B53C B6 ( ) Vol.53 No D11 JournalofXiamenUniversity (NaturalScience) Nov.2014 doi: /j.issn P > J M (^ < C,^ ^

Size: px

Start display at page:

Download "B53C B6 ( ) Vol.53 No D11 JournalofXiamenUniversity (NaturalScience) Nov.2014 doi: /j.issn P > J M (^ < C,^ ^"

秒蹒诸
5 years ago
Views:

1 B53C B6 ( ) Vol.53 No.6 04D JournalofXiamenUniversity (NaturalScience) Nov.04 doi:0.6043/j.issn P > J M (^ < C,^ ^ ) GH:: ) ZM 7 XY ) V,3 c : "Z ) 7: ", W c : "E U / ). 3E )W, U) 9, IFW 9 )< ;-,E << ), I <!) <. J N DR, < ) U3.4%~5.0%)O. =IJ:W ; c ; J KL MN: TP39 &7O : A & N: (04) STUV <)=>?@A, 0 90D,PeterBrownI [] IF U ): 7 ", ). M < I 6, : 7 C ) ` D, [ (JHU) ) : 7 < [ 3] KFIBM MD) *IF U ) : 7 Egypt [4]. < IBM : 7 ")O I, F : 7!.M : ) 7 NF $. : G ), U " 7 C.bR 犲 P FU 犳,( U 犳 ;< FbR 犲, )b )!Z 犳犲. a,kw ) Pr( 犲狘犳 )= Pr ( 犳狘犲 ) Pr( 犲 ), Pr( 犳 ) )br! Z ) U 犳 < T X :03 6 PQR: 948 Fb( ) 犿犪犻犾 :stdin@xmu.edu.cn )br ^ 犲, N/ 7/Pr( 犳 ) br 犲 (, c X %/ 7/ Pr( 犳犲 )Pr( 犲 ): ^ 犲 =argmax Pr( 犳狘犲 ) Pr( 犲 ), 犲 7 Brown I N : 7 ) [].,Pr( 犲 )ZbR ) I+), N ";Pr( 犳犲 )Z br 犲 U 犳 ), N ". " b R (, U (, + ) Z b R I )?,S F E ) a?; " U b R U (, + F 7 )K,S F?.,: 7 KW/ 37 : ) ")E. 狀 " C E. ) ")E, ) 狆 ( 犳犼犲犪犼 )I. 3)_,U GH < X ) E. Z 7 ") 6F (non deterministic polynomial,np),k W * _ 5 6, K Q) $< ) E. B C Pr(.)N U )=, 狆 (.)N " () ; 犳犑?N 犑 )U 犳犑 )-., 犲?N )br 犲犑 )-.. ) 犪, W U B 犼 7 br B 犪犼 7 ) c(. c "K

2 B6 5 : J)W < 805 犑犑犑 Pr( 犳狘犲 )= Pr( 犳, 犪狘犲 ). 犑犪 IBM c " U ) E,M )O : "Z ), Marcu) ` ) : " [5],Keohn ) pharaoh : 7 : [6]. " E U I+ ),( L U ) ;, 4 3 N C.; U Monday 7, ) Z 4 Monday F Sun day, "! N.NF 7, ) c ", 7, 3 T), Monday Sunday 7, "! I 7. L )!Z c [7] ): 7 ". ; 7 ( 犳犑, 犲 ),4 / 8 犽 7, 7 P 犅, 4 /8 犽 7 -.: 犑犑 Pr( 犳狘犲 )= Pr( 犅狘犲 ) Pr( 犳狘犅, 犲 )= 犅 α( 犲 ) 犓犓 Pr( 珟犳狘珓犲 ). 犅 U K / 8 U ) α( 犲 ),!4 c/n 77/, 9 9 ) c. 9 ) c,e N c, < ). c Z 7G ( 犉犑,, 珦犃 ), 3 珦犃 ZU -. 犉犑 br -. ) c(. 珦犃 Z 7 犑 ( +)), *) N, T ) 7 Z c), Z W E4AE ), ZNF c ) %. J M J M>PbDE 4 犞 7 / 犆 7,π Z 7`, 4 狑犻 3 犮犻.! 狀 - ) 狀 ":Pr( 狑狀狑 )=Pr( 狑狀犮狀 )Pr( 犮狀犮狀 - ). br`n: 犔 (π)= 犜犜 log( 狆 ( 狑犻狘犮犻 ) 狆 ( 犮犻狘犮犻 -))= 狑犻 = ( 狑犻 )log 狆 ( 狑犻 )+ 狆 ( 犮犻 - 犮犻 ) 犻狆犮犻犮犻 - 狆 ( 犮犻 - 犮犻 ) log 狆 ( 犮犻 -) 狆 ( 犮犻 ) =- 犎 ( 狑 )+ ( 犮犮 )., 犎 ( 狑 )Z ) J, 7, < π /, 7 / Z P ); ( 犮, 犮 )Z )>? [8]. ( ) U 6, WangI [9] IF <GH)agglomerativeinformationbotle neck(aib)< C ; ), U Wang I [0] a 狀 ") IF <9. M W ) b R Z < ) c (,4 < ) 4 /. 狀 ",W )br`n [] : 犜犔 (π,π 犉 )= 犜 + [ log[ 狆 ( 犲犻狘犻 ) 犜犉犻 = 犜犉狆 ( 犻狘犻 -)]+ log[ 狆 ( 犳犼狘犉犼 ) 犼 = 狆 ( 犉犼狘犪 )]], 犼 3 π π 犉 < ) /, 犜犜犉 W c < )7. D7 6 7 c 犪犼,D 7 犳犼 3 c) 犲犪犻 9. 犜犔 (π,π 犉 )= 犜 + [ log[ 狆 ( 犲犻狘犻 ) 犜犉犻 = 犜犉狆 ( 犻狘犻 -)]+ log[ 狆 ( 犳犼狘犉犼 ) 犼 = 狆 ( 犉犼狘犪 )]]= 犼犜 + 犜犉 [ 犖 ( )log 犖 ( )- 犖 ( )log 犖 ( )+ 犖 ( 犲 )log 犖 ( 犲 ) 犲犖 ( 犪, 犉 )log 犖 ( 犪, 犉 )- 犪犉犖 ( 犪 )log 犖 ( 犪 )- 犖 ( 犉 )log 犖 ( 犉 )+ 犪犉犖 ( 犳 )log 犖 ( 犳 ) ]. 犳犎 ( 犲 ) 犎 ( 犳 )N < ) J, (, )N )>?, ( 犪, 犉 )N < )>?.N X W /, X %br`: 犌 (π,π 犉 )= 犖 ( )log 犖 ( )- 犖 ( )log 犖 ( )+ ( 犪, 犉 ) 犪, 犉犖 log 犖 ( 犪, 犉 )- 犪犖 ( 犪 )log 犖 ( 犪 )- 犖 ( 犉 )log 犖 ( 犉 ). 犉

3 K(4O ) 04D 3 M 9 < ZQ D S, < ) c,/ < <I XJ) 7 C,X < Q ) 7. 犉 / <, ) / N 犆犆犉, )7/ N 犞犞犉.W 9 hierarchy bicluster)< W : : < ) 7 犆, 犆犉 :W /π,π 犉 ) < )D7 ) - -,/ 犛, 犛犉 7 $. ) 犛 I X 犆 7,D7 c/n 7,D7 / ;N,,, 狀, 犛,. 3) 犛犉 I X 犆犉 7,D7 c/n 7,D7 / ;N 犉, 犉,, 犉犿, 犛犉,. 4) 犛 > 0, 3 I X ) 7 犲 /N 7 狓,,,, 狀 <I 7 犻, 狓犻,bR` X, 狓犻, R;N 犻, 犛, 7. ^ 犻 =argmax 犔 (π,π 犉 ), 犻 {,,, 犆 }, 犻犻 = 犻狓 }. 5) 犛犉 U *, 3 I X ) 7 犳 /N 7 犉狓, 犉, 犉,, 犉犿 <I 7 犉犻, 犉狓犉犻,b R ` X, 犉狓犉犻, R;N 犉犻, 犛犉, 7. ^ 犻 =argmax 犔 (π,π 犉 ), 犻 {,,, 犆犉 }, 犻犉犻 = 犉犻犉狓 }. 6) KBC4) 5), 7 U) / 4 ). 4 M 9 Z = 4 D 7 % / ; 7, D 7 < I 7, 7 ) A 7,bR ` + X.W 9, D 4 < )D7 A br` 犌 (π,π 犉 ) + X ) []. W 9 flat bicluster)< W : : < ) 7 犆, 犆犉 :W /π,π 犉 )4 < )D7 / ) % 犆, 犆犉 7, < ) / R ;N,,, 狀犉, 犉,, 犉犿. ) )D7 犲,4 犲 A br` ^ 犲 =argmax 犔 (π,π 犉 ), 犲, 犲犻, 犻 = 犻 - 犲, 犲犼 = 犼犲,3 π N4 犲犻 A 犼, ) /. 3) 犉 )D7 犳,4 犳 A br` ^ 犳 =argmax 犔 (π,π 犉 ), 犳犉, 犳犉犻, 犉犻 = 犉犻 - 犳, 犳犉犼 = 犉犼犲,3 π 犉 N 4 犳犉犻 A 犉犼, 犉 ) /. 4) K) 3), U 7 A. 5 M > / <W ZW 犌 (π,π 犉 ) NbR `, ),a X ) /. Z 9,4 ) D7 % 7,X 7 7X. c, ) /, 9 ZU )., E I <! ),3 Z a 9 < % 9 ).K W 9, ),3< W : : < ) 7 犆, 犆犉 :W /π,π 犉 ) < )D7 ) - -,/ 犛, 犛犉 7 $. ) 犛 I X 犆 7,D7 c/n 7,D7 / ;N,,, 狀, 犛,. 3) 犛犉 I X 犆犉 7,D7 c/n 7,D7 / ;N 犉, 犉,, 犉犿, 犛犉,. 4) 犛 > 0, 3 I X ) 7 犲 /N 7 狓,,,, 狀 <I 7 犻, 狓犻,bR` X, 狓犻, R;N 犻, 犛, 7. ^ 犻 =argmax 犔 (π,π 犉 ), 犻 {,,, 犆 }, 犻

4 B6 5 : J)W < 807 犻 = 犻狓 }. 5) 犛犉 U *, 3 I X ) 7 犳 /N 7 犉狓, 犉, 犉,, 犉犿 <I 7 犉犻, 犉狓犉犻,b R ` X, 犉狓犉犻, R;N 犉犻, 犛犉, 7. ^ 犻 =argmax 犔 (π,π 犉 ), 犻 {,,, 犆犉 }, 犻犉犻 = 犉犻犉狓 }. 6) KBC4) 5), 7 U) / 4 ). 7) )D7 犲,4 犲 A br` ^ 犲 =argmax 犔 (π,π 犉 ), 犲, 犲犻, 犻 = 犻 - 犲, 犲犼 = 犼犲,3 π N4 犲犻 A 犼, ) /. 8) 犉 )D7 犳,4 犳 A br` ^ 犳 =arg max 犔 (π,π 犉 ), 犳犉, 犳犉犻, 犉犻 = 犉犻 - 犳犳, 犉犼 = 犉犼犲,3 π 犉 N 4 犳犉犻 A 犉犼, 犉 ) /. 9) K7) 8), U 7 A. 3 +F3<W <,a W E C.W E N M,D F 0737,3 E N 45838, E N < ) 6 c(, c a Egypt: 7 ) GIZA = W E C c E, ) c, C W.GIZA )7/ c ^. ^ GIZA E ) c Fig. AlignmentresultgeneratedbyGIZA W )?:perp 犫犻 = - 犔 (π,π ) 犉, W C D. 犔 (π,π 犉 )= 犜犜 + 犜 [ log 犜犉犉犻 = [ 狆 ( 犲犻狘犻 ) 狆 ( 犻狘犻 -)]+ log[ 狆 ( 犳犼狘犉犼 ) 狆 ( )] 犉犼狘犪犼 = 犼 ] ZW )br`.perp 犫犻 ), M ) π 6) ( 犉 ) π 犉,W ). NF 6 ), F ), CW ),3 E N4896, E N503; 3< / C F 5, ) 7 N 00 ~400. K, <,9 ) 9 ). < X ) < c< ). 9 W )7/ (/N007 ) Tab. Partialresultgeneratedbyhierarchicalalgorithm E E : : : 3 : 4 : : : 3 : 4 everyone eve rybody every thingeachan yone anybody somebody ev eryeverywhere someone eve ryday welgood bad beter best great greatest worse cool rough cold cute nice low fine strong hotwarm don t didn t wouldn t shouldn t doesn t won t donttneedn t doesen tdon t too very quite rathersopret ty Too highly absolutely greatly sorely tremely ex F X ) O = H,= = K I J c Z K Z = O I J O O O

5 K(4O ) 04D 9 )7/ (/ 007 ) Tab. Partialresultgeneratedbyflatalgorithm E E : : : 3 : 4 : : : 3 : 4 know knows trust believe understood hear under stand remem ber realize known heard knewbelieved welgood bad biggest beter best great Great nice worst badly worse finest eficient fool ish infromafterat don t didn t before on be never wouldn t hind by into shouldn tcan t under upon haven tdoesn t near besidea couldn t won t mong across cannot hardly during be mustn thasn t tweenbelow hadn t M F ; ; + Z c ) / M I 3 N! " #, $ % & ' 3 3<< )? Tab.3 Comparisonofperplexityinthreealgorithms 7? < (5.0%) (4.8%) (4.53%) (3.48%) : <. )$ 9, < W? J)^/. 9 4,K W G I U 7 b(, c 9 +IF ).9 9 F 6), 9 6, Tab.4 Flatalgorithmtimeconsuming 7 /s /s % KWGI, << ) U ), c ) c "U ). E I) <, ` / U, J F ` / ), KK ) Tab.5 Timeconsumingcomparison betweenhierarchicalandflatalgorithms h < :00,00,300, , X ), E I ) < F X ), 9 <, F 3.4%~5.0%. Z Q I ) Z,9 < ) < 9 <, Z 9 9 F 6), &7: [] `.: 7 [J]. E K,003,7 (4):. [] BrownPF,CockeJ,PietraSA D,etal.Astatisticalap proachto machinetranslation [J].ComputationalLin guistics,990,6(): [3] KnightK.Astatisticalmttutorialworkbook [R].Mary land:jhusummerworkshop,999. [4] CenterforLanguageand Speech ProcessingatJohns Hopkins University.The EGYPT statistical machine translationtoolkit[eb/ol].[999].[ ].ht tp: old site.clsp.jhu.edu/ws99/projects/mt/toolkit/. [5] Marcu D,Wong W.A phrase based,joint probability modelforstatisticalmachinetranslation [C] Proceed ingsoftheconferenceonempiricalmethodsinnatural LanguageProcessing.Philadelphia:NSF,00: [6] Koehn P.Pharaoh:a beam search decoderforphrase

6 B6 5 : J)W < 809 basedstatisticalmachinetranslation modelsusermanual anddescription [J].MachineTranslation:From RealUs ersto Research Lecture Notesin Computer Science, 004,365:5 4. [7] OchFJ.Statisticalmachinetranslation:fromsingle word modelstoalignmenttemplates[d].deutschland:techni caluniversityaachen,00. [8] BrownPF,PietraVJD.Class basedn gram modelsof naturallanguage[j].computationallinguistics,99,8 (4): [9] WangLei,ZhangJianjia,ZhouLuping,etal.Afastap proximateaibalgorithmfordistributionalwordclustering [C] Proceedingsofthe03IEEEConferenceonCom puter Vision and Patern Recognition.[S.l.]:IEEE, 03: [0] Wang W,VergyriD.Theuseofwordn gramsandparts ofspeechforhierarchicalclusterlanguagemodeling[c] ProceedingsoftheIEEEInternationalConferenceon Acoustic,SpeechandSignalProcessing.[S.l.]:IEEE, 006: [] OchFJ.Aneficientmethodfordeterminingbilingual wordclasses[c] ProceedingEACL 99Proceedingsof theninthconferenceoneuropeanchapteroftheasso ciationforcomputationallinguistics.new York:ACM Press,999:7 76. 犅犻犾犻狀犵狌犪犾犠狅狉犱犆犾狌狊狋犲狉犻狀犵犃犾犵狅狉犻狋犺犿犅犪狊犲犱狅狀狀犳狅狉犿犪狋犻狅狀狀狋狉狅狆狔 ZHANGZhen chang (ColegeofComputerandInformationScience,FujianAgricultureandForestryUniversity,Fuzhou35000,China) 犃犫狊狋狉犪犮狋 : Statisticsmethodconstitutetheprimaryresearchdirectionintheareaofmachinetranslation.Thealignmenttemplateis oneofthebeststatisticsmodel,andbilingualwordclusteringplaysakeyroleintheparametertrainingofalignmenttemplate.inthis paper,weproposeahybridclusteralgorithm basedonhierarchicalandflatalgorithm.thenewhybridalgorithmexhibits3.4% ~ 5.0% performanceimprovementthanflatalgorithmaccordingtotheinformationentropy. 犓犲狔狑狅狉犱狊 : bilingualwordclustering ;alignmenttemplate;informationentropy

( ) Vol.54 No JournalofXiamenUniversity (NaturalScience) Mar.2015 doi: /j.issn R, ( ]a G,9O ) 56

( ) Vol.54 No JournalofXiamenUniversity (NaturalScience) Mar.2015 doi: /j.issn R, ( ]a G,9O ) 56 154 12 ( ) Vol.54 No.2 2015 3 JournalofXiamenUniversity (NaturalScience) Mar.2015 doi:10.6043/j.issn.0438 0479.2015.02.018 + 9 R, ( ]a G,9O 361005) 56: % 5 ) B` M 7 #