”Aftersummarizingandanalyzingtraditionalmethodsoffeatureselection，thisdissertationproposedaflexiblefeatureselectionmethodaimingattopicgranularityThismethodautomaticallydividesoriginaluserrequirementsin

7、tudetopic(requirementwithbiggranularity)anddetailedtopic(requirementwithsmallgranularity)，andthenselfdeteminesthebestfeatureselectionmethodaccordingtodifferenttopicgranularityThisdissertationalsoinvestigatesthesmoothingt

8、echniquesintermweightingThisdissertationdiscussestheproblemlearningfromuncertaintyinformationinAdaptiveFiltering，andproposedanadaptivelearningmethodusinguncertaintyinformationItthoroughlyexplorestheeffecttoperformancebya

9、pplyingdifferentmethodsdealingwiththeunjudgeddocumentsforprofileupdating，andfinallyachievesthemostrobustandeffectiveprofileupdatingmethodThresholdoptimizationisoneofthemostimportantanddifficultproblemsinAdaptiveFiltering

10、Thisdissertationpointsoutgeneraldrawbacksofcurrentthresholdoptimizationmethods，andthenproposedatarget—orientedthresholdoptimizationmethod，takingtheevaluationmeasureitselfasthetargetfunctionforoptimizingMeanwhile，thedisse

11、rtationalsoaddressesholisticandlocaltargetfimctionoptimizationstrategies，summarizestheiradvantagesanddisadvantages，andthencomparesintheroundthedifferentperformanceswiththemethodsthatholisticorlocaltargetoptimizationcontr