机器学习演算法特徵的选取与组合概要

機器學習演算法特徵的選取與組合臺灣大學資訊工程學系高紹航臺灣大學外國語文學系高照明詞義辨識PresentedbyPattyLiu本文由高照明老師與高紹航同學撰寫並發表於ROCLING2007，再由PattyLiu於SLPLabmeeting時報告OutlineWordSenseDisambiguationSenseval-2BayesianClassificationForwardSequentialSelectionAlgorithmThefeaturesweappliedResultWordSenseDisambiguationAwordmayhavemorethenonesenseEx.Bank-銀行,河堤,庫ThetaskofWSDistoautomaticallyidentifythecorrectsenseinagivencontext.Senseval-2Publishedin2001Senseval-2Englishlexicalsample73differenttargetwords,includingnouns,verbs,andadjectives.Senseval://193.133.140.102/senseval2/CorporaofSenseval-2Competition：..\corpora\english-lex-sample\train\eng-lex-sample.training.xmlinstanceid=art.40001docsrc=bnc_ACN_245answerinstance=art.40001senseid=art%1:06:00::/contextTheirmultiscreenprojectionsofslidesandfilmloopshavefeaturedinorbitalparties,attheAstoriaandHeaven,inRifatOzbek's1988/89fashionshows,andatEnergy'srecentDocklandsall-dayer.FromtheirresidencyattheFridgeduringthefirstsummeroflove,Halousedslideandfilmprojectorstothrowupacollageofop-artpatterns,filmloopsofdancerslikeE-BoyandWumni,anduniquefractalsderivedfromvideofeedback.&bquo;We'renotawareofcreatingavisualidentifyforthehousescene,becausewe'rerightinthere.Weseeadanceratarave,filmhimlaterthatweek,andprojecthimatthenextrave.&equo;[hi]BenLewis[/hi]Halocanbecontactedon0717383248.[ptr][/p][caption]headArt/headyoucandancetofromthecreativegroupcalledHalo[/caption][/div2][div2][head]/context/instanceBayesianClassificationSupposethetargetwordhasksenses,s1,s2,…,skFinds’suchthatismaximum,cisthecontextorfeaturesofthetargetword)](log)|(max[logarg'iisPscPs)|'(csPP(s’|c))](log)|([logmaxarg)()|(maxarg)()()|(maxarg)()(maxargkkSkkSkkSkSSPScPSPScPcPSPScPcPScPkkkkForwardSequentialSelectionAlgorithmUsedinfeatureselectionFirstletFirstaddthebestfeatureintoSandtheniterativelyaddintoSthebestfeatureintheremainingfeaturesetuntiltheperformancecannotbeimproved.ThefinalSisapproximatelythebestfeaturesetSThefeaturesweappliedWetried9feature,namedF1,F2,…F9Thefeaturesweapplied-F1Thewordsaroundthetargetwordexcludingstopwordssuchas“is”,“a”Bestwindowsizeis3WindowSizePrecision(%)152.7254.2354.6454.6554.1Thefeaturesweapplied-F2SimilartoF1,butincludetheinformationofrelativepositionofthetargetwordIncludestopwordsForexample,“Theartofdesign”{(The,-1),(of,1),(design,2)}WindowSizeTestofF2Bestwindowsizeis1WindowSizePrecision(%)154.9253.6351.1447.9Thefeaturesweapplied-F3SimilartoF2,butusepart-of-speechinstead“Theartofdesign”design:(n,2)Bestwindowsizeis1WindowSizePrecision(%)144.6235.5330.7427.7Thefeaturesweapplied-F4Ngramscontainingthetargetword.“Theartofdesign”{(The-art),(art-of),(The-art-of),(art-of-design),(The-art-of-design)}Bestwindowsizeis3WindowSizePrecision(%)148.2256.9357.8457.8557.8Thefeaturesweapplied-F5SimilartoF4,butusepart-of-speechinsteadsuchas(n-prep-n)forart-of-designBestwindowsizeis4WindowSizePrecision(%)148.2252.1353.8454.2554.1Thefeaturesweapplied-F6UsewordsketchinthesketchenginetoextractallpossiblecollocationsinvolvingthetargetwordBestwindowsizeis5Bestdependencytypeis{modifiers,object,n_modifier,a_modifier,and/or,modifier}WindowSizeTestofF6WindowSizePrecision(%)WindowSizePrecision(%)150.51151.6251.11251.4351.61351.4451.81451.1552.01550.8652.0751.8851.5951.41051.5MinimumSalienceTestofF6MinimumSaliencePrecision(%)0.052.01.051.82.051.83.051.3F6Step1TypePrecision(%)TypePrecision(%)object49.2and/or49.4object_of47.5pp*49.4subject48.4possessor46.6subject_of48.0possessed47.6a_modifier48.1modifier48.4n_modifier49.0part*48.5modifies50.1*comp_of48.3*comp48.2F6Step2TypePrecision(%)TypePrecision(%)object51.1and/or50.8object_of50.1pp*50.9subject50.2possessor50.1subject_of50.1possessed50.0a_modifier50.3modifier50.2n_modifier50.7part*50.2*comp50.1*comp_of50.1F6Step3TypePrecision(%)TypePrecision(%)object_of51.2and/or51.5subject51.1pp*51.4subject_of51.1possessor51.1a_modifier51.3possessed51.0n_modifier51.5modifier51.2*comp51.1part*51.1*comp_of51.2F6Step4TypePrecision(%)TypePrecision(%)object_of51.6and/or51.7subject51.5pp*51.6subject_of51.5possessor51.5a_modifier51.7possessed51.5*comp51.4modifier51.4*comp_of51.4part*51.3F6Step5TypePrecision(%)TypePrecision(%)object_of51.6and/or51.9subject51.7pp*51.8subject_of51.7possessor51.6*comp_of51.7possessed51.7*comp51.8modifier51.8part*51.7F6Step6TypePrecision(%)TypePrecision(%)object_of51.9pp*51.9subject51.9possessor51.9subjec