



下載本文檔
版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進行舉報或認領(lǐng)
文檔簡介
關(guān)系tri-training:利用無標記數(shù)據(jù)學習一階規(guī)則Tri-Training:UsingUnlabeledDatatoLearnFirst-OrderRules
Abstract:Knowledgeengineeringhasemergedasasignificantbranchofartificialintelligence,howevermanualknowledgeengineeringislaborintensiveandtimeconsuming.Toreducetheeffortexpendedbyexperts,tri-trainingisintroducedtoautomaticallylearnfirst-orderrulesfromunlabeleddata.Tri-trainingemploysthreeclassifiersinaniterativeprocessthatusesunlabeleddataandself-trainingtogeneratelabelsthatarethenusedtocorrectmislabeledinstances.Ourpaperpresentsastudyoftri-trainingbycomparingitwithotherexistingtechniques.Experimentsusingtworealworlddatasetsshowthattri-trainingperformswellandconsistentlyyieldshigherpredictionaccuracythansingle-sourcelearning,baggingandboostingmodels.
Keywords:knowledgeengineering,artificialintelligence,tri-training,first-orderrules,unlabeleddata,self-training
Introduction:KnowledgeengineeringisanimportantbranchofArtificialIntelligence(AI)thatdealswithdevelopingandmanagingknowledge-basedsystems.Itistypicallyalaboriousandtimeconsumingtask.Traditionalknowledgeengineeringgenerallyrequiresmanualacquisitionofknowledgefromexpertswhichiscostlyandinefficient.Toreducethecostandtimeassociatedwithmanualknowledgeengineering,researchersareexploringwaystoautomaticallylearnfromunlabeleddata.
Oneofthemethodsproposedrecentlyistri-training.Tri-trainingisasemi-supervisedapproachthatusesunlabeleddatatolearnfirst-orderrules.Itutilizestheuncertainlabelingofthreebaseclassifiersandtheirconsensustolabeladditionalunlabeledinstances.Ititerativelybuildsthreeclassifiersonlabeledandpartiallylabeleddata.Italsoemploysself-trainingtoimprovetheaccuracyoftheclassifiers.
...(restofthepaperomitted)Tri-traininghasbeenusedinmanyapplicationssuchastextcategorization,NamedEntityRecognition(NER),andpredictinguserpreferences.Intextcategorization,tri-trainingcanbeusedtoclassifyshorttextsordocumentsintomultiplecategories.Itachievesahigherpredictionaccuracythansingle-sourcelearningandbaggingmodels.Tri-traininghasalsobeenappliedinNERtaskforautomaticallyidentifyingentitiesfromtext.Itimprovesthelabelingprocessbyprovidinglabelsfornoisydatawhichisknowntoreducethelabelingcostsignificantly.Anotherapplicationoftri-trainingistopredictusers'preferencesbylearningfromuserbehaviors.Thismethodhasbeenusedtobuildrecommendersystemsthatcanaccuratelypredictuser'spreferencesbasedonthehistoryoftheiractivities.
Overall,tri-trainingisapromisingapproachforsemi-supervisedlearningwithunlabeleddata.Experimentsusingtworealworlddatasetshaveshownthattri-trainingperformswellandconsistentlyyieldshigherpredictionaccuracythansingle-sourcelearning,baggingandboostingmodels.However,furtherresearchisneededtoimprovethescalabilityandrobustnessofthealgorithm.Thisincludesdevelopingalgorithmstoselectinformativelabelsandfeaturelearningtechniquestogeneraterepresentativefeatures.Inadditiontotheresearchmentionedabove,thereisaneedforfurtherexplorationontheuseofmoreadvancedmachinelearningtechniquesfortri-training.Forexample,deeplearningtechniquessuchasconvolutionalneuralnetworksandrecurrentneuralnetworkscanbeusedtoimprovetheaccuracyofthemodels.Anotherpromisingareaofresearchisthedevelopmentoftransferlearningalgorithmswhichallowknowledgeacquiredfromonedomaintobeappliedtoanotherdomain.Thiscanreducetheamountofdatarequiredforeffectivelearningandcanimprovetheaccuracyofthemodelsignificantly.
Additionally,currenttri-trainingalgorithmsrelyonafixedsetofinputfeatures.Tofurtherimprovetheaccuracyofthemodel,featureselectionalgorithmscanbeusedtoselectthemostinformativefeatures.Thiswillreducethenumberofirrelevantfeaturesandcanimprovetheaccuracyofthemodel.Finally,animportantareaofresearchisthedevelopmentofalgorithmsthatcanautomaticallydeterminethebestcombinationoflabelsandfeatureswhichwouldenableamoreefficientknowledgeacquisitionprocess.
Inconclusion,tri-trainingisapowerfulsemi-supervisedapproachthatcanbeusedtolearnfromunlabeleddata.Withfurtherresearchandimprovements,ithasthepotentialtoreducethecostandtimeassociatedwithknowledgeengineering.Tri-trainingisaneffectivesemi-supervisedlearningapproachwhichutilizesunlabeleddatatogeneratefirst-orderrules.Ithasbeenusedforavarietyoftaskssuchastextcategorization,NamedEntityRecognition(NER),anduserpreferenceprediction.Experimentshavedemonstratedthattri-trainingcansignificantlyoutperformsingle-sourcelearning,baggingandboostingmodels.
Furtherresearchisneededtoimprovethescalabilityandrobustnessofthealgorithm.Somepotentialareasofexplorationincludetheuseofdeeplearningtechniques,transferlearningalgorithmsandfeatureselectionalgorithms.Additionally,algorithmstoautomaticallydeterminethebestcombinationoflabelsandfeaturescouldfurtherimprovetheaccuracy
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負責。
- 6. 下載文件中如有侵權(quán)或不適當內(nèi)容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 2025年電梯廣告投放合同
- 2025深圳經(jīng)濟特區(qū)股權(quán)轉(zhuǎn)讓合同范本
- 2024年榆林高專附中教師招聘真題
- 購房定金協(xié)合同范本
- 2024年紹興嵊州市人民醫(yī)院招聘真題
- 2024年平湖市市屬事業(yè)單位考試真題
- 2024年樂山市五通橋區(qū)招聘事業(yè)單位工作人員真題
- 設(shè)立分公司合作合同(2025年版)
- 2024年安仁職業(yè)中專專任教師招聘真題
- 2024年安徽亳州技師學院專任教師招聘真題
- JJF 1271-2010公路運輸模擬試驗臺校準規(guī)范
- GB/T 22795-2008混凝土用膨脹型錨栓型式與尺寸
- GB/T 19851.15-2007中小學體育器材和場地第15部分:足球門
- GB/T 10095.1-2001漸開線圓柱齒輪精度第1部分:輪齒同側(cè)齒面偏差的定義和允許值
- 全國戒毒醫(yī)療機構(gòu)名單
- ICU 呼吸機相關(guān)性肺炎預防措施執(zhí)行核查表
- 汽車吊檢測保養(yǎng)記錄
- 市政工程安全臺賬表
- 航天模型的設(shè)計、制作與比賽課件
- 公路工程施工現(xiàn)場安全檢查手冊
- 高考倒計時60天課件
評論
0/150
提交評論