ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚HMM뫍CRFs퓚탅쾢돩좡펦폃훐뗄뇈뷏퇐뺿췵ꆡꆡ뗋죽뫨(쓏뺩듳톧탅쾢맜샭쾵ꆡ쓏뺩210093)ꆡꆡꆾ햪튪ꆿꆡ퓚뇈뷏HMM뫍CRFs쫽톧샭싛뗄믹뒡짏,럖뇰쳡돶믹폚HMM듊뷇즫뇪힢뫍믹폚CRFsퟖ뷇즫뇪힢뗄죋쏻쪵쳥돩좡쒣탍,늢춨맽뾪럅탔닢쫔뫍쪵볹펦폃솽듎퇩횤ꆢ뇈뷏솽헟뗄폐킧탔,듓뛸퓚쪵볹훐횤쏷듓샭싛뇈뷏훐뗃돶뗄뷡싛:CRFs뷏횮HMM룼쫊뫏폚뷢뻶탲쇐뇪힢믲뛔쿳럖샠컊쳢ꆣꆡꆡꆾ맘볼듊ꆿꆡHMMꆡCRFsꆡ탅쾢돩좡ꆡ죋쏻쪵쳥돩좡ꆡ뷇즫뇪힢ꆡ쳘헷ꆡꆡꆾ럖샠뫅ꆿꆡTP311ComparativeStudyonHMMandCRFsApplyinginInformationExtractionWangHaoꆡDengSanhong(DepartmentofInformationManagement,NanjingUniversity,Nanjing210093,China)ꆡꆡꆾAbstractꆿꆡThispaperbringsforwardtwomodelsforperson-nameentityextractionbasedonthecomparisonofmaththeorybetweenHMMandCRFs,oneusingwordrolelabelbasedHMMandtheotherusingcharacterrolelabelbasedCRFs,thenvalidatesandcomparestheeffectofbothbyopen-testingandapplyinginpractice,andtherebyprovesinpracticethatCRFsisfitterforsequencelabelingandobjectclassifyingthanHMM.ꆡꆡꆾKeywordsꆿꆡHMMꆡCRFsꆡInformationextractionꆡPerson-nameentityextractionꆡRolelabelꆡFeature뻶컄놾훐뗄듺듊횸돆컊쳢;1ꆡ틽ꆡ퇔ꆡꆡ(5)쟩뷚쒣냥(ScenarioTemplate),벴룹뻝펦폃쒿뇪뚨틥ꆡꆡ쯦ퟅ췸싧벼쫵뗄랢햹웕벰,죋샠탅쾢헽퓚훰붥쫽ퟖ죎컱뿲볜,폃폚쳘뚨쇬폲뗄탅쾢쪶뇰뫍ퟩ횯ꆣ뮯ꆢ뗧ퟓ뮯ꆣ떽20쫀볍80쓪듺쒩돶쿖쇋ꆰ탅쾢놬햨ꆱ쿖ꆡꆡ탅쾢돩좡럖캪쿔탔횪쪶뗄돩좡뫍틾탔횪쪶뗄췚뻲ꆣ쿳,죋쏇훰붥놻ꆰ퇍쎻ꆱ퓚탅쾢쫽뻝뗄ꆰ몣퇳ꆱ훐ꆣ죧뫎톸쿔탔횪쪶돩좡쫇횸믹폚맦퓲믲믹폚춳볆뗄랽쪽[2]쪵쿖쏼쯙뗘듓몣솿탅쾢훐믱좡폐폃횪쪶,죧뫎돤럖쳡룟탅쾢샻쏻쪵쳥벰웤쿠맘탅쾢뗄돩좡;틾탔횪쪶뗄췚뻲퓲쫇닉폃폃싊,죧뫎뷢뻶ꆰ탅쾢놬햨ꆱ뫍ꆰ횪쪶?랦ꆱ뗄쎬뛜,쯹폐쫽뻝췚뻲(DataMining)벼쫵,듓컄놾훐쳡좡쪵쳥뫍쪵쳥헢킩뚼쫇퓚ꆰ몣솿탅쾢캪쳘헷뗄탅쾢쪱듺ꆱ쳡돶뗄탂뿎횮볤틾님뗄뻟폐튻뚨뿉탅뛈뗄맘쾵ꆣ믹폚춳볆뗄랽램쫇쳢ꆣ탅쾢돩좡벼쫵쫇뷢뻶짏쫶컊쳢뗄폐킧랽램ꆣ횸톧쾰뺭맽뇪힢뗄폯쇏뿢,믱뗃폯퇔톧맦싉횪쪶,닉폃튻ꆡꆡ탅쾢돩좡벼쫵튻냣놻럖뷢캪5룶닣듎[1]:뚨뗄춳볆쯣램,쪶뇰싺ퟣ횸뚨횵뗄쏼쏻쪵쳥[2]ꆣ틾십ꆡꆡ(1)쏼쏻쪵쳥(NamedEntity)쪶뇰,훷튪쫇죋쏻ꆢ뗘쏻ꆢ믺뛻뿉럲쒣탍(HiddenMarkovModel,HMM[3])뫍쳵볾쯦믺릹쏻ꆢ닺욷쏻뗈쏻듊탔쳵쒿,틔벰죕웚ꆢ쪱볤ꆢ쫽ퟖꆢ폊볾뗘횷ꆢ뎡(ConditionalRandomFields,CR[4]Fs)쫇탅쾢돩좡벼쫵믵뇒뗈폐쏻듊뗄폐킧쪶뇰뫍럖샠;훐솽훖뎣폃뗄룅싊춳볆쒣탍,뿉틔폐킧뷢뻶탅쾢돩좡훐ꆡꆡ(2)쒣냥튪쯘(TemplateElement)쪶뇰,벴펦폃쒣냥맦퓲탲쇐뇪힢믲뛔쿳럖샠컊쳢ꆣ놾컄쫔춼뛔HMM뫍CRFs뗄랽램쯑쯷뫍쪶뇰쏼쏻쪵쳥쿠맘탅쾢,듋쪱뒦샭뗄춨뎣쫇튻뷸탐좫쏦웊컶,듓쫽톧샭싛ꆢ뾪럅탔쪵퇩틔벰쪵볹펦폃훐풪맘쾵;뇈뷏솽헟닮틬,캪톡퓱뫏샭뗄룅싊춳볆쒣탍쪵쿖뻟쳥뗄ꆡꆡ(3)쒣냥맘쾵(TemplateRelation)돩좡,벴펦폃쒣냥쯑쯷탅쾢돩좡릤ퟷ쳡릩닎뾼틀뻝ꆣ뫍돩좡쪵쳥폫쪵쳥횮볤뗄맘쾵,뒦샭뗄춨뎣쫇뛾풪맘쾵;ꆡꆡ(4)춬횸맘쾵2ꆡHMM뫍CRFs뗄샭싛쒣탍(Co-reference)쳦뮻,쪵쿖횸듺쿻뷢,벴뷢ꆡꆡ쫕룥죕웚:ꆡ틾십뛻뿉럲쒣탍ꆡꆡ쫕탞룄룥죕웚:2007-10-24ꆡꆡ틾십뛻뿉럲쒣탍쫇횸횧돖캪맛달탲쇐듓ힴ첬벯뫏훐ꆤ57ꆤ' 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚톡퓱뻟폐ퟮ듳뿉쓜탔(ퟮ듳룅싊)ힴ첬탲쇐맽돌뗄쒣탍ꆣꆡꆡ(2)맘폚맛달탲쇐w훐캻훃캪i-1뫍i뗄ힴ첬뇪볇t뗄캪듋틽죫쳵볾룅싊P(T|W),뇭쪾듊뒮W퓚ힴ첬뒮T쿂틆쳘헷fꆣj뗄룅싊,퓲HMM뗄쒿뗄벴캪쟳쪹P(T|W)횵ퟮ듳뗄T,ꆡꆡ쳘헷좨훘ꛋ뫍ꛌ퓲뿉틔닉폃룐횪믺(Perceptron)믲jj볇캪[5]:GIS(GeneralizedIterativeScaling)ꆢIIS(ImprovedIterativeTꇤ=arg#maxP(T|W)(1)Scaling)뗈뗼듺쯣램,춨맽뛔틑횪쫂쪵뷸탐통솷맀쯣ꆡꆡ뛔P(T|W)쪹폃쳵볾룅싊릫쪽,룹뻝놴튶쮹(Bayes)돶살ꆣ릫쪽뛔웤뷸탐뮯볲,좻뫳룹뻝HMM뛀솢탔볙짨뫍뛾풪폯1ꆡ죧맻tu=t,늢쟒wv=wgt(t,w(틥맘쾵u,wv)=,릫쪽(1)ퟮ훕뿉뮯볲캪:0ꆡꆡꆡꆡꆡ웤쯻7)1ꆡ죧맻tu=t1,늢쟒tv=t2Tꇤ=arg#min{ꇆm[-lnP(w|t)-lnP(t|tiiii-1)]}(2)f(ttti=11,t2u,v,w)=(ꆡꆡ웤훐,듊폯쳘뚨ힴ첬쿂룅싊뫍ힴ첬횮볤틆룅싊뿉0ꆡꆡꆡꆡꆡ웤쯻8)틔뷼쯆뗘듓듳맦쒣폯쇏뿢훐춳볆맀쯣ꆡꆡ뛔뇈HMM,CRFs뗄쳘뗣훷튪뇭쿖퓚::쇏훐ꆡꆡ(1)릫쪽(6)뷏횮릫쪽(5)웤쫽톧쒣탍뢴퓓뛈쏷쿔퓶듳;w뗄ힴ첬캪t뗄듎쫽C(w,t)iiiiP(w|t)ꇖ통솷폯ii통솷폯쇏훐ힴ첬t퓙쿖뗄ퟜ듎쫽=Cꆡꆡ(2)폫HMM뷶벯돉쇋2룶횸뚨쳘헷,닎쫽통솷뷶탨춳볆(t)ii3룶욵듎횵쿠뇈,CRFs쓜릻퓚춬튻룶쒣탍훐컞쿞훆벯돉늻춬(3)쳘헷,쳘뇰쫇뿉볓죫풶뻠샫풼쫸,룼쓜뷒쪾폯퇔톧쳘헷,떫P(t|tii-1)ꇖ통솷폯쇏훐ힴ첬t횮잰뗄ힴ첬쫇tC(tii-1뗄듎쫽i-1,t)i통솷폯쇏훐ힴ첬듎쫽=ti-1돶쿖뗄ퟜC(ti-1)CRFs벯돉뛠훖쳘헷짺돉돉잧짏췲룶쳘헷몯쫽,떼훂웤좨훘통(4)솷뗄볆쯣솿튲쫇쿠떱듳뗄;ꆡꆡ폚쫇릫쪽(2)벴뿉틔뮯캪:ꆡꆡ(3)CRFs닉폃솪뫏쳵볾룅싊P(T|W)붨쒣,뇜쏢쇋HMM(w,t)C(tTꇤiii-1,t)i=뗄뛀솢탔볙짨뫍뛾풪볙짨,듓쫽톧붨쒣뷇뛈뛸퇔,CRFs뷏횮arg#min{-ꇆmC[ln+ln]}(5)i=1C(t)C(tii-1)ꆡꆡHMM뻟폐룼뿉뾿룼뫏샭뗄쫽톧췆떼;HMM쯣램뇈뷏돉쫬,퓙볓짏틗폚통솷뗈쳘뗣놻맣랺ꆡꆡ(4)CRFs놣쇴쇋HMM훐뗄횮잰뇪볇뗄ힴ첬뛔떱잰ힴ첬펦폃폚ퟔ좻폯퇔뒦샭쇬폲ꆣ좻뛸폖틲캪HMM쫇짺돉쒣뇪볇뗄펰쿬,쪹쳘헷뗄톡퓱룼캪뫏샭;탍,짐듦퓚훮뛠컊쳢:ꆡꆡ(5)HMM쫇폐쿲춼쒣탍,춨맽Viterbi쯣램쯑쯷떽떱잰뛔ꆡꆡ(1)쯼쎻폐횱뷓뛔쳵볾룅싊P(T|W)붨쒣,뛸쫇닉폃솪뫏쿳캪횹뗄ퟮ병슷뺶,늻뾼싇횮뫳뛔쿳벰웤뇪볇룅싊,뛸CRFs룅싊P(T,W)듺쳦,죏캪맛달횵평ힴ첬짺돉;퓲닉폃컞쿲춼쒣탍,쫇뛔헻룶뇪볇탲쇐쟳뷢솪뫏룅싊,퓚헻룶ꆡꆡ(2)HMM뛀솢탔볙짨죏캪떱잰맛달횵뷶폫떱잰ힴ첬쿠탲쇐랶캧쓚맩튻뮯,뷏HMM뻟폐룼뫏샭뗄쫽톧샭싛믹뒡,춬맘,뫶싔쇋웤폫횮잰ힴ첬뫍횮잰맛달횵횮볤뗄쿠맘탔;쪱튲뇜쏢쇋틲쟳뷢뻖늿맛달횵룅싊쯹듸살뗄뇪볇욫훃(Labelꆡꆡ(3)뷶틔짺돉ힴ첬뫍떱잰맛달횵ퟷ캪뻶뚨쳘헷,늻틗죚B[6,8]ias)컊쳢ꆣ뫏룷훖쳘헷뫍뎤뻠샫틀삵맘쾵ꆣꆡꆡ룹뻝짏쫶럖컶,뇊헟죏캪,CRFs뷏HMM룼쫊뫏폚뷢ꆡꆡHMM짏쫶좱뗣떼훂웤늢늻쫇ퟮ쫊뫏폃폚뷢뻶탲쇐뻶탲쇐뇪힢믲뛔쿳럖샠컊쳢ꆣ쿂쏦뇊헟춨맽헢솽훖쒣탍뇪힢컊쳢뗄룅싊춳볆쒣탍ꆣ퓚죋쏻쪵쳥돩좡훐뻟쳥펦폃뗄뇈뷏,살뷸튻늽퇩횤짏쫶ꆡ쳵볾쯦믺뎡럖컶ꆣꆡꆡ쳵볾쯦믺뎡(CRFs)쫇튻훖뗤탍뗄탲쇐뇪힢에뇰쒣탍,쫇퓚룸뚨맛달탲쇐뗄쳵볾쿂,볆쯣헻룶맛달탲쇐ힴ첬3ꆡ믹폚HMM뗄듊뷇즫뇪힢죋쏻쪵쳥쪶뇰뇪볇뗄솪뫏쳵볾룅싊럖늼뗄컞쿲춼쒣탍ꆣ룹뻝ퟮ듳풭ꆡ쒣탍뚨틥샭[6]뫍컞쿲춼샭싛,CRFs뗄탎쪽뿉틔폃튻룶솪뫏쳵볾룅ꆡꆡ믹폚HMM뗄듊뷇즫뇪힢죋쏻쪵쳥쪶뇰쒣탍(Model싊럖늼P(T|W)살뇭쪾,뚨틥캪[7]:forPerson-NameRecognitionBasedonWordRoleLabelingP(t|w)=Cexp(ꇆꇆꛋf(tꛌg(t,w))ji-1,t,w)+ꇆꇆjijjiijijUsingHMM,HMM_WRL_PnR)죧춼1쯹쪾ꆣ믹놾쮼슷(6)쫇:쫗쿈룹뻝통솷폯쇏훐돶쿖죋쏻뗄릹돉맦퓲,뚨틥웤릹ꆡꆡ웤훐,C=1Þꇆexp(ꇆꇆꛋf(tjji-1,t,w)+ꇆꇆꛌgijjtijij돉뷇즫;룹뻝릫쪽(3)뫍(4)듓통솷폯쇏훐춳볆듊폯퓚쳘(t,w)),쪹뗃ꇆ[p(t|w)]=1ꆣ퓚CRFs훐,쳘헷놻럖캪i뚨뷇즫쿂돶쿖뗄듎쫽ꆢ쳘뚨뷇즫뗄돶쿖듎쫽틔벰죎틢솽t솽샠:뷇즫횮볤뗄쿠쇚듎쫽뗈볆쯣솿;좻뫳닉폃믹폚듊뗤뗄헽ꆡꆡ(1)맘폚캻훃캪i뗄맛달탲쇐w뫍ힴ첬뇪볇t뗄ힴ첬쳘쿲ퟮ듳욥엤쯣램살쪵쿖닢쫔컄놾듖럖틔탎돉듊맛달탲헷g;j쇐,펦폃횮잰듓통솷폯쇏훐믱뗃뗄춳볆솿,볆쯣맛달탲쇐ꆤ58ꆤ' 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚훐듊폯쳘뚨ힴ첬쿂룅싊뫍ힴ첬횮볤뗄틆룅싊,닉폃Vit2춼2쯹쪾:erbi쯣램캪맛달탲쇐쯑쯷ퟮ병뷇즫슷뺶(룅싊뫍ퟮ듳),믱뗃HMM뷇즫솴;ퟮ뫳춨맽뛔HMM뷇즫솴뷸탐뷇즫닰럖뫳,믹폚죋쏻쒣쪽욥엤듓듊뒮맛달탲쇐훐돩좡돶럻뫏횸뚨쒣쪽뗄죋쏻쪵쳥ꆣꆡꆡ쿂쏦틔ꆰ놾놨뫏럊1퓂3죕뗧볇헟쇵뷜놨뗀ꆱ캪샽,살쮵쏷HMM_WRL_PnR뗄뒦샭맽돌ꆣ쫗쿈뛔뻤ퟓ뷸탐컄놾듖럖,탎돉듊뒮탲쇐ꆰ놾놨뫏럊1퓂3죕뗧볇헟춼2ꆡ죋쏻뷇즫톧쾰맽돌쇵뷜놨뗀ꆱꆣꆡꆡퟮ뫳춳볆짺돉<듊폯,뷇즫>폐탲뛔,뗃떽3룶죋쏻쪶뇰ꆡ쒣쪽뮯죋쏻쪶뇰듊뗤:춳볆듊폯wퟷ캪t뷇즫돶쿖뗄듎쫽,C(w,t)ퟷ캪듊폯iiii뷇즫듊뗤;샛볆쯹폐늻춬뷇즫뗄돶쿖듎쫽C(t)ퟷ캪뷇즫듊i뗤;틔벰춳볆죎틢솽뷇즫폐탲쿠쇚뗄듎쫽C(ti-1,t)ퟷ캪뷇즫i틆듊뗤ꆣꆡꆡ(3)쒣쪽욥엤쯣램,죋쏻뷇즫쒣쪽욥엤뗄쪵쿖ꆣ잰컄룹뻝쳘쫢떽튻냣뗄맩쓉램,듓듳맦쒣폯쇏훐랢쿖훐맺죋쏻뗄튻냣폯램쳘헷;듋뫳벴뿉닉폃튻냣떽쳘쫢뗄퇝틯췆샭램,죏캪ퟱ춼1ꆡ믹폚HMM뗄듊뷇즫뇪힢죋쏻쪵쳥쪶뇰톭짏쫶폯램쳘헷뗄뷇즫ퟩ뫏벴캪훐맺죋쏻,뻟쳥늽훨죧쿂:쒣탍(HMM_WRL_PnR)ꆡꆡꋙ펦폃Viterbi쯣램,볆쯣뺭맽듖럖믱뗃뗄훐컄듊뒮쿠뛔펦뗄ퟮꆡꆡ죧춼1쯹쪾,룹뻝HMM병(룅싊ퟮ듳믲릫쪽(6)뗄횵ퟮ킡)뷇즫뒮(HMM솴[5]),죧잰샽뗄_WRL_PnR쪵쿖훐컄컄놾훐뗄죋쏻쪵쳥쪶뇰,훷튪폐3쿮릤ퟷHMM솴캪ꆰAAAAAAAAABEAꆱ;:ꆡꆡꆡꆡꋚU뫍V뷇즫쫇죋쏻뷇즫뫍튻냣뷇즫뗄ퟩ뫏,탨튪붫ퟩ뫏뷇즫벰(1)뷇즫뚨틥ꆣ뾼달훐컄폯쇏훐뗄죋쏻릹돉,랢쿖훐컄죋쏻퓚뷡릹짏뻟폐튻뚨뗄맦싉폫웤뛔펦뗄듊뒮듊wi뛏쇑;:ꆡꆡꋛ룹뻝죋쏻뷇즫뚨틥,죏캪HMM솴훐죎뫎욥엤BBCD,BBE,ꆡꆡꋙ훐맺죋쏻튻냣ퟮ뛠평4룶ퟖ(탕+탕+쏻+쏻)ퟩ돉;BBZ,BBG,BCD,BXD,BE,BG,BZ,CD,FB,XD,Y,Z,B,W뗈뷇즫쒣쪽ꆡꆡꋚ훐맺죋쏻뗄ퟩ뫏쒣쪽퓚쫽솿짏뻟폐폐쿞탔;웤뛔펦뗄듊폯ퟩ뫏벴캪죋쏻ꆣꆡꆡꋛ훐컄죋쏻뗄잰뫳쇙뷧듊뛔죋쏻뻟폐튻뚨뗄쳡쪾ퟷ폃;ꆡꆡꋜ틴틫췢살죋쏻퓚랭틫폃ퟖ짏뻟폐튻뚨뗄훘뢴탔ꆣ4ꆡ믹폚CRFs뗄ퟖ뷇즫뇪힢죋쏻쪵쳥쪶뇰ꆡꆡ룹뻝짏쫶맦싉,뿉틔ퟜ뷡돶HMM_WRL_PnR훐뗄듊뷇즫벯뫏,죧뇭1쯹쪾ꆡ쒣탍뚨틥:ꆡꆡ뇭1ꆡ훐컄죋쏻쪶뇰HMM훐뗄ힴ첬벯뫏(릹돉뷇즫벯뫏ꆡꆡ룹뻝릫쪽(6),튪볆쯣맛달횵쳵볾쿂ힴ첬횵뗄솪뫏룅)싊럖늼,탨튪릹퓬쳘헷몯쫽늢볆쯣웤좨훘ꆣ믹폚CRFs쪵탲뷇즫뷢쫍쪾샽탲뷇즫뷢쫍쪾샽쿖탲쇐뇪힢,맘볼퓚폚뷢뻶4룶컊쳢:맛달횵뗄짨뚨;ힴ1B탕쫏샮쫀쏱,뚫랽쮷9ZCD돉듊샮돉닅,췵쏱몽첬횵뗄뚨틥;쳘헷몯쫽뗄톡퓱;좨훘뗄통솷ꆣ2C쮫쏻쫗ퟖ샮쫀쏱10K죋쏻뗄짏컄샏쪦돆퓞뫺욽ꆡꆡ믹폚CRFs뗄ퟖ뷇즫뇪힢죋쏻쪵쳥쪶뇰쒣탍(Model3D쮫쏻캲ퟖ샮쫀쏱11L죋쏻뗄쿂컄샮쫀쏱랢뚯탾커쏅횮뇤forPerson-NameRecognitionBasedonCharacterRoleLa24E떥쏻췵12M솽룶죋쏻횮볤뗄듊췵뫍췵쫇탖쏃belingUsingCRFs,CRFs_CRL_PnR)죧춼3쯹쪾,믹놾쮼5F탕잰ힺ킡췵,샏헅13UKB돉듊뷌맙뛔냗햹쳃쮵슷쫇:쫗쿈,뛔훐컄컄놾뷸탐풭ퟓ쟐럖;웤듎,믹폚ꆰퟖꆱ풭6G탕뫳ힺ샮ퟜ,쪼탲쇐퇜짺맛달탲쇐,룹뻝맛달탲쇐뗄좡횵붨솢헫뛔죋췵늩14VDL믲EL돉듊뗋펱뎬짺잰폃맽뗄뚫컷쏻쪶뇰뗄짏쿂컄쳘헷쒣냥,늢내쒣냥맦퓲듓통솷폯쇏훐7XBC돉듊웫맺쏷,닩쏷좨15W웤쯼탕쏻듊뿋쇖뛙8YBCD믲BE싞엌,샮쳡좡쳘헷몯쫽;퓙룹뻝통솷폯쇏훐돶쿖뗄죋쏻릹돉맦퓲,돉듊쯄맢16A웤쯼듊닎볓펭탂췭믡뚨틥웤릹돉뷇즫벯뫏;샻폃통솷폯쇏훐뗄맛달탲쇐뫍뷇ꆡꆡ(2)뷇즫듊뗤짺돉쯣램ꆣ룹뻝릫쪽(5),탨튪뛔통솷폯쇏즫탲쇐,믹폚CRFs통솷쳡좡뗄쳘헷몯쫽뗄좨훘틔릹붨폃뿢뷸탐춳볆,ퟜ뷡훐컄죋쏻뗄튻냣폯램맦싉,틔뇣쪹폃헢킩춳폚죋쏻쪶뇰뗄CRFsퟖ뷇즫뇪힢쒣탍,춨맽뗷헻맛달횵좡볆쫽뻝맀쯣맛달듊뒮뗄ퟮ병뷇즫솴ꆣ탨튪춳볆뗄쫽뻝냼삨:횵ꆢ쳘헷쒣냥룶쫽틔벰뷇즫벯뫏뚨틥살뗷뷚쳘헷몯쫽벰룷듊떣죎룷뷇즫뗄듎쫽,룷뷇즫퓚통솷폯쇏훐돶쿖뗄ퟜ듎쫽,웤좨훘틔쪵쿖ퟮ뫏샭뗄붨쒣;룹뻝붨솢뗄쒣탍,볆쯣닢쫔틔벰뷇즫벯뫏훐죎틢솽뷇즫횮볤쮳탲쿠쇚듎쫽ꆣ뻟쳥맽돌죧폯쇏훐쎿튻룶(ퟩ)맛달횵퓚웤짏쿂컄뮷뺳훐놻쪶뇰캪쒳ꆤ59ꆤ' 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚훖뷇즫뗄솪뫏뿉탅뛈,늢톡퓱뿉탅뛈ퟮ듳뗄뷇즫샠뇰ퟷꆡꆡ(2)쳘헷쒣냥뗄톡퓱ꆣ뷶춨맽떱잰뛔쿳뗄맛달횵ퟩ쯹랴캪맛달횵뗄뷇즫뇪힢;ퟮ뫳룹뻝죋쏻쪵쳥뗄뷇즫ퟩ뫏듓펳뗄쳘헷탅쾢뇏뺹폐쿞,폐뇘튪볓죫웤짏쿂컄탅쾢살잿뮯쳘풭쪼맛달탲쇐훐돩좡돶럻뫏횸뚨쒣쪽뗄죋쏻쪵쳥ꆣ헷ꆣ놾컄쓢닉폃뻖늿짏쿂컄탅쾢살풼쫸떱잰뛔쿳ꆣ뻖늿짏쿂ꆡꆡ룹뻝죋쏻쪶뇰맽돌,짏쫶4룶맘볼컊쳢뻍뮯캪컄탅쾢쫇횸틔떱잰맛달뛔쿳캪훐탄,쿲잰믲쿲뫳솬탸톡좡튻:ꆡꆡ(1)틔ꆰퟖꆱ캪떥캻맛달탲쇐ퟩ뗄퇜짺뚨ퟖ뎤랶캧뗄짏쿂컄탅쾢ퟷ캪떱잰뛔쿳뗄쳘헷탅쾢,헢룶뻖;ꆡꆡ(2)훐컄ꆰퟖꆱ뷇즫벯뫏뗄뚨틥늿솬탸랶캧돆캪ퟖ뎤뒰뿚[9]ꆣ죧맻붫맛달횵ퟩ뾴ퟷ쫇튻훖뫡;ꆡꆡ(3)룹뻝맛달횵업쇐맦퓲,톡퓱튻뚨뗄맛달횵ퟩ뫏ퟷ캪쿲뛠풪쳘헷,쓇쎴ퟖ뎤뒰뿚뻍쳡릩쇋튻훖ퟝ쿲짏쿂컄뛠풪풼쳘헷몯쫽쫸ꆣ룹뻝맛달탲쇐ퟩ틔벰뻖늿짏쿂컄탅쾢뗄쳘헷풼쫸,CRFs;ꆡꆡ(4)룹뻝통솷폯쇏톡퓱쳘헷몯쫽늢볆쯣웤좨훘쾵쫽ꆣ_CRL_PnR톡퓱뫏쫊뗄맛달탲쇐ퟩ뫏뫍튻뚨ퟖ뎤뗄짏쿂컄뒰뿚(놾컄닉폃3ퟖ뎤뒰뿚),붨솢쇋죧뇭3쯹쪾뗄쳘헷쒣냥](FeatureTem[10plate)ꆣ뇭훐C뇭쪾떱잰ퟖ,F뇭쪾틴틫죋쏻폃ퟖ쳘헷,S뇭쪾탕쫏쳘헷,P뇭쪾잰쇙뷧ퟖ쳘헷,Q뇭쪾뫳쇙뷧ퟖ쳘헷,W(4)뇭쪾4듊캻뇪힢쳘헷,R뇭쪾뇪힢뷇즫쳘헷,n뇭쪾ퟖ뎤뒰뿚캻틆좡횵ꆣꆡꆡ뇭3ꆡCRFs_CRL_PnR죋쏻쪵쳥쪶뇰훐뗄쳘헷쒣냥맛달횵뇪힢쳘헷쳘헷N-gram쳘헷쒣냥튻풪쳘헷Cn,Fn,Sn,Pn,Qn,Bn,En,Wn,CnFnSnPnQnBnEnWn,n=-1,0,1춼C3ꆡ믹폚CRFs뗄ퟖ뷇즫뇪힢죋쏻쪵쳥쪶뇰n-1Cn,Fn-1Fn,Sn-1Sn,Pn-1Pn,Qn-1Qn,Bn-1CꆢFꆢSꆢPꆢQꆢBꆢR뛾풪쳘헷Bn,En-1En,Wn-1Wn,n=0,1;C쒣탍n-2Cn,Fn-2Fn,(CRFs_CRL_PnR)Sn-2Sn,Pn-2Pn,Qn-2Qn,Bn-2Bn,En-2En,Wn-2EꆢW(4)Wn,n=1;ꆡ죋쏻뷇즫뗄뚨틥n-2Cn-1Cn,Fn-2Fn-1Fn,Sn-2Sn-1Sn,Pn-2Pn-1ꆡꆡ룹뻝죽풪쳘헷Pn,Qn-2Qn-1Qn,Bn-2Bn-1Bn,En-2En-1En,뷚쯹ퟜ뷡뗄훐컄죋쏻릹돉맦퓲,뷡뫏뾼싇Wn-1Wn,n=1듋뒦뛔훐컄떥ꆰퟖꆱ뷸탐뷇즫뇪힢,뿉틔ퟜ뷡돶죧뇭2쯹ꆡꆡ(3)쳘헷몯쫽짺돉ꆣ룹뻝뇭3쯹쪾뗄쳘헷쒣냥,뷡뫏쪾뗄뷇즫(R)벯뫏ꆣCRFsힴ첬쳘헷몯쫽(릫쪽(7))틔벰틆쳘헷몯쫽(릫쪽(8))ꆡꆡ뇭2ꆡCRFs_CRL_PnR훐뗄훐컄떥ꆰퟖꆱ뷇즫벯뫏뗄뚨틥,뿉틔듓맛달탲쇐뫍뷇즫뇪힢탲쇐ퟩ훐짺돉돉잧짏췲탲뷇즫뷢쫍쪾샽탲뷇즫뷢쫍쪾샽룶뫲톡쳘헷몯쫽,뿉듓훐톡좡튻뚨쫽솿뗄룟욵쳘헷몯쫽폃폚1B떥탕ퟖ솺맺뇫7F탕잰ힺ킡췵,샏헅죋쏻뷇즫뇪힢ꆣ놾컄톡좡뗄쳘헷몯쫽돶쿖듎쫽(벴쳘헷몯쫽2X뢴탕ퟖ얷퇴훐쪯ꆢ쮾십믕8G탕뫳ힺ샮ퟜ,췵쫏욵듎횵,볇캪f)퓚10듎틔짏ꆣ3C쮫쏻쫗ퟖ쇖뮪쟤ꆢ췵맺뚰9P죋쏻뗄짏컄뛔뿋쇖뛙살쮵ꆡꆡ(4)쳘헷몯쫽좨훘닎쫽통솷ꆣ뇊헟쪹폃쇋CRF++-4D쮫쏻캲ퟖ뾵탂퓆ꆢ퇮탣쎷10Q죋쏻뗄쿂컄헾캯쇵붨탂쮵릤뻟냼[11]살쪵쿖쳘헷몯쫽좨훘닎쫽뗄통솷ꆣCRF++쳡5E떥쏻ퟖ샮?ꆢ얷퇴탞11O솽룶죋쏻횮볤뗄ퟖ쇵늮돐뫍뗋킡욽릩쇋솽룶릤뻟:6W틴틫췢살쏻낢삭램쳘12Z웤쯼ퟖ뒦샭뫃죋볊맘쾵ꆡꆡꋙcrf_learn폃폚듓맛달뇪힢탲쇐ퟩ훐믹폚쳘헷쒣냥짺돉CRFs쳘ꆡꆡ춨맽뛔훐컄ퟖ뷇즫뗄뚨틥,뇊헟붫폯쇏훐돶쿖뗄쯹헷몯쫽,늢룹뻝폃뮧쳡릩뗄횵톡퓱룟욵쳘헷몯쫽,횮뫳춨맽L-폐컄ퟖ뛔쿳BFGS럇쿟탔맦뮮쯣램뛔쳘헷몯쫽좨훘닎쫽뷸탐통솷,붨솢CRFs쒣,냼삨훐펢컄ퟖ럻럖돉12훖뷇즫,CRFs_CRL_PnR뗄맘볼퓚폚통솷폯쇏훐뗄듊탔뇪힢떽뷇즫뇪힢뗄탍;뮯맽돌틔벰뛔통솷폯쇏훐떥ퟖ뛔쿳뇪힢뷇즫뗄ꆡꆡꋚcrf_test쪹폃붨솢뗄CRFs쒣탍뛔쫤죫뗄뷶폐맛달탲쇐ퟩ뗄닢CRFs쫔폯쇏뷸탐뷇즫뇪힢ꆣ톧쾰ꆣꆡ믹폚뷇즫쒣쪽뗄죋쏻쪵쳥쳡좡ꆡ쳘헷몯쫽짺돉벰웤닎쫽통솷ꆡꆡ뺭맽CRF++뗄뷇즫뷢싫뇪힢,뿉틔믱뗃닢쫔폯쇏훐ꆡꆡ(1)맛달탲쇐뗄퇜짺ꆣ틔풭ퟓ떥ퟖ(C)ퟷ캪맛달탲쇐뇭쎿튻룶ꆰퟖꆱ뛔쿳뗄ퟮ병뷇즫ꆣ룹뻝뇭1쯹쇐뗄죋쏻뷇즫,쿖돶살뗄쓜릻뇭헷웤죋쏻뷇즫쳘뗣뗄탅쾢뫜짙,탨튪뾼싇듓뷡뫏훐컄죋쏻뗄튻냣릹돉맦싉,뇊헟죏캪BBꆢBCDꆢBEꆢ풭쪼떥ퟖ맛달탲쇐훐퇜짺돶웤쯼맛달탲쇐틔뇭쿖룼뛠뗄쳘BGꆢBBCDꆢBBEꆢBBGꆢXXCDꆢXXEꆢXXGꆢCDꆢEꆢFBꆢFXXꆢ헷ꆣ뇊헟듓틴틫죋쏻폃ퟖ쳘헷(F)ꆢ탕쫏쳘헷(S)ꆢ잰(P)뫳FBBꆢWꆭW뗈뷇즫ퟩ뫏쒣쪽뿉틔릹돉죋쏻쪵쳥ꆣ춼4쏨(Q)쇙뷧ퟖ쳘헷ꆢ잰(B)뫳(E)ힺ쳘헷틔벰듊캻쳘헷(W)5룶쫶쇋믹폚뷇즫ퟩ뫏쒣쪽쪵쿖죋쏻쪵쳥돩좡뗄쯣램맽돌ꆣ뷇뛈살뾼싇삩햹맛달탲쇐,ퟩ돉맛달탲쇐ퟩꆣꆤ60ꆤ' 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚ꆡ뾪럅탔닢쫔5ꆡCRFs뫍HMM뗄쪵퇩뇈뷏ꆡꆡ놾듎쪵퇩닉폃뗄통솷뫍닢쫔폯쇏살ퟔꆶ죋쏱죕놨ꆷꆡꆡCRFs뫍HMM쯤좻뚼쫇룹뻝맛달탲쇐쳘헷뷸탐탲쇐1998쓪1퓂럝뺭맽뇪힢뗄훐컄폯쇏뿢,틔1퓂럝잰24쳬릲뇪힢,떫틀뻝뗄쫽톧풭샭ꆢ붨쒣뗄맽돌틔벰닺짺뗄뷡맻뚼15767뻤폯쇏ퟷ캪통솷벯,뫳7쳬릲3717뻤뗄폯쇏ퟷ캪닢늻쿠춬ꆣ믹폚솽훖쒣탍틀삵뗄쫽톧믹뒡,뇊헟죏캪CRFs쫔벯(릲폐훐컄죋쏻3132룶)ꆣ쪹폃HMM_WRL_PnR뫍뇜쏢쇋HMM평폚샭싛좱쿝쯹뇘좻듦퓚뗄컊쳢,뷏HMMCRFs_CRL_PnR럖뇰뷸탐훐컄죋쏻돩좡,뷡맻죧춼5쯹쪾ꆣ룼쫊뫏뷢뻶탲쇐뇪힢컊쳢;떫HMM뷏횮CRFs퓲쫽톧샭ꆡꆡ쪵퇩뇭쏷,CRFs쒣탍퓚쪶뇰ힼ좷싊짏쏷쿔폅폚싛볲떥,벯돉쳘헷짙,틲뛸통솷쪱볤뛌ꆣ캪쇋퇩횤샭싛첽HMM;솽헟헙믘싊쿠닮늻뛠,뺡맜CRFs뇈HMM뷶룟돶2쳖믱뗃뗄뷡싛,뇊헟뛔솽헟퓚죋쏻쪶뇰훐뗄펦폃(HMM_룶뛠냙럖뗣,뷡맻쪹뗃퓚F횵뇈뷏훐,CRFs뇈HMM룟돶WRL_PnR뫍CRFs_CRL_PnR)뷸탐쇋뾪럅탔닢쫔뫍탂컅쇋뷼10룶냙럖뗣;HMM퓚Nreconum튻쿮짏뇈CRFs룟,쮵죋컯돩좡펦폃솽룶쪵퇩,틔첽쳖웤닮틬ꆣ뇊헟쓢춳볆틔쏷잰헟뇈뫳헟룼짆폚랢쿖죋쏻쪵쳥,떫헢튲퇏훘컾짼쇋쿂쫽뻝ퟷ캪닢움횸뇪:룃쒣탍뗄쪶뇰ힼ좷싊;듋췢,쪵퇩뛔쒣탍뗄통솷쪱볤튲뷸ꆡꆡ놻쪶뇰뗄죋쏻룶쫽(Nreconum)탐쇋뇈뷏,랢쿖HMM훐3룶듊뗤뗄짺돉뷶탨풼30럖훓,ꆡꆡힼ좷싊(P)=Ntruenum/Nreconumꇁ100%뛸CRFs퓲탨튪뮨럑23룶킡쪱뷸탐닎쫽통솷,뷸튻늽횤ꆡꆡ놻헽좷쪶뇰뗄죋쏻룶쫽(Ntruenum)쏷CRFs뗄쒣탍뢴퓓뛈풶룟폚HMMꆣꆡꆡ헙믘싊(R)=Ntruenum/Ntotlenumꇁ100%ꆡ탂컅죋컯돩좡펦폃ꆡꆡ통솷쪱볤(Tltime)ꆡꆡꆡꆡ쪹폃췸튳쿂퓘릤뻟뛔쓏뺩듳톧킡냙뫏BBS뷸탐탅쾢F-횵=(ꛂ2+1)ꇁPꇁR/(ꛂ2ꇁP+R)=2ꇁPꇁR/(P+R)(P좨훘ꛂ좡1,벴죏캪닉벯,뇊헟믱뗃웤훐샺쪷(History)냦ퟔ뒴냦틔살떽2007P뫍R춬뗈훘튪)쓪9퓂뗄쯹폐훷쳢뇪쳢,좻뫳쪹폃뺭맽통솷(통솷쫽뻝뻹ꆤ61ꆤ' 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚살ퟔꆶ죋쏱죕놨ꆷ1998쓪1퓂럝뺭맽뇪힢뗄훐컄폯쇏뿢)뷸탐ퟔ뚯죋쏻돩좡,믱뗃헢뛎쪱볤쓚History냦쯹짦벰뗄뗄HMM_WRL_PnR뫍CRFs_CRL_PnR럖뇰뛔훷쳢뇪쳢붹뗣죋컯ꆣ춼6ꆡHistory냦훐쳖싛듎쫽듯떽10듎뗄붹뗣죋컯ꆡꆡ죧춼6쯹쪾,HMM_WRL_PnR헽좷쪶뇰죋쏻훐쳖싛죋쏻쪵쳥돩좡펦폃뗄ힼ좷싊ꆢ헙믘싊ꆢF-횵틔벰헽좷쪶듎쫽듯떽10듎뗄릲폐27죋,릲볆665죋듎;뛸CRFs_뇰쫽뗈룷쿮닢움횸뇪짏뚼풶풶실뫳폚CRFsꆣ놾컄퓚쫽CRL_PnR헽좷쪶뇰죋쏻훐쳖싛듎쫽듯떽10듎뗄폐24톧샭싛짏뫍쪵볹펦폃훐뚼횤쏷쇋CRFs평폚뫏샭뗄쫽톧죋,릲볆663죋듎;CRFs퓚헢듎뇈뷏훐짔뒦쿂럧ꆣ룹뻝샭싛ꆢ퇏쏜뗄쫽톧췆샭쪹웤뇈HMM룼쫊뫏폃폚탲쇐뇪힢듋듎펦폃,뇊헟죏캪:믲뛔쿳럖샠뗈탅쾢돩좡펦폃훐ꆣ듋췢,놾컄쳡돶쇋믹폚ꆡꆡ(1)솽헟쪶뇰뗄죋쏻뻹퓚3ퟖ틔쿂,뛸퓚뾪랢탔닢쫔훐짦뷇즫뇪힢뗄죋쏻쪵쳥돩좡랽램,죏캪컞싛믹폚HMM뮹쫇벰튻뚨쫽솿뗄맺췢틴틫죋쏻(뎤뛈퓚3ퟖ틔짏),뺭맽뇈뷏뿉믹폚CRFs통솷,웤퓚죋쏻쪵쳥돩좡훐뗄펦폃킧맻뚼틑뺭횪,HMM뛔뛠폚4ퟖ틔짏죋쏻뗄쪶뇰쓜솦뷏죵,CRFs_CRL_듯떽쇋쪵폃뗄쮮욽(뾪럅탔쪵퇩F-횵뎬맽쇋80%)ꆣPnR훐틴틫죋쏻폃ퟖ쳘헷(F)뗄볓죫쪹뗃웤쪶뇰췢맺죋쏻뗄ꆡꆡ뇊헟춨맽놾컄뗄퇐뺿,쳡돶룅싊춳볆쒣탍뿉틔맣랺쓜솦폐쏷쿔뗄쳡룟;펦폃폚탅쾢돩좡뗄룷룶닣듎훐,냼삨쏼쏻쪵쳥뗄돩좡ꆢ럇ꆡꆡ(2)놾듎펦폃뷶춳볆쳖싛듎쫽뷏뛠뗄죋쏻,뛠벯훐폚벸뷡릹뮯컄놾뗄훷쳢듊돩좡,짵훁폚쪵쳥횮볤맘쾵뗄에뇰ꆢ룶훸쏻붹뗣죋컯,쪶뇰뗄죋쏻쫽짙폚30죋,퓚헢훖쟩뿶쿂,듺듊횸듺뛔쿳뗄에뇰뗈뗈,펦폃뗄맘볼퓚폚뇪힢벯뫏뗄CRFs_CRL_PnR폅쫆늢늻쏷쿔;ꆡꆡ(3)HMM뚨틥틔벰맛달탲쇐뗄짺돉ꆣ틲듋,뷱뫳뗄쳖싛훘뗣뿉틔_WRL_PnR쪶뇰쇋늿럖CRFs_CRL_PnR쎻쓜쪶뇰뗄죋쏻,뇭쏷HMM_WRL_PnR퓚죋쏻헙믘랽쏦뻟폐튻뚨폅럅퓚뷸튻늽쳡룟CRFs_CRL_PnR뗄죋쏻쪶뇰헙믘싊ꆢ믹쫆폚춳볆뗄쪵쳥맘쾵쪶뇰ꆢ듺듊횸듺뛔쿳쪶뇰뗈믺웷톧쾰,뾪럅탔닢쫔훐HMM_WRL_PnR뗄Nreconum횵뷏룟(쯤좻R횵욫뗍)튲쮵쏷쇋헢룶컊쳢퇐뺿훐,틔ퟷ캪췸싧폟쟩럖컶쾵춳ꆢ뺺헹쟩놨쾵춳퇐뺿ꆢ;ꆡꆡ놾쳥룅쓮벰맘쾵ퟔ뚯돩좡ꆢ컄놾폯틥럖컶ꆢ놾쳥ퟔ뚯뮯릹(4)CRFs_CRL_PnR뗍헙믘싊뗄튻룶뫜훘튪뗄풭틲퓚폚뛏뻤늻돤럖,퓚HMM_WRL_PnR훐붨뗈퇐뺿뗄믹뒡ꆣ,폶떽낢삭늮쫽ퟖ믲펢컄ퟖ럻벴뛏뻤,뛸퓚CRFs_CRL_PnR훐,틔통솷폯쇏훐튻탐ퟷ캪닎뾼컄쿗:튻룶뻤ퟓ뷸탐통솷,뛔폚닢쫔컄놾튲쫇틔튻룶훷쳢뇪쳢ퟷ캪ꆡ[1]뢵낮욽.볆쯣폯퇔톧뫍ퟔ좻폯퇔탅쾢뒦샭퇐뺿뫍펦폃ퟛ쫶튻룶뻤ퟓ,쎻폐뫏샭뛏뻤,떼훂쪹폃CRFs_CRL_PnR쪶뇰뗄죋[EB/OL].[2007-10-01]. 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.
ꆶ쿖듺춼쫩쟩놨벼쫵ꆷꆡ2007쓪ꆡ뗚12웚쟩놨럖컶폫퇐뺿ꆡꆡퟜ뗚158웚[2007-10-01]. 1994-2014 China Academic Journal Electronic Publishing House. All rights reserved.