%!PS-Adobe-2.0 %%Creator: dvipsk 5.58f Copyright 1986, 1994 Radical Eye Software %%Title: iccs99.dvi %%Pages: 8 %%PageOrder: Ascend %%BoundingBox: 0 0 612 792 %%EndComments %DVIPSCommandLine: dvips -o iccs99.ps iccs99 %DVIPSParameters: dpi=300, compressed, comments removed %DVIPSSource: TeX output 1999.08.21:1908 %%BeginProcSet: texc.pro /TeXDict 250 dict def TeXDict begin /N{def}def /B{bind def}N /S{exch}N /X{S N}B /TR{translate}N /isls false N /vsize 11 72 mul N /hsize 8.5 72 mul N /landplus90{false}def /@rigin{isls{[0 landplus90{1 -1}{-1 1} ifelse 0 0 0]concat}if 72 Resolution div 72 VResolution div neg scale isls{landplus90{VResolution 72 div vsize mul 0 exch}{Resolution -72 div hsize mul 0}ifelse TR}if Resolution VResolution vsize -72 div 1 add mul TR matrix currentmatrix dup dup 4 get round 4 exch put dup dup 5 get round 5 exch put setmatrix}N /@landscape{/isls true N}B /@manualfeed{ statusdict /manualfeed true put}B /@copies{/#copies X}B /FMat[1 0 0 -1 0 0]N /FBB[0 0 0 0]N /nn 0 N /IE 0 N /ctr 0 N /df-tail{/nn 8 dict N nn begin /FontType 3 N /FontMatrix fntrx N /FontBBox FBB N string /base X array /BitMaps X /BuildChar{CharBuilder}N /Encoding IE N end dup{/foo setfont}2 array copy cvx N load 0 nn put /ctr 0 N[}B /df{/sf 1 N /fntrx FMat N df-tail}B /dfs{div /sf X /fntrx[sf 0 0 sf neg 0 0]N df-tail}B /E{ pop nn dup definefont setfont}B /ch-width{ch-data dup length 5 sub get} B /ch-height{ch-data dup length 4 sub get}B /ch-xoff{128 ch-data dup length 3 sub get sub}B /ch-yoff{ch-data dup length 2 sub get 127 sub}B /ch-dx{ch-data dup length 1 sub get}B /ch-image{ch-data dup type /stringtype ne{ctr get /ctr ctr 1 add N}if}B /id 0 N /rw 0 N /rc 0 N /gp 0 N /cp 0 N /G 0 N /sf 0 N /CharBuilder{save 3 1 roll S dup /base get 2 index get S /BitMaps get S get /ch-data X pop /ctr 0 N ch-dx 0 ch-xoff ch-yoff ch-height sub ch-xoff ch-width add ch-yoff setcachedevice ch-width ch-height true[1 0 0 -1 -.1 ch-xoff sub ch-yoff .1 add]/id ch-image N /rw ch-width 7 add 8 idiv string N /rc 0 N /gp 0 N /cp 0 N{ rc 0 ne{rc 1 sub /rc X rw}{G}ifelse}imagemask restore}B /G{{id gp get /gp gp 1 add N dup 18 mod S 18 idiv pl S get exec}loop}B /adv{cp add /cp X}B /chg{rw cp id gp 4 index getinterval putinterval dup gp add /gp X adv}B /nd{/cp 0 N rw exit}B /lsh{rw cp 2 copy get dup 0 eq{pop 1}{dup 255 eq{pop 254}{dup dup add 255 and S 1 and or}ifelse}ifelse put 1 adv} B /rsh{rw cp 2 copy get dup 0 eq{pop 128}{dup 255 eq{pop 127}{dup 2 idiv S 128 and or}ifelse}ifelse put 1 adv}B /clr{rw cp 2 index string putinterval adv}B /set{rw cp fillstr 0 4 index getinterval putinterval adv}B /fillstr 18 string 0 1 17{2 copy 255 put pop}for N /pl[{adv 1 chg} {adv 1 chg nd}{1 add chg}{1 add chg nd}{adv lsh}{adv lsh nd}{adv rsh}{ adv rsh nd}{1 add adv}{/rc X nd}{1 add set}{1 add clr}{adv 2 chg}{adv 2 chg nd}{pop nd}]dup{bind pop}forall N /D{/cc X dup type /stringtype ne{] }if nn /base get cc ctr put nn /BitMaps get S ctr S sf 1 ne{dup dup length 1 sub dup 2 index S get sf div put}if put /ctr ctr 1 add N}B /I{ cc 1 add D}B /bop{userdict /bop-hook known{bop-hook}if /SI save N @rigin 0 0 moveto /V matrix currentmatrix dup 1 get dup mul exch 0 get dup mul add .99 lt{/QV}{/RV}ifelse load def pop pop}N /eop{SI restore showpage userdict /eop-hook known{eop-hook}if}N /@start{userdict /start-hook known{start-hook}if pop /VResolution X /Resolution X 1000 div /DVImag X /IE 256 array N 0 1 255{IE S 1 string dup 0 3 index put cvn put}for 65781.76 div /vsize X 65781.76 div /hsize X}N /p{show}N /RMat[1 0 0 -1 0 0]N /BDot 260 string N /rulex 0 N /ruley 0 N /v{/ruley X /rulex X V}B /V {}B /RV statusdict begin /product where{pop product dup length 7 ge{0 7 getinterval dup(Display)eq exch 0 4 getinterval(NeXT)eq or}{pop false} ifelse}{false}ifelse end{{gsave TR -.1 -.1 TR 1 1 scale rulex ruley false RMat{BDot}imagemask grestore}}{{gsave TR -.1 -.1 TR rulex ruley scale 1 1 false RMat{BDot}imagemask grestore}}ifelse B /QV{gsave transform round exch round exch itransform moveto rulex 0 rlineto 0 ruley neg rlineto rulex neg 0 rlineto fill grestore}B /a{moveto}B /delta 0 N /tail{dup /delta X 0 rmoveto}B /M{S p delta add tail}B /b{S p tail} B /c{-4 M}B /d{-3 M}B /e{-2 M}B /f{-1 M}B /g{0 M}B /h{1 M}B /i{2 M}B /j{ 3 M}B /k{4 M}B /w{0 rmoveto}B /l{p -4 w}B /m{p -3 w}B /n{p -2 w}B /o{p -1 w}B /q{p 1 w}B /r{p 2 w}B /s{p 3 w}B /t{p 4 w}B /x{0 S rmoveto}B /y{ 3 2 roll p a}B /bos{/SS save N}B /eos{SS restore}B end %%EndProcSet %%BeginProcSet: special.pro TeXDict begin /SDict 200 dict N SDict begin /@SpecialDefaults{/hs 612 N /vs 792 N /ho 0 N /vo 0 N /hsc 1 N /vsc 1 N /ang 0 N /CLIP 0 N /rwiSeen false N /rhiSeen false N /letter{}N /note{}N /a4{}N /legal{}N}B /@scaleunit 100 N /@hscale{@scaleunit div /hsc X}B /@vscale{@scaleunit div /vsc X}B /@hsize{/hs X /CLIP 1 N}B /@vsize{/vs X /CLIP 1 N}B /@clip{ /CLIP 2 N}B /@hoffset{/ho X}B /@voffset{/vo X}B /@angle{/ang X}B /@rwi{ 10 div /rwi X /rwiSeen true N}B /@rhi{10 div /rhi X /rhiSeen true N}B /@llx{/llx X}B /@lly{/lly X}B /@urx{/urx X}B /@ury{/ury X}B /magscale true def end /@MacSetUp{userdict /md known{userdict /md get type /dicttype eq{userdict begin md length 10 add md maxlength ge{/md md dup length 20 add dict copy def}if end md begin /letter{}N /note{}N /legal{} N /od{txpose 1 0 mtx defaultmatrix dtransform S atan/pa X newpath clippath mark{transform{itransform moveto}}{transform{itransform lineto} }{6 -2 roll transform 6 -2 roll transform 6 -2 roll transform{ itransform 6 2 roll itransform 6 2 roll itransform 6 2 roll curveto}}{{ closepath}}pathforall newpath counttomark array astore /gc xdf pop ct 39 0 put 10 fz 0 fs 2 F/|______Courier fnt invertflag{PaintBlack}if}N /txpose{pxs pys scale ppr aload pop por{noflips{pop S neg S TR pop 1 -1 scale}if xflip yflip and{pop S neg S TR 180 rotate 1 -1 scale ppr 3 get ppr 1 get neg sub neg ppr 2 get ppr 0 get neg sub neg TR}if xflip yflip not and{pop S neg S TR pop 180 rotate ppr 3 get ppr 1 get neg sub neg 0 TR}if yflip xflip not and{ppr 1 get neg ppr 0 get neg TR}if}{noflips{TR pop pop 270 rotate 1 -1 scale}if xflip yflip and{TR pop pop 90 rotate 1 -1 scale ppr 3 get ppr 1 get neg sub neg ppr 2 get ppr 0 get neg sub neg TR}if xflip yflip not and{TR pop pop 90 rotate ppr 3 get ppr 1 get neg sub neg 0 TR}if yflip xflip not and{TR pop pop 270 rotate ppr 2 get ppr 0 get neg sub neg 0 S TR}if}ifelse scaleby96{ppr aload pop 4 -1 roll add 2 div 3 1 roll add 2 div 2 copy TR .96 dup scale neg S neg S TR}if}N /cp {pop pop showpage pm restore}N end}if}if}N /normalscale{Resolution 72 div VResolution 72 div neg scale magscale{DVImag dup scale}if 0 setgray} N /psfts{S 65781.76 div N}N /startTexFig{/psf$SavedState save N userdict maxlength dict begin /magscale true def normalscale currentpoint TR /psf$ury psfts /psf$urx psfts /psf$lly psfts /psf$llx psfts /psf$y psfts /psf$x psfts currentpoint /psf$cy X /psf$cx X /psf$sx psf$x psf$urx psf$llx sub div N /psf$sy psf$y psf$ury psf$lly sub div N psf$sx psf$sy scale psf$cx psf$sx div psf$llx sub psf$cy psf$sy div psf$ury sub TR /setpagedevice{pop}N statusdict begin /letter{}N /lettertray{}N /legal{} N /legaltray{}N /a4{}N /a4tray{}N /b4{}N /b4tray{}N /b5{}N /b5tray{}N /showpage{}N /erasepage{}N /copypage{}N end /letter{}N /lettertray{}N /legal{}N /legaltray{}N /a4{}N /a4tray{}N /b4{}N /b4tray{}N /b5{}N /b5tray{}N /showpage{}N /erasepage{}N /copypage{}N /p 3 def @MacSetUp}N /doclip{psf$llx psf$lly psf$urx psf$ury currentpoint 6 2 roll newpath 4 copy 4 2 roll moveto 6 -1 roll S lineto S lineto S lineto closepath clip newpath moveto}N /endTexFig{end psf$SavedState restore}N /@beginspecial{ SDict begin /SpecialSave save N gsave normalscale currentpoint TR @SpecialDefaults count /ocount X /dcount countdictstack N}N /@setspecial {CLIP 1 eq{newpath 0 0 moveto hs 0 rlineto 0 vs rlineto hs neg 0 rlineto closepath clip}if ho vo TR hsc vsc scale ang rotate rwiSeen{rwi urx llx sub div rhiSeen{rhi ury lly sub div}{dup}ifelse scale llx neg lly neg TR }{rhiSeen{rhi ury lly sub div dup scale llx neg lly neg TR}if}ifelse CLIP 2 eq{newpath llx lly moveto urx lly lineto urx ury lineto llx ury lineto closepath clip}if /setpagedevice{pop}N statusdict begin /letter{} N /lettertray{}N /legal{}N /legaltray{}N /a4{}N /a4tray{}N /b4{}N /b4tray{}N /b5{}N /b5tray{}N /showpage{}N /erasepage{}N /copypage{}N end /letter{}N /lettertray{}N /legal{}N /legaltray{}N /a4{}N /a4tray{}N /b4{ }N /b4tray{}N /b5{}N /b5tray{}N /showpage{}N /erasepage{}N /copypage{}N newpath}N /@endspecial{count ocount sub{pop}repeat countdictstack dcount sub{end}repeat grestore SpecialSave restore end}N /@defspecial{SDict begin}N /@fedspecial{end}B /li{lineto}B /rl{rlineto}B /rc{rcurveto}B /np {/SaveX currentpoint /SaveY X N 1 setlinecap newpath}N /st{stroke SaveX SaveY moveto}N /fil{fill SaveX SaveY moveto}N /ellipse{/endangle X /startangle X /yrad X /xrad X /savematrix matrix currentmatrix N TR xrad yrad scale 0 0 1 startangle endangle arc savematrix setmatrix}N end %%EndProcSet TeXDict begin 40258431 52099146 1000 300 300 (iccs99.dvi) @start /Fa 2 51 df<1218127812981218AC12FF08107D8F0F>49 D<121FEA6180EA40C0EA806012C01200A213C0EA0180EA030012065AEA10201220EA7FC0 12FF0B107F8F0F>I E /Fb 3 52 df<1206120E12FE120EB1EAFFE00B157D9412>49 DII E /Fc 3 107 df<3807FFF83800E00E14071403A2EA01C01407A2140E38 03803CEBFFF0A2EB803C3807001C140EA3000E131CA214381470381C01E0B5120018177F 961B>66 D<120313801300C7FCA6121C12241246A25A120C5AA31231A21232A2121C0917 7F960C>105 D<1318133813101300A6EA01C0EA0220EA0430A2EA08601200A313C0A4EA 0180A4EA630012E312C612780D1D80960E>I E /Fd 10 106 df<0040130400C0130C00 6013186C13306C13606C13C03806018038030300EA0186EA00CC13781330A2137813CCEA 0186EA030338060180380C00C048136048133048131848130C0040130416187A9623>2 D15 D24 D<1506A381A216801501ED00C01660167016 18B8FCA2C912181670166016C0ED018015031600A21506A328187E962D>33 D50 D68 D<130F1338136013E0EA01C0AFEA0380EA0700121E12F8121E 1207EA0380EA01C0AFEA00E013601338130F102D7DA117>102 D<12F8121E1207EA0380 EA01C0AFEA00E013601338130F1338136013E0EA01C0AFEA0380EA0700121E12F8102D7D A117>I<1320136013C0A3EA0180A2EA0300A31206A25AA25AA35AA25AA35AA21260A37E A27EA37EA27EA27EA3EA0180A2EA00C0A3136013200B2E7CA112>I<12C0A21260A37EA2 7EA37EA27EA27EA3EA0180A2EA00C0A31360A213C0A3EA0180A2EA0300A31206A25AA25A A35AA25AA35AA20B2E7EA112>I E /Fe 8 121 df<127012F8A3127005057C840D>58 D<127012F012F8A212781208A31210A31220A21240050E7C840D>I<9039FFF01FF89039 0FC007809138800600010713046E5A5D01035B6E5A010113C0ECF18002F3C7FCEB00F214 FC1478147CA314BEEB011EEB021F1304EB0C0F01187FEB100701207F1340EB8003D80100 7F00071301001F497E39FFC01FFE251F7F9E26>88 D<39FFF001FF390F80007815301540 000714C06D13800003EB0100EBE0025C00015BEBF018141000005B6D5AEB7880137D91C7 FC7F133CA213381378A3137013F0A4485AEA1FFE201F7F9E1A>I<147C14CEEB019E1303 140CEB0700A4130EA3EBFFF0EB0E00A25BA55BA55BA55BA45B1201EA3180127948C7FC12 62123C17297E9F16>102 DI<13E01201A2EA00C01300A7120E1213EA23801243A3EA870012 07A2120EA25AA21320EA3840A31380EA1900120E0B1F7E9E10>105 D<3803C1C0380C622038103470EB38F012201460384070001200A35BA314203861C04012 F1148012E238446300EA383C14147E931A>120 D E /Ff 38 122 df12 D<91393FC0FF809139E0 E38040903A0181E600E00103EBEC01EC80DC903A07001C00C0031813001538A2130EA25D 90B7128090390E007003EE07005BA215E0160EA25BA24A485AA301701540EE3880EC0380 A301E0EC1900ED000E4A90C7FC13C0140612013831860E38798F0C38F31E1838620C3038 3C07C02B29829F28>14 D<130113021304130813101320136013C0EA0180A2EA03005A12 06120E120C121C12181238A212301270A21260A212E0A25AAD12401260A212207EA27E10 2E79A113>40 D<13107F7F130613021303A37F1480A71303A31400A35BA21306A2130E13 0CA2131C1318133813301370136013E05B485A90C7FC5A12065A5A5A5A1280112E80A113 >I45 D<127012F8A212F012E005057B840E>I<1302A2 1306130E133C13DCEA031C12001338A41370A413E0A4EA01C0A4EA0380A41207EAFFF80F 1E7B9D17>49 D51 D53 D56 D<131FEB7180EBC0C0EA0180000313E0EA07005A1301121EA314C0EA1C03A31307EB0F80 120CEA06373803C700EA000F130EA25B1260EAF0385BEAE060EA80C0EA4380003EC7FC13 1F7B9D17>I<14021406A2140E141EA2143F142F144F14CF148FEB010FA21302A2130413 08A20110138014071320EB3FFFEB40071380A2EA0100A2120212061204001E14C039FF80 7FF81D207E9F22>65 D67 D<3801FFF038001F00131EA35BA45BA45BA4 485AA4485AA4485AA4120FEAFFF0141F7D9E12>73 D<3801FFF8D8001FC7FC131EA35BA4 5BA45BA4485AA315803903C00100A25C140238078006A25C141C380F0078B512F8191F7D 9E1D>76 DI<01FFEB3F E0011FEB0F001504EB1780A201275BEB23C0A3903841E010A214F0134001805B1478A348 486C5AA3141E00025CA2140FA24891C7FC80A2120C001C1302EAFF80231F7D9E22>I<48 B5128039001E00E015701538153C5BA4491378A215F015E09038F003C0EC0F00EBFFFC01 F0C7FC485AA4485AA4485AA4120FEAFFF01E1F7D9E1F>80 D<393FFC0FF83907C003C090 38800100A3380F0002A4001E5BA4485BA4485BA4485BA35CA200705B49C7FCEA3002EA38 04EA0C18EA07E01D20779E22>85 D97 DI<137EEA01C138030080EA0E07 121E001C1300EA3C0248C7FCA35AA5EA70011302EA3004EA1838EA07C011147C9315>I< 1478EB03F8EB0070A414E0A4EB01C0A213F1EA038938070780EA0E03121C123C38380700 1278A3EAF00EA31420EB1C40A2EA703C135C38308C80380F070015207C9F17>I<137CEA 01C2EA0701120E121C123CEA3802EA780CEA7BF0EA7C0012F0A4127013011302EA3804EA 1838EA07C010147C9315>I<1478EB019CEB033CA2EB07181400A2130EA5EBFFE0EB1C00 A45BA55BA55BA5485AA35B1231007BC7FC12F31266123C1629829F0E>III<13C0EA01E0A213C0C7FC A7120E12131223EA4380EA4700A21287120EA35AA3EA38401380A21270EA31001232121C 0B1F7C9E0E>I108 D<391C0F80F0392630C318394740640C903880680EEB0070A200 8E495A120EA34848485AA3ED70803A3803807100A215E115623970070064D83003133821 147C9325>I<381C0F80382630C0384740601380EB0070A2008E13E0120EA3381C01C0A3 EB038400381388A2EB0708EB031000701330383001C016147C931A>I<137CEA01C33803 0180000E13C0121E001C13E0123C1278A338F003C0A3EB07801400EA700F130EEA3018EA 1870EA07C013147C9317>I<3801C1E0380262183804741C1378EB701EA2EA08E01200A3 3801C03CA3143838038078147014E0EBC1C038072380EB1E0090C7FCA2120EA45AA2B47E 171D809317>I114 D<13FCEA0302EA0601EA0C03130713061300EA0F8013F0EA07 F8EA03FCEA003E130E1270EAF00CA2EAE008EA4010EA2060EA1F8010147D9313>II<000E13C0001313E0382301C0EA4381EA4701A238870380120EA338 1C0700A31410EB0E201218A2381C1E40EA0C263807C38014147C9318>I<000E13C00013 13E0382301C0EA4381EA4701A238870380120EA3381C0700A4130E1218A2EA1C1EEA0C3C EA07DCEA001CA25B12F05BEAE060485AEA4380003EC7FC131D7C9316>121 D E /Fg 44 122 df<903A01FF803FE0011F9038E3FFF8903A7F80FFF01E903AFE007F80 1FD801F849485A000349EC7F80D807F05BA30200EC3F00171E94C7FCA4B91280A33B07F0 00FE003FB3A33C7FFF0FFFE3FFF8A3352A7FA939>14 D<130E131E137EEA07FE12FFA212 F81200B3AB387FFFFEA317277BA622>49 DII<140E141E143E147E14FEA213011303EB077E130EA2131C1338137013E0A2EA01 C0EA0380EA0700120EA25A5A5A5AB612F8A3C7EAFE00A890383FFFF8A31D277EA622>I< 00081303381F803FEBFFFE5C5C5C5C148049C7FC001CC8FCA6EB7FC0381DFFF8381F80FC 381E003E0008133FC71380EC1FC0A215E0A21218127C12FEA315C05A0078EB3F80A26CEB 7F00381F01FE380FFFF800035BC613801B277DA622>II<1238123E003FB512F0A34814E015C015 8015003870000EA25C485B5C5CC6485AA2495A130791C7FC5B5B131E133EA2137E137CA2 13FCA41201A76C5A13701C297CA822>II<121C123E127FEA FF80A3EA7F00123E121CC7FCA9121C123E127FEAFF80A3EA7F00123E121C091B7B9A13> 58 D65 D<91393FF00180903903FFFE03010FEBFF8790393FF007DF9039FF8001FF4848C7127F48 48143FD807F0141F000F150F48481407A2485A1603127F5B93C7FC12FFA9127FA26DEC03 80123FA26C7EEE07006C7E0007150ED803FC141E6C6C5C6C6C6C13F890393FF007E0010F B55A010391C7FC9038003FF829297CA832>67 DI70 D73 D76 D79 DI82 D<90387F80603903FFF0E04813F9380F807F381F001F003E1307481303140112FCA21400 7EA26C140013C0EA7FFEEBFFE06C13FC6C7F6CEBFF806C14C07EC66C13E0010713F0EB00 7FEC0FF8140714030060130112E0A36C14F0A26C13036C14E0B4EB07C09038E01F8000F3 B5120000E15B38C01FF01D297CA826>I<007FB712C0A39039803FC03FD87E0014070078 1503A20070150100F016E0A2481500A5C71500B3A4017FB512E0A32B287EA730>I 87 D<3803FF80000F13F0381F01F8383F807EA280D81F001380120EC7FCA3EB0FFF90B5 FC3807FC3FEA0FE0EA3F8013005A12FEA4007E137F007F13DF393F839FFC380FFF0F3801 FC031E1B7E9A21>97 DIIIII<9038FF81 F00003EBE7F8390FC1FE7C381F80FC9038007C3848EB7E1048EB7F00A66C137E6C137CEB 80FC380FC1F8381FFFE0001813800038C8FCA2123C123E383FFFF86C13FF15806C14C06C 14E0001F14F0383C0007007CEB01F8481300A4007CEB01F0003C14E0001FEB07C0390FC0 1F803903FFFE0038007FF01E287E9A22>II<1207EA0F80EA1F C0EA3FE0A3EA1FC0EA0F80EA0700C7FCA7EAFFE0A3120FB3A3EAFFFEA30F2B7DAA14>I< EAFFE0A3120FACEC1FFCA3EC0780EC0F00141E5C5C14E0EBE3C013E7EBEFE013FF80EBF3 F813E3EBC1FCEBC0FE147FA2EC3F80EC1FC0140F15E039FFFC3FFEA31F2A7EA924>107 DI<3BFFC07F800FF0903AC1FFE03FFC903A C383F0707E3B0FC603F8C07F903ACC01F9803F01D8D9FF00138001F05BA201E05BB03CFF FE1FFFC3FFF8A3351B7D9A3A>I<38FFC07F9038C1FFC09038C787E0390FCE03F013D881 13F0A213E0B03AFFFE3FFF80A3211B7D9A26>II<38FFE1FE9038E7FF809038FE07 E0390FF803F0496C7E496C7E818181A21680A716005DA25D4A5A01F05B6D485A9038FE0F E09038EFFF80D9E1FCC7FC01E0C8FCA9EAFFFEA321277E9A26>I<38FFC1F0EBC7FCEBCE 3E380FD87FA213F0143E141CEBE000B0B5FCA3181B7E9A1C>114 D<3803FE30380FFFF0EA1E03EA380048137012F0A27E00FE1300EAFFE0EA7FFEEBFF806C 13E06C13F0000713F8C6FCEB07FCEA600000E0137C143C7E14387E6C137038FF01E038F7 FFC000C11300161B7E9A1B>I<1370A413F0A312011203A21207381FFFF0B5FCA23807F0 00AD1438A61203EBF870000113603800FFC0EB1F8015267FA51B>I<39FFE03FF8A3000F 1303B214071207140F3A03F03BFF803801FFF338003FC3211B7D9A26>I<3AFFFE03FF80 A33A0FF0007800000714706D13F000035CEBFC0100015CA26C6C485AA2EBFF07017F90C7 FC148FEB3F8E14CEEB1FDCA2EB0FF8A36D5AA26D5AA26D5A211B7F9A24>I<3BFFFC3FFC 0FFEA33B0FE007E000E03B07F003F001C0A29039F807F80300031680A23B01FC0EFC0700 A291381EFE0F3A00FE1C7E0EA29039FF383F1E017F141C147890393FF01FB8A29138E00F F8011F5CA26D486C5AA36D486C5AA22F1B7F9A32>I<3AFFFE03FF80A33A0FF000780000 0714706D13F000035CEBFC0100015CA26C6C485AA2EBFF07017F90C7FC148FEB3F8E14CE EB1FDCA2EB0FF8A36D5AA26D5AA26D5AA2495AA2EA3807007C90C8FCEAFE0F130E5B133C EA7C78EA3FE0EA0F8021277F9A24>121 D E /Fh 80 125 df<90381F83E09038F06E30 3901C07878380380F8903800F03048EB7000A7B612803907007000B2383FE3FF1D20809F 1B>11 D<133FEBE0C0EA01C0380381E0EA0701A290C7FCA6B512E0EA0700B2383FC3FC16 20809F19>I<90381F81F89038F04F043901C07C06390380F80FEB00F05A0270C7FCA6B7 FC3907007007B23A3FE3FE3FE02320809F26>14 D34 D37 D<127012F812FCA212741204A31208A21210A212201240060E7C9F0D>39 D<13401380EA01005A12061204120C5AA212381230A212701260A412E0AC1260A4127012 30A212381218A27E120412067E7EEA008013400A2E7BA112>I<7E12407E12307E120812 0C7EA212077EA213801201A413C0AC1380A412031300A25A1206A25A120812185A12205A 5A0A2E7EA112>I<127012F012F8A212781208A31210A31220A21240050E7C840D>44 DI<127012F8A3127005057C840D>I<144014C0EB0180A3EB0300 A31306A25BA35BA35BA25BA35BA3485AA348C7FCA21206A35AA35AA25AA35AA35AA2122D 7EA117>II<13801203120F12F31203B3A6EA 07C0EAFFFE0F1E7C9D17>III<1306A2130EA2131E132EA2134E138EA2EA 010E1202A212041208A212101220A2124012C0B512F038000E00A7EBFFE0141E7F9D17> II<137CEA 0182EA0701380E0380EA0C0712183838030090C7FC12781270A2EAF1F0EAF21CEAF406EA F807EB0380A200F013C0A51270A214801238EB07001218EA0C0E6C5AEA01F0121F7E9D17 >I<1240387FFFE014C0A23840008038800100A21302485AA25B5BA25BA21360A213E05B 1201A41203A76C5A131F7E9D17>III<127012F8A312 701200AA127012F8A3127005147C930D>I<127012F8A312701200AA127012F012F8A212 781208A31210A31220A21240051D7C930D>I<007FB512F8B612FCC9FCA8B612FC6C14F8 1E0C7E9023>61 D63 DI<5B 497EA3497EA3EB09E0A3EB10F0A3EB2078A3497EA2EBC03EEB801EA248B5FCEB000FA200 02EB0780A348EB03C0A2120C001E14E039FF801FFE1F207F9F22>II<90380FE0109038381C309038E002703803C001 39078000F048C71270121E15305A1510127C127800F81400A91278007C1410123CA26C14 20A27E6C6C13406C6C13803900E00300EB380CEB0FF01C217E9F21>IIII<90380FE0109038381C309038E00270 3803C00139078000F048C71270121E15305A1510127C127800F81400A7EC3FFEEC01F000 781300127C123CA27EA27E6C7E3903C001703900E002309038380C1090380FF0001F217E 9F24>I<39FFF07FF8390F000780AD90B5FCEB0007AF39FFF07FF81D1F7E9E22>II<3807FFC038003E00131EB3A3122012F8A3EAF01CEA 403CEA6038EA1070EA0FC012207F9E17>I<39FFF007FC390F0003E0EC0180150014025C 5C5C5C5C5C49C7FC5B497E130FEB13C0EB21E01341EB80F0EB0078A28080A280EC0780A2 EC03C015E015F039FFF01FFE1F1F7E9E23>IIIIII82 D<3803F040380C0CC0EA1803EA3001EA6000A212E01440A36C13007E 127CEA7F80EA3FF86CB4FC00071380C613C0EB1FE013031301EB00F014707EA46C136014 E06C13C038F8018038C60300EA81FC14217E9F19>I<007FB512E038780F010060EB0060 00401420A200C0143000801410A400001400B3497E3803FFFC1C1F7E9E21>I<39FFF00F F8390F0003E0EC0080B3A46CEB01001380120314026C6C5A6C6C5AEB3830EB0FC01D207E 9E22>I<3BFFF07FF81FF03B1F000FC007C06C903907800180170015C001805C00071502 EC09E013C000035DEC19F01410D801E05CA2EC2078D800F05CA2EC403C01785CA2EC801E 017C1460013C144090383D000F133F6D5CA2011E1307010E91C7FCA2010C7F010413022C 207F9E2F>87 D<39FFF001FF391F800078000F146012076D1340000314807F3901F00100 1200EBF802EB7C06EB3C04EB3E08131EEB1F10EB0FB0EB07A014E06D5AACEB3FFC201F7F 9E22>89 D<12FFA212C0B3B3A512FFA2082D7CA10D>91 DI<12FFA21203B3B3A512FF A2082D80A10D>I97 D<121C12FC121CAA137CEA 1D87381E0180EB00C0001C13E01470A21478A6147014F014E0001E13C0381A0180381987 00EA107C15207E9F19>IIII<137CEA01C6EA030F1207EA0E061300A7EAFF F0EA0E00B2EA7FE01020809F0E>I<14E03803E330EA0E3CEA1C1C38380E00EA780FA5EA 380E6C5AEA1E38EA33E00020C7FCA21230A2EA3FFE381FFF8014C0383001E03860007048 1330A4006013606C13C0381C03803803FC00141F7F9417>I<121C12FC121CAA137C1386 EA1D03001E1380A2121CAE38FF8FF014207E9F19>I<1238127CA31238C7FCA6121C12FC 121CB1EAFF80091F7F9E0C>I<13E0EA01F0A3EA00E01300A61370EA07F012001370B3A3 1260EAF06013C0EA6180EA3F000C28829E0E>I<121C12FC121CAAEB1FE0EB0780EB0600 13045B5B5B136013E0EA1DF0EA1E70EA1C38133C131C7F130F7F148014C038FF9FF01420 7E9F18>I<121C12FC121CB3ABEAFF8009207F9F0C>I<391C3E03E039FCC30C30391D0390 38391E01E01CA2001C13C0AE3AFF8FF8FF8021147E9326>IIII<3801F04038070CC0EA0E02EA1C03EA380112 78127012F0A6127012781238EA1C03EA0C05EA0709EA01F1EA0001A8EB0FF8151D7F9318 >III<1202A31206A2120EA2123EEAFFF8EA0E00AB1304A5EA070812 03EA01F00E1C7F9B12>I<381C0380EAFC1FEA1C03AE1307120CEA061B3803E3F014147E 9319>I<38FF83F8383E00E0001C13C06C1380A338070100A21383EA0382A2EA01C4A213 E4EA00E8A21370A3132015147F9318>I<39FF9FE1FC393C078070391C030060EC802000 0E1440A214C0D80704138014E0A239038861001471A23801D032143A143E3800E01CA2EB 6018EB40081E147F9321>I<38FF87F8381E03C0380E0180EB0300EA0702EA0384EA01C8 13D8EA00F01370137813F8139CEA010E1202EA060738040380000C13C0003C13E038FE07 FC16147F9318>I<38FF83F8383E00E0001C13C06C1380A338070100A21383EA0382A2EA 01C4A213E4EA00E8A21370A31320A25BA3EAF080A200F1C7FC1262123C151D7F9318>I< EA7FFFEA700E1260EA401C133813781370EA00E0120113C0EA038012071301120E121EEA 1C03EA3802EA7006130EEAFFFE10147F9314>III E /Fi 23 119 df68 D76 D 82 D<3801FE023807FF86381F01FE383C007E007C131E0078130EA200F81306A27E1400 B4FC13E06CB4FC14C06C13F06C13F86C13FC000313FEEA003F1303EB007F143FA200C013 1FA36C131EA26C133C12FCB413F838C7FFE00080138018227DA11F>I<007FB61280A239 7E03F80F00781407007014030060140100E015C0A200C01400A400001500B3A248B512F0 A222227EA127>II97 DI100 D<13FE3807FF80380F87 C0381E01E0003E13F0EA7C0014F812FCA2B5FCA200FCC7FCA3127CA2127E003E13186C13 30380FC0703803FFC0C6130015167E951A>II<3801FE0F3907FFBF8038 0F87C7381F03E7391E01E000003E7FA5001E5BEA1F03380F87C0EBFF80D809FEC7FC0018 C8FCA2121C381FFFE06C13F86C13FE001F7F383C003F48EB0F80481307A40078EB0F006C 131E001F137C6CB45A000113C019217F951C>II<121C123E127FA3123E121CC7FCA7 B4FCA2121FB2EAFFE0A20B247EA310>I108 D<38FF07E0EB1FF8381F307CEB403CEB803EA21300AE39FFE1FFC0A21A167E951F>110 D<13FE3807FFC0380F83E0381E00F0003E13F848137CA300FC137EA7007C137CA26C13F8 381F01F0380F83E03807FFC03800FE0017167E951C>I<38FF0FE0EB3FF8381FE07CEB80 3E497E1580A2EC0FC0A8EC1F80A29038803F00EBC03EEBE0FCEB3FF8EB0FC090C8FCA8EA FFE0A21A207E951F>I114 DI<487EA41203A21207A2120F123FB5FCA2EA0F80ABEB8180A5EB8300EA07C3EA 03FEEA00F811207F9F16>I<38FF01FEA2381F003EAF147E14FE380F81BE3907FF3FC0EA 01FC1A167E951F>I<39FFE01FE0A2391F800700000F1306EBC00E0007130C13E000035B A26C6C5AA26C6C5AA2EB7CC0A2137F6D5AA26DC7FCA2130EA21B167F951E>I E /Fj 23 121 df67 D 73 D<3803FFF83800701C1406140713E0A43801C00EA2141C143838038060EBFF80EB80 00A248C7FCA4120EA45AB47E181A7D991A>80 D83 D97 D99 DII<1307EB0980131BEB3B00133813301370A4EA07FFEA00E0A5 485AA5485AA490C7FC5AA21206126612E412CC1270112181990C>I<13F338038B803806 0700120E120C121CEA380EA4EA301CA3EA183C5BEA07B8EA0038A25B1260EAE0E0EAC1C0 007FC7FC11177E8F12>II<1203120712061200 A61238124C124E128E129CA2121C1238A212701272A212E212E41264123808197C980C> I<121F1207A3120EA4121CA41238A41270A412E4A412E81230081A7D990A>108 D<38307C1E38598663399E0783801403129CA239380E0700A3140ED8701C1340A2141C15 8038E0380C39601807001A107C8F1F>III113 D II<1206120EA45AA2EAFFC0EA1C005AA45AA412E1A312E212E4 12380A177C960D>III120 D E /Fk 19 119 df<1380EA010012025A120C120812185AA35AA412E0AA1260A47EA37E 1208120C12047E7EEA008009267D9B0F>40 D<7E12407E7E12181208120C7EA37EA41380 AA1300A41206A35A1208121812105A5A5A09267E9B0F>I<126012F0A212701210A31220 A21240A2040B7D830B>44 D<126012F0A2126004047D830B>46 D<12035AB4FC1207B3A2 EA7FF80D187D9713>49 D57 D73 D83 D97 D<12FC121CA913FCEA1D07381E0380381C01C0130014E0A6EB01C01480381E0300EA1906 EA10F8131A809915>I<133F1307A9EA03E7EA0C17EA180F487E127012E0A6126012706C 5AEA1C373807C7E0131A7F9915>100 DI<1218123CA212181200A612 FC121CAE12FF081A80990A>105 D110 D112 D114 DI<1208A41218A21238EAFFC0EA3800 A81320A41218EA1C40EA07800B177F960F>I<38FF0F80383C0700EA1C061304A26C5AA2 6C5AA3EA03A0A2EA01C0A36C5A11107F8F14>118 D E end %%EndProlog %%BeginSetup %%Feature: *Resolution 300dpi TeXDict begin %%EndSetup %%Page: 1 1 1 0 bop 697 207 a Fk(In)13 b Fj(Pr)n(o)n(c)n(e)n(e)n(dings)e(of)i(the)f (Sixth)g(International)e(Col)r(lo)n(quium)i(on)h(Co)n(gnitive)f(Scienc) n(e)p Fk(,)1203 253 y(San)i(Sebastian,)g(Spain,)h(1999)e(\(in)o(vited)i (pap)q(er\).)298 509 y Fi(The)j(Role)g(of)h(Unlab)r(eled)e(Data)i(in)f (Sup)r(ervised)f(Learning)777 610 y Fh(T)l(om)e(M.)f(Mitc)o(hell)668 667 y(Sc)o(ho)q(ol)i(of)e(Computer)h(Science)673 723 y(Carnegie)h(Mellon)g(Univ)o(ersit)o(y)713 836 y(T)l(om.Mitc)o(hell@cm) o(u.edu)861 949 y(ICCS-99)848 1006 y(Ma)o(y)f(1999)-59 1149 y Fg(1)69 b(In)n(tro)r(duction)-59 1250 y Fh(Most)13 b(computational)i(mo)q(dels)h(of)e(sup)q(ervised)i(learning)g(rely)f (only)g(on)g(lab)q(eled)h(training)g(examples,)f(and)f(ignore)-59 1307 y(the)h(p)q(ossible)i(role)f(of)f(unlab)q(eled)j(data.)i(This)c (is)f(true)h(b)q(oth)f(for)g(cognitiv)o(e)h(science)h(mo)q(dels)f(of)f (learning)h(suc)o(h)g(as)-59 1363 y(SO)o(AR)h([New)o(ell)g(1990])e(and) h(A)o(CT-R)g([Anderson,)h(et)f(al.)23 b(1995],)14 b(and)j(for)e(mac)o (hine)i(learning)h(and)e(data)f(mining)-59 1420 y(algorithms)f(suc)o(h) h(as)g(decision)h(tree)e(learning)i(and)f(inductiv)o(e)h(logic)g (programming)d(\(see,)i(e.g.,)e([Mitc)o(hell)j(1997]\).)-59 1476 y(In)21 b(this)h(pap)q(er)f(w)o(e)f(consider)i(the)f(p)q(oten)o (tial)g(role)g(of)g Ff(unlab)n(ele)n(d)e Fh(data)h(in)i(sup)q(ervised)g (learning.)38 b(W)l(e)21 b(presen)o(t)-59 1533 y(an)d(algorithm)g(and)g (exp)q(erimen)o(tal)h(results)g(demonstrating)e(that)g(unlab)q(eled)k (data)c(can)h(signi\014can)o(tly)i(impro)o(v)o(e)-59 1589 y(learning)15 b(accuracy)f(in)h(certain)f(practical)g(problems.)20 b(W)l(e)14 b(then)h(iden)o(tify)g(the)f(abstract)e(problem)j(structure) e(that)-59 1645 y(enables)18 b(the)f(algorithm)h(to)e(successfully)j (utilize)h(this)d(unlab)q(eled)j(data,)d(and)g(pro)o(v)o(e)f(that)h (unlab)q(eled)j(data)c(will)-59 1702 y(b)q(o)q(ost)h(learning)h (accuracy)f(for)f(problems)i(in)f(this)h(class.)25 b(The)17 b(problem)h(class)f(w)o(e)g(iden)o(tify)h(includes)h(problems)-59 1758 y(where)f(the)g(features)g(describing)i(the)e(examples)g(are)g (redundan)o(tly)h(su\016cien)o(t)g(for)e(classifying)i(the)f(example;)i (a)-59 1815 y(notion)f(w)o(e)f(mak)o(e)g(precise)i(in)g(the)f(pap)q (er.)31 b(This)19 b(problem)g(class)g(includes)i(man)o(y)e(natural)f (learning)i(problems)-59 1871 y(faced)g(b)o(y)g(h)o(umans,)h(suc)o(h)f (as)g(learning)h(a)f(seman)o(tic)g(lexicon)h(o)o(v)o(er)e(noun)i (phrases)f(in)h(natural)f(language,)h(and)-59 1928 y(learning)14 b(to)e(recognize)i(ob)s(jects)f(from)f(m)o(ultiple)j(sensor)d(inputs.) 20 b(W)l(e)14 b(argue)e(that)g(mo)q(dels)i(of)f(h)o(uman)g(and)g (animal)-59 1984 y(learning)19 b(should)f(consider)h(more)e(strongly)h (the)f(p)q(oten)o(tial)i(role)f(of)f(unlab)q(eled)j(data,)e(and)g(that) f(man)o(y)g(natural)-59 2041 y(learning)f(problems)g(\014t)f(the)h (class)f(w)o(e)g(iden)o(tify)l(.)-59 2184 y Fg(2)69 b(A)15 b(Sup)r(ervised)g(Learning)h(Problem:)25 b(Learning)16 b(to)f(Classify)f(W)-6 b(eb)16 b(P)n(ages)-59 2285 y Fh(T)l(o)j(illustrate)h(the)f(role)g(of)g(unlab)q(eled)j(data)c(in)i (sup)q(ervised)h(learning,)g(consider)f(the)f(problem)h(of)e(learning)i (to)-59 2342 y(classify)12 b(pages)e(of)h(h)o(yp)q(ertext)g(from)f(the) h(w)o(orld)g(wide)h(w)o(eb,)f(giv)o(en)g(lab)q(eled)i(training)f(data)e (consisting)i(of)f(individual)-59 2398 y(w)o(eb)k(pages)h(along)f(with) h(their)g(correct)f(classi\014cations)i(\(Figure)f(1\).)k(W)l(e)c(ha)o (v)o(e)f(b)q(een)i(studying)f(this)g(problem)g(as)-59 2455 y(part)g(of)h(our)g(larger)g(researc)o(h)g(goal)f(of)h (automatically)g(extracting)g(information)h(from)e(the)h(w)o(eb)g([Cra) o(v)o(en)f(et)h(al.)934 2581 y(1)p eop %%Page: 2 2 2 1 bop 420 266 a 16577003 10940820 5262540 13287915 34666987 32956661 startTexFig 420 266 a %%BeginDocument: /usr0/mitchell/pictures/page-links.ps /arrowHeight 10 def /arrowWidth 5 def /IdrawDict 52 dict def IdrawDict begin /reencodeISO { dup dup findfont dup length dict begin { 1 index /FID ne { def }{ pop pop } ifelse } forall /Encoding ISOLatin1Encoding def currentdict end definefont } def /ISOLatin1Encoding [ /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /space/exclam/quotedbl/numbersign/dollar/percent/ampersand/quoteright /parenleft/parenright/asterisk/plus/comma/minus/period/slash /zero/one/two/three/four/five/six/seven/eight/nine/colon/semicolon /less/equal/greater/question/at/A/B/C/D/E/F/G/H/I/J/K/L/M/N /O/P/Q/R/S/T/U/V/W/X/Y/Z/bracketleft/backslash/bracketright /asciicircum/underscore/quoteleft/a/b/c/d/e/f/g/h/i/j/k/l/m /n/o/p/q/r/s/t/u/v/w/x/y/z/braceleft/bar/braceright/asciitilde /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef /.notdef/dotlessi/grave/acute/circumflex/tilde/macron/breve /dotaccent/dieresis/.notdef/ring/cedilla/.notdef/hungarumlaut /ogonek/caron/space/exclamdown/cent/sterling/currency/yen/brokenbar /section/dieresis/copyright/ordfeminine/guillemotleft/logicalnot /hyphen/registered/macron/degree/plusminus/twosuperior/threesuperior /acute/mu/paragraph/periodcentered/cedilla/onesuperior/ordmasculine /guillemotright/onequarter/onehalf/threequarters/questiondown /Agrave/Aacute/Acircumflex/Atilde/Adieresis/Aring/AE/Ccedilla /Egrave/Eacute/Ecircumflex/Edieresis/Igrave/Iacute/Icircumflex /Idieresis/Eth/Ntilde/Ograve/Oacute/Ocircumflex/Otilde/Odieresis /multiply/Oslash/Ugrave/Uacute/Ucircumflex/Udieresis/Yacute /Thorn/germandbls/agrave/aacute/acircumflex/atilde/adieresis /aring/ae/ccedilla/egrave/eacute/ecircumflex/edieresis/igrave /iacute/icircumflex/idieresis/eth/ntilde/ograve/oacute/ocircumflex /otilde/odieresis/divide/oslash/ugrave/uacute/ucircumflex/udieresis /yacute/thorn/ydieresis ] def /Times-Bold reencodeISO def /Times-Roman reencodeISO def /none null def /numGraphicParameters 17 def /stringLimit 65535 def /Begin { save numGraphicParameters dict begin } def /End { end restore } def /SetB { dup type /nulltype eq { pop false /brushRightArrow idef false /brushLeftArrow idef true /brushNone idef } { /brushDashOffset idef /brushDashArray idef 0 ne /brushRightArrow idef 0 ne /brushLeftArrow idef /brushWidth idef false /brushNone idef } ifelse } def /SetCFg { /fgblue idef /fggreen idef /fgred idef } def /SetCBg { /bgblue idef /bggreen idef /bgred idef } def /SetF { /printSize idef /printFont idef } def /SetP { dup type /nulltype eq { pop true /patternNone idef } { dup -1 eq { /patternGrayLevel idef /patternString idef } { /patternGrayLevel idef } ifelse false /patternNone idef } ifelse } def /BSpl { 0 begin storexyn newpath n 1 gt { 0 0 0 0 0 0 1 1 true subspline n 2 gt { 0 0 0 0 1 1 2 2 false subspline 1 1 n 3 sub { /i exch def i 1 sub dup i dup i 1 add dup i 2 add dup false subspline } for n 3 sub dup n 2 sub dup n 1 sub dup 2 copy false subspline } if n 2 sub dup n 1 sub dup 2 copy 2 copy false subspline patternNone not brushLeftArrow not brushRightArrow not and and { ifill } if brushNone not { istroke } if 0 0 1 1 leftarrow n 2 sub dup n 1 sub dup rightarrow } if end } dup 0 4 dict put def /Circ { newpath 0 360 arc patternNone not { ifill } if brushNone not { istroke } if } def /CBSpl { 0 begin dup 2 gt { storexyn newpath n 1 sub dup 0 0 1 1 2 2 true subspline 1 1 n 3 sub { /i exch def i 1 sub dup i dup i 1 add dup i 2 add dup false subspline } for n 3 sub dup n 2 sub dup n 1 sub dup 0 0 false subspline n 2 sub dup n 1 sub dup 0 0 1 1 false subspline patternNone not { ifill } if brushNone not { istroke } if } { Poly } ifelse end } dup 0 4 dict put def /Elli { 0 begin newpath 4 2 roll translate scale 0 0 1 0 360 arc patternNone not { ifill } if brushNone not { istroke } if end } dup 0 1 dict put def /Line { 0 begin 2 storexyn newpath x 0 get y 0 get moveto x 1 get y 1 get lineto brushNone not { istroke } if 0 0 1 1 leftarrow 0 0 1 1 rightarrow end } dup 0 4 dict put def /MLine { 0 begin storexyn newpath n 1 gt { x 0 get y 0 get moveto 1 1 n 1 sub { /i exch def x i get y i get lineto } for patternNone not brushLeftArrow not brushRightArrow not and and { ifill } if brushNone not { istroke } if 0 0 1 1 leftarrow n 2 sub dup n 1 sub dup rightarrow } if end } dup 0 4 dict put def /Poly { 3 1 roll newpath moveto -1 add { lineto } repeat closepath patternNone not { ifill } if brushNone not { istroke } if } def /Rect { 0 begin /t exch def /r exch def /b exch def /l exch def newpath l b moveto l t lineto r t lineto r b lineto closepath patternNone not { ifill } if brushNone not { istroke } if end } dup 0 4 dict put def /Text { ishow } def /idef { dup where { pop pop pop } { exch def } ifelse } def /ifill { 0 begin gsave patternGrayLevel -1 ne { fgred bgred fgred sub patternGrayLevel mul add fggreen bggreen fggreen sub patternGrayLevel mul add fgblue bgblue fgblue sub patternGrayLevel mul add setrgbcolor eofill } { eoclip originalCTM setmatrix pathbbox /t exch def /r exch def /b exch def /l exch def /w r l sub ceiling cvi def /h t b sub ceiling cvi def /imageByteWidth w 8 div ceiling cvi def /imageHeight h def bgred bggreen bgblue setrgbcolor eofill fgred fggreen fgblue setrgbcolor w 0 gt h 0 gt and { l b translate w h scale w h true [w 0 0 h neg 0 h] { patternproc } imagemask } if } ifelse grestore end } dup 0 8 dict put def /istroke { gsave brushDashOffset -1 eq { [] 0 setdash 1 setgray } { brushDashArray brushDashOffset setdash fgred fggreen fgblue setrgbcolor } ifelse brushWidth setlinewidth originalCTM setmatrix stroke grestore } def /ishow { 0 begin gsave fgred fggreen fgblue setrgbcolor /fontDict printFont printSize scalefont dup setfont def /descender fontDict begin 0 [FontBBox] 1 get FontMatrix end transform exch pop def /vertoffset 1 printSize sub descender sub def { 0 vertoffset moveto show /vertoffset vertoffset printSize sub def } forall grestore end } dup 0 3 dict put def /patternproc { 0 begin /patternByteLength patternString length def /patternHeight patternByteLength 8 mul sqrt cvi def /patternWidth patternHeight def /patternByteWidth patternWidth 8 idiv def /imageByteMaxLength imageByteWidth imageHeight mul stringLimit patternByteWidth sub min def /imageMaxHeight imageByteMaxLength imageByteWidth idiv patternHeight idiv patternHeight mul patternHeight max def /imageHeight imageHeight imageMaxHeight sub store /imageString imageByteWidth imageMaxHeight mul patternByteWidth add string def 0 1 imageMaxHeight 1 sub { /y exch def /patternRow y patternByteWidth mul patternByteLength mod def /patternRowString patternString patternRow patternByteWidth getinterval def /imageRow y imageByteWidth mul def 0 patternByteWidth imageByteWidth 1 sub { /x exch def imageString imageRow x add patternRowString putinterval } for } for imageString end } dup 0 12 dict put def /min { dup 3 2 roll dup 4 3 roll lt { exch } if pop } def /max { dup 3 2 roll dup 4 3 roll gt { exch } if pop } def /midpoint { 0 begin /y1 exch def /x1 exch def /y0 exch def /x0 exch def x0 x1 add 2 div y0 y1 add 2 div end } dup 0 4 dict put def /thirdpoint { 0 begin /y1 exch def /x1 exch def /y0 exch def /x0 exch def x0 2 mul x1 add 3 div y0 2 mul y1 add 3 div end } dup 0 4 dict put def /subspline { 0 begin /movetoNeeded exch def y exch get /y3 exch def x exch get /x3 exch def y exch get /y2 exch def x exch get /x2 exch def y exch get /y1 exch def x exch get /x1 exch def y exch get /y0 exch def x exch get /x0 exch def x1 y1 x2 y2 thirdpoint /p1y exch def /p1x exch def x2 y2 x1 y1 thirdpoint /p2y exch def /p2x exch def x1 y1 x0 y0 thirdpoint p1x p1y midpoint /p0y exch def /p0x exch def x2 y2 x3 y3 thirdpoint p2x p2y midpoint /p3y exch def /p3x exch def movetoNeeded { p0x p0y moveto } if p1x p1y p2x p2y p3x p3y curveto end } dup 0 17 dict put def /storexyn { /n exch def /y n array def /x n array def n 1 sub -1 0 { /i exch def y i 3 2 roll put x i 3 2 roll put } for } def /SSten { fgred fggreen fgblue setrgbcolor dup true exch 1 0 0 -1 0 6 -1 roll matrix astore } def /FSten { dup 3 -1 roll dup 4 1 roll exch newpath 0 0 moveto dup 0 exch lineto exch dup 3 1 roll exch lineto 0 lineto closepath bgred bggreen bgblue setrgbcolor eofill SSten } def /Rast { exch dup 3 1 roll 1 0 0 -1 0 6 -1 roll matrix astore } def /arrowhead { 0 begin transform originalCTM itransform /taily exch def /tailx exch def transform originalCTM itransform /tipy exch def /tipx exch def /dy tipy taily sub def /dx tipx tailx sub def /angle dx 0 ne dy 0 ne or { dy dx atan } { 90 } ifelse def gsave originalCTM setmatrix tipx tipy translate angle rotate newpath arrowHeight neg arrowWidth 2 div moveto 0 0 lineto arrowHeight neg arrowWidth 2 div neg lineto patternNone not { originalCTM setmatrix /padtip arrowHeight 2 exp 0.25 arrowWidth 2 exp mul add sqrt brushWidth mul arrowWidth div def /padtail brushWidth 2 div def tipx tipy translate angle rotate padtip 0 translate arrowHeight padtip add padtail add arrowHeight div dup scale arrowheadpath ifill } if brushNone not { originalCTM setmatrix tipx tipy translate angle rotate arrowheadpath istroke } if grestore end } dup 0 9 dict put def /arrowheadpath { newpath arrowHeight neg arrowWidth 2 div moveto 0 0 lineto arrowHeight neg arrowWidth 2 div neg lineto } def /leftarrow { 0 begin y exch get /taily exch def x exch get /tailx exch def y exch get /tipy exch def x exch get /tipx exch def brushLeftArrow { tipx tipy tailx taily arrowhead } if end } dup 0 4 dict put def /rightarrow { 0 begin y exch get /tipy exch def x exch get /tipx exch def y exch get /taily exch def x exch get /tailx exch def brushRightArrow { tipx tipy tailx taily arrowhead } if end } dup 0 4 dict put def %I Idraw 10 Grid 8 8 Begin %I b u %I cfg u %I cbg u %I f u %I p u %I t [ 0.799705 0 0 0.799705 0 0 ] concat /originalCTM matrix currentmatrix def Begin %I Rect %I b 65535 1 0 0 [] 0 SetB %I cfg Blue 0 0 1 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 167 81 407 369 Rect End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-bold-r-normal-*-14-*-*-*-*-*-*-* Times-Bold 14 SetF %I t [ 1 0 0 1 256 534 ] concat %I [ (Professor John Smith) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 272 499 ] concat %I [ (Hi! I teach computer courses and) (advise students, including) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 288 459 ] concat %I [ (Mary Kay) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 288 443 ] concat %I [ (Bill Blue) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 272 411 ] concat %I [ (In my spare time I work on a number) (of research projects, including) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 288 379 ] concat %I [ (- machine learning for web) ( page classification) () (- active learning for robot control) ] Text End Begin %I Text %I cfg Blue 0 0 1 SetCFg %I f -*-times-medium-r-normal-*-12-*-*-*-*-*-*-* Times-Roman 12 SetF %I t [ 1 0 0 1 288 315 ] concat %I [ (- software engineering) ] Text End Begin %I Line %I b 65535 1 0 0 [] 0 SetB %I cfg Blue 0 0 1 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 215 257 255 257 Line %I 1 End Begin %I Line %I b 65535 1 0 0 [] 0 SetB %I cfg Blue 0 0 1 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 170 ] concat %I 319 161 383 161 Line %I 1 End Begin %I Text %I cfg Red 1 0 0 SetCFg %I f -*-times-medium-r-normal-*-14-*-*-*-*-*-*-* Times-Roman 14 SetF %I t [ 1 0 0 1 103 622 ] concat %I [ (My Advisor) ] Text End Begin %I Text %I cfg Red 1 0 0 SetCFg %I f -*-times-medium-r-normal-*-14-*-*-*-*-*-*-* Times-Roman 14 SetF %I t [ 1 0 0 1 562 550 ] concat %I [ (Professor Smith) ] Text End Begin %I Line %I b 65535 2 0 0 [] 0 SetB %I cfg Red 1 0 0 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 30 433 98 433 Line %I 1 End Begin %I Line %I b 65535 2 0 0 [] 0 SetB %I cfg Red 1 0 0 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 488 360 584 360 Line %I 1 End Begin %I Line %I b 65535 1 0 1 [] 0 SetB %I cfg Black 0 0 0 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 59 432 166 328 Line %I 1 End Begin %I Line %I b 65535 1 0 1 [] 0 SetB %I cfg Black 0 0 0 SetCFg %I cbg White 1 1 1 SetCBg none SetP %I p n %I t [ 1 -0 -0 1 73 175 ] concat %I 516 359 406 241 Line %I 1 End End %I eop showpage end %%EndDocument endTexFig -59 1057 a Fh(Figure)20 b(1:)29 b(T)l(raining)20 b(example)h(of)e(a)h (\\facult)o(y)f(home)h(page.")33 b(The)20 b(task)f(of)g(classifying)i (a)f(w)o(eb)g(page)f(can)h(b)q(e)-59 1113 y(ac)o(hiev)o(ed)h(b)o(y)e (considering)j(just)d(the)h(w)o(ords)f(on)h(the)g(w)o(eb)g(page.)33 b(Alternativ)o(ely)l(,)22 b(the)e(page)g(can)g(b)q(e)g(classi\014ed)-59 1170 y(using)14 b(only)f(the)g(w)o(ords)f(on)h(h)o(yp)q(erlinks)i(that) d(p)q(oin)o(t)h(to)f(the)h(w)o(eb)g(page)g(\(e.g.,)f(\\m)o(y)g (advisor")h(\\Professor)e(Smith"\).)-59 1226 y(When)17 b(examples)g(are)f(describ)q(ed)i(b)o(y)f(suc)o(h)f Ff(r)n(e)n (dundantly)h(su\016cient)g(fe)n(atur)n(es)p Fh(,)f(unlab)q(eled)j(data) d(can)g(b)q(e)h(used)g(to)-59 1282 y(b)q(o)q(ost)e(the)g(accuracy)g (for)g(sup)q(ervised)i(learning.)-59 1420 y(1998].)g(F)l(or)12 b(example,)h(one)g(sp)q(eci\014c)h(task)e(w)o(e)g(ha)o(v)o(e)g (considered)i(is)e(training)h(a)f(system)g(to)g(classify)h(w)o(eb)g (pages)f(in)o(to)-59 1476 y(categories)17 b(suc)o(h)h(as)f(\\studen)o (t)h(home)f(page,")g(\\facult)o(y)h(home)f(page,")h(\\home)f(page)g(of) h(an)f(academic)h(course,")-59 1533 y(etc.)12 1589 y(This)e(w)o(eb)f (page)g(classi\014cation)i(problem)f(can)g(b)q(e)g(form)o(ulated)f(as)g (a)g(t)o(ypical)i(sup)q(ervised)g(learning)f(problem.)-59 1646 y(By)g(\\sup)q(ervised)i(learning)f(problem")g(w)o(e)f(mean)g (that)g(there)g(is)h(some)f(set)g Fe(X)k Fh(of)c(instances,)g(and)h (there)f(is)h(some)-59 1702 y(target)12 b(function)j Fe(f)i Fh(:)c Fe(X)j Fd(!)d Fe(Y)24 b Fh(to)13 b(b)q(e)h(learned,)g (giv)o(en)g(a)g(set)f(of)g(training)h(examples)h(of)e(the)g(form)g Fd(fh)p Fe(x)1698 1709 y Fc(i)1712 1702 y Fe(;)8 b(f)d Fh(\()p Fe(x)1804 1709 y Fc(i)1817 1702 y Fh(\))p Fd(ig)p Fh(.)18 b(In)-59 1759 y(our)13 b(case,)h Fe(X)i Fh(is)e(the)g(set)f(of) g(all)h(w)o(eb)g(pages,)f Fe(Y)24 b Fh(is)14 b(the)f(set)g(of)g(p)q (ossible)i(classi\014cations)g(\(e.g.,)d(\\studen)o(t")h(\\facult)o(y") -59 1815 y(\\neither"\),)20 b(and)g(the)g(function)g Fe(f)25 b Fh(to)19 b(b)q(e)h(learned)h(in)f(the)g(function)g(that)f (maps)g(an)h(arbitrary)f(page)g Fe(x)h Fh(to)f(its)-59 1872 y(correct)d(classi\014cation.)25 b(F)l(or)16 b(the)h Fe(i)p Fh(th)f(lab)q(eled)j(training)e(example,)g Fe(x)1167 1879 y Fc(i)1198 1872 y Fh(is)g(a)f(particular)i(w)o(eb)e(page,)g(and)h Fe(f)5 b Fh(\()p Fe(x)1870 1879 y Fc(i)1884 1872 y Fh(\))16 b(is)-59 1928 y(the)f(correct)g(classi\014cation)i(for)d Fe(x)532 1935 y Fc(i)546 1928 y Fh(,)h(pro)o(vided)h(b)o(y)f(an)g (external)h(teac)o(her.)12 1985 y(As)h(discussed)h(in)g([Cra)o(v)o(en)e (et)h(al.)26 b(1998],)15 b(a)i(n)o(um)o(b)q(er)g(of)g(sup)q(ervised)i (learning)f(algorithms)f(can)g(b)q(e)h(applied)-59 2041 y(to)13 b(the)h(problem)h(of)f(learning)h(to)e(classify)i(w)o(eb)f (pages.)19 b(One)c(common)f(approac)o(h)f(is)i(to)e(\014rst)h(represen) o(t)g(eac)o(h)g(w)o(eb)-59 2097 y(page)f(b)o(y)g(a)g(large)h(feature)f (v)o(ector,)f(where)i(eac)o(h)f(p)q(ossible)i(w)o(ord)e(in)h(the)f (language)g(corresp)q(onds)h(to)f(a)g(feature,)g(and)-59 2154 y(the)h(v)m(alue)i(of)e(the)h(feature)f(is)h(the)f(n)o(um)o(b)q (er)h(of)f(times)h(the)g(w)o(ord)e(o)q(ccurs)i(in)g(the)g(w)o(eb)f (page.)20 b(This)15 b(\\bag)f(of)g(w)o(ords")-59 2210 y(represen)o(tation)g(ob)o(viously)h(ignores)g(information)f(ab)q(out)g (the)g(sequence)i(in)f(whic)o(h)g(w)o(ords)e(o)q(ccur.)20 b(Ho)o(w)o(ev)o(er,)13 b(this)-59 2267 y(represen)o(tation)i(of)g(w)o (eb)g(pages)g(as)f(large)h(feature)g(v)o(ectors)f(enables)j(us)e(to)f (apply)i(standard)f(sup)q(ervised)i(learning)-59 2323 y(algorithms)11 b(for)f(classi\014cation)i(of)e(feature)g(v)o(ectors.) 18 b(F)l(or)10 b(example,)i(if)f(w)o(e)f(train)h(a)f(naiv)o(e)i(Ba)o(y) o(es)e(classi\014er)i([Mitc)o(hell)-59 2380 y(1997])h(on)h(a)g(set)h (of)f(appro)o(ximately)g(4,000)f(lab)q(eled)k(w)o(eb)d(pages,)g(w)o(e)g (ac)o(hiev)o(e)i(accuracies)f(of)f(appro)o(ximately)g(70\045)-59 2436 y(in)i(classifying)h(new)e(w)o(eb)g(pages)g(in)o(to)g(the)h (categories)f(men)o(tioned)h(ab)q(o)o(v)o(e)e([Cra)o(v)o(en)g(et)i (al.,)e(1997].)934 2581 y(2)p eop %%Page: 3 3 3 2 bop -59 297 a Fh(Giv)o(en:)9 353 y Fd(\017)23 b Fh(set)15 b(L)g(of)g(lab)q(eled)i(training)f(examples)9 447 y Fd(\017)23 b Fh(set)15 b(U)g(of)g(unlab)q(eled)i(examples)-59 586 y(Lo)q(op:)9 680 y Fd(\017)23 b Fh(Learn)15 b(h)o(yp)q(erlink-based)j (classi\014er)e(H)f(from)g(L)9 773 y Fd(\017)23 b Fh(Learn)15 b(full-text)h(classi\014er)h(F)e(from)f(L)9 867 y Fd(\017)23 b Fh(Allo)o(w)15 b(H)h(to)e(lab)q(el)j Ff(p)e Fh(p)q(ositiv)o(e)i(and)e Ff(n)g Fh(negativ)o(e)g(examples)h(from)e(U)9 961 y Fd(\017)23 b Fh(Allo)o(w)15 b(F)g(to)g(lab)q(el)i Ff(p)e Fh(p)q(ositiv)o(e)h(and)g Ff(n)e Fh(negativ)o(e)i(examples)g(from)e(U)9 1055 y Fd(\017)23 b Fh(Add)15 b(these)h(self-lab)q(eled)i(examples)e(to)e(L) -59 1190 y(T)l(able)j(1:)k(A)16 b(Cotraining)g(algorithm,)g(for)g (training)g(t)o(w)o(o)f(classi\014ers)i(and)f(using)h(unlab)q(eled)i (data)c(to)g(b)q(o)q(ost)h(their)-59 1247 y(accuracies.)12 1375 y(While)g(w)o(e)e(can)h(ac)o(hiev)o(e)g(reasonable)h(accuracy)e (after)g(training)i(on)f(thousands)f(of)h(hand)g(lab)q(eled)i(w)o(eb)d (pages,)-59 1432 y(in)i(this)g(pap)q(er)h(w)o(e)e(are)g(in)o(terested)h (in)h(the)f(question)g(of)f(ho)o(w)g(unlab)q(eled)j(data)d(can)h(b)q(e) g(useful.)23 b(In)16 b(fact,)f(this)h(is)g(a)-59 1488 y(crucial)e(question)g(in)f(our)g(w)o(eb)g(page)f(classi\014cation)i (task,)f(b)q(ecause)g(it)h(is)f(costly)g(to)f(hand)h(lab)q(el)i (thousands)d(of)h(w)o(eb)-59 1545 y(pages,)i(and)g(b)q(ecause)h(it)g (is)f(v)o(ery)g(easy)g(to)g(obtain)g(h)o(undreds)h(of)f(millions)i(of)e (unlab)q(eled)j(pages)d(from)f(the)i(w)o(eb.)12 1601 y(Ho)o(w)11 b(can)h(w)o(e)g(use)h(unlab)q(eled)i(w)o(eb)d(pages)g(to)f (more)h(accurately)h(learn)g(to)e(classify)i(future)f(pages?)20 b(The)12 b(answ)o(er)-59 1658 y(can)f(b)q(e)h(seen)g(b)o(y)f(\014rst)g (observing)g(that)g(the)g(w)o(eb)g(con)o(tains)g(an)g(in)o(teresting)h (kind)g(of)f(redundan)o(t)h(information)f(ab)q(out)-59 1714 y(eac)o(h)j(w)o(eb)g(page,)g(as)g(sho)o(wn)g(in)h(\014gure)g(1.)k (As)14 b(sho)o(wn)g(in)h(this)g(\014gure,)f(w)o(e)g(could)h(classify)g (the)g(w)o(eb)f(page)g(either)h(b)o(y)-59 1771 y(considering)f(the)e(w) o(ords)g(on)g(the)h(page)f(itself)i(\(as)d(suggested)i(in)g(the)f (paragraph)g(ab)q(o)o(v)o(e\),)g(or)g(w)o(e)g(could)h(classify)h(the) -59 1827 y(page)j(b)o(y)g(ignoring)g(the)g(w)o(ords)g(on)g(the)g(page)g (and)g(instead)g(considering)i(the)e(w)o(ords)f(on)h(h)o(yp)q(erlinks)i (that)d(p)q(oin)o(t)-59 1883 y(to)d(the)h(page.)20 b(In)14 b(man)o(y)g(cases,)g(the)g(w)o(ords)f(on)h(the)g(h)o(yp)q(erlinks)i (will)f(b)q(e)g(su\016cien)o(t)g(to)e(classify)i(the)f(example,)h(and) -59 1940 y(the)f(w)o(ords)f(on)h(the)g(page)g(itself)h(will)h(also)e(b) q(e)h(su\016cien)o(t.)20 b(In)15 b(suc)o(h)f(cases,)g(w)o(e)g(will)h (sa)o(y)f(these)g(t)o(w)o(o)f(sets)h(of)f(features)-59 1996 y({)k(the)g(h)o(yp)q(erlink)i(w)o(ords,)d(and)i(the)f(w)o(eb)g (page)g(w)o(ords)f({)h(are)g Ff(r)n(e)n(dundantly)g(su\016cient)g Fh(to)f(classify)i(the)f(example.)-59 2053 y(W)l(e)d(will)i(see)e(b)q (elo)o(w)h(that)e(in)i(general)f(if)h(the)f(instances)g Fe(X)j Fh(for)d(some)g(sup)q(ervised)h(learning)g(task)f(can)g(b)q(e)g (factored)-59 2109 y(in)o(to)h(sets)g(of)g(redundan)o(tly)h(su\016cien) o(t)g(features,)f(then)g(unlab)q(eled)j(data)d(can)g(b)q(e)h(used)g(to) f(b)q(o)q(ost)g(the)g(accuracy)g(of)-59 2166 y(classi\014ers)h(trained) g(with)g(limited)h(lab)q(eled)g(data.)-59 2309 y Fg(3)69 b(Learning)23 b(with)f(Redundan)n(tly)h(Su\016cien)n(t)e(F)-6 b(eatures)12 2410 y Fh(Ho)o(w)13 b(can)h(w)o(e)g(use)g(unlab)q(eled)j (data)d(to)f(b)q(o)q(ost)h(the)g(learning)h(accuracy)g(for)e(our)h(w)o (eb)g(classi\014cation)i(problem?)934 2581 y(3)p eop %%Page: 4 4 4 3 bop 1 267 1888 2 v 0 324 2 57 v 437 324 V 446 324 V 471 307 a Fh(P)o(age-based)15 b(classi\014er)p 899 324 V 51 w(Hyp)q(erlink-based)j(classi\014er)p 1455 324 V 51 w(Com)o(bined)e(classi\014er)p 1888 324 V 1 325 1888 2 v 0 382 2 57 v 26 365 a(Sup)q(ervised)h(training)p 437 382 V 446 382 V 221 w(12.9)p 899 382 V 423 w(12.4)p 1455 382 V 412 w(11.1)p 1888 382 V 1 384 1888 2 v 0 440 2 57 v 26 423 a(Co-training)p 437 440 V 446 440 V 391 w(6.2)p 899 440 V 435 w(11.6)p 1455 440 V 423 w(5.0)p 1888 440 V 1 442 1888 2 v -59 519 a(T)l(able)i(2:)25 b(Error)17 b(rate)h(in)h(p)q(ercen)o(t)g(for)e(classifying)j(w)o(eb)e (pages)g(as)f(course)i(home)f(pages.)28 b(The)19 b(top)e(ro)o(w)g(sho)o (ws)-59 575 y(errors)d(when)i(training)g(on)f(only)h(the)f(lab)q(eled)i (examples.)k(Bottom)14 b(ro)o(w)g(sho)o(ws)g(errors)h(when)h (co-training,)f(using)-59 632 y(b)q(oth)h(lab)q(eled)i(and)f(unlab)q (eled)h(examples.)24 b(The)16 b(righ)o(tmost)f(column)i(sho)o(ws)e(the) i(classi\014cation)g(accuracy)f(when)-59 688 y(the)h(t)o(w)o(o)g (classi\014ers)h(are)f(pro)o(vided)h(equal)h(v)o(otes)d(on)i(the)f (\014nal)h(classi\014cation.)28 b(The)18 b(results)g(rep)q(orted)f (here)h(are)-59 745 y(the)d(a)o(v)o(erage)f(of)h(\014v)o(e)g(indep)q (enden)o(t)j(runs)d(of)g(the)g(algorithm)h(with)f(di\013eren)o(t)h (random)e(starting)h(examples.)-59 882 y(The)h(k)o(ey)g(idea)g(is)h (that)e(w)o(e)h(will)h(train)f(t)o(w)o(o)f(indep)q(enden)o(t)j (classi\014ers)f(rather)e(than)h(one.)22 b(One)17 b(classi\014er)g (will)g(use)-59 939 y(only)d(the)f(w)o(ords)f(on)i(the)f(w)o(eb)g (page,)g(and)h(the)f(other)g(classi\014er)h(will)h(use)f(only)f(the)h (w)o(ords)e(on)h(the)h(h)o(yp)q(erlinks.)21 b(W)l(e)-59 995 y(b)q(egin)16 b(b)o(y)e(training)h(b)q(oth)g(classi\014ers)g(using) g(whatev)o(er)f(lab)q(eled)j(training)e(examples)g(are)f(a)o(v)m (ailable.)21 b(Presumably)-59 1052 y(this)13 b(will)g(result)g(in)g(t)o (w)o(o)e(classi\014ers)i(that)f(are)g(imp)q(erfect,)h(but)f(b)q(etter)g (than)g(random.)19 b(No)o(w)11 b(w)o(e)h(use)h(the)f(unlab)q(eled)-59 1108 y(data)f(as)g(follo)o(ws:)18 b(eac)o(h)11 b(classi\014er)i(is)e (allo)o(w)o(ed)h(to)f(examine)h(the)g(unlab)q(eled)i(data)c(and)i(to)e (pic)o(k)j(its)e(most)g(con\014den)o(tly)-59 1165 y(predicted)i(p)q (ositiv)o(e)f(and)g(negativ)o(e)f(examples,)h(and)g(add)g(these)f(to)g (the)g(set)g(of)g(lab)q(eled)j(examples.)19 b(In)12 b(other)f(w)o (ords,)-59 1221 y(eac)o(h)16 b(classi\014er)h(is)f(allo)o(w)o(ed)h(to)e (augmen)o(t)g(the)h(p)q(o)q(ol)h(of)e(lab)q(eled)j(examples.)23 b(Both)15 b(classi\014ers)i(are)f(no)o(w)f(retrained)-59 1277 y(on)j(this)g(augmen)o(ted)g(set)g(of)f(lab)q(eled)j(examples,)f (and)f(the)g(pro)q(cess)g(is)h(rep)q(eated)f(as)g(long)g(as)f(desired.) 29 b(T)l(able)19 b(1)-59 1334 y(summarizes)d(this)f(co-training)h (algorithm.)12 1390 y(Wh)o(y)g(should)j(this)e(co-training)i(algorithm) e(lead)h(to)f(more)f(accurate)h(classi\014ers?)29 b(The)17 b(in)o(tuition)i(is)f(that)e(if)-59 1447 y(the)i(h)o(yp)q(erlink)i (classi\014er)f(\014nds)g(an)f(\\easily)h(classi\014ed")g(h)o(yp)q (erlink)h(in)f(the)f(unlab)q(eled)j(data)d(\(e.g.,)f(one)h(that)f(is) -59 1503 y(quite)g(similar)g(to)e(one)h(of)f(the)h(lab)q(eled)i (examples)f(on)f(whic)o(h)g(it)g(w)o(as)f(trained\),)h(the)g(w)o(eb)g (page)g(that)f(it)h(p)q(oin)o(ts)g(to)-59 1560 y(will)g(b)q(e)f(added)h (to)e(the)g(lab)q(eled)j(p)q(o)q(ol)e(of)g(examples)g(as)f(w)o(ell.)21 b(Of)14 b(course)h(just)f(b)q(ecause)i(the)f(h)o(yp)q(erlink)h(happ)q (ened)-59 1616 y(to)h(b)q(e)g(easy)h(to)e(classify)i(do)q(es)g(not)f (mean)g(the)g(w)o(eb)h(page)f(will)i(b)q(e)f(easily)g(classi\014ed)h(b) o(y)e(the)g(other)g(classi\014er.)28 b(If)-59 1673 y(not,)13 b(then)h(the)g(h)o(yp)q(erlink)i(classi\014er)e(has)g(added)g(useful)h (training)f(information)g(to)f(impro)o(v)o(e)h(the)g(other)f (classi\014er.)-59 1729 y(Similarly)l(,)20 b(the)d(w)o(eb)h(page)f (classi\014er)i(can)f(add)f(examples)i(that)d(are)h(easy)h(for)f(it)g (to)g(classify)l(,)i(but)e(that)g(pro)o(vide)-59 1786 y(useful)f(information)g(to)e(impro)o(v)o(e)h(the)h(accuracy)f(of)g (the)g(h)o(yp)q(erlink)i(classi\014er.)12 1842 y(In)c(exp)q(erimen)o (ts,)i(w)o(e)e(ha)o(v)o(e)g(found)h(that)e(this)i(Cotraining)g (algorithm)f(do)q(es)h(impro)o(v)o(e)f(classi\014cation)i(accuracy)-59 1898 y(when)23 b(learning)g(to)f(classify)h(w)o(eb)f(pages.)41 b(In)23 b(one)f(exp)q(erimen)o(t)h([Blum)g(and)g(Mitc)o(hell,)i(1998],) d(summarized)-59 1955 y(in)f(T)l(able)g(2,)f(w)o(e)g(trained)h(a)e (classi\014er)j(to)d(lab)q(el)j(w)o(eb)e(pages)g(as)g(home)g(pages)g (of)f(academic)i(courses.)35 b(In)20 b(this)-59 2011 y(exp)q(erimen)o(t)14 b(w)o(e)f(pro)o(vided)h(just)e(16)h(lab)q(eled)i (examples,)f(and)f(appro)o(ximately)g(800)f(unlab)q(eled)k(pages)d(dra) o(wn)g(from)-59 2068 y(computer)g(science)h(departmen)o(t)e(w)o(eb)h (sites.)19 b(On)14 b(eac)o(h)f(iteration)g(of)f(co-training,)h(eac)o(h) g(classi\014er)h(w)o(as)e(allo)o(w)o(ed)h(to)-59 2124 y(add)i(1)h(new)f(p)q(ositiv)o(e)h(and)g(3)f(new)h(negativ)o(e)f (examples)h(to)f(the)g(p)q(o)q(ol)h(of)f(lab)q(eled)j(examples.)j (After)15 b(30)g(iterations)-59 2181 y(of)g(the)g(cotraining)g (algorithm,)g(the)g(accuracy)g(of)g(the)g(com)o(bined)h(classi\014er)g (w)o(as)f(95\045,)f(compared)h(to)f(89\045)g(when)-59 2237 y(only)f(the)g(lab)q(eled)h(data)e(w)o(as)g(used.)19 b(In)14 b(this)f(case,)f(the)h(impact)g(of)f(cotraining)h(w)o(as)f(to)f (reduce)j(the)e(error)g(b)o(y)h(b)q(etter)-59 2294 y(than)i(a)g(factor) f(of)h(t)o(w)o(o.)934 2581 y(4)p eop %%Page: 5 5 5 4 bop -59 311 a Fg(4)69 b(F)-6 b(ormal)23 b(Results)-59 413 y Fh(Giv)o(en)e(the)f(exp)q(erimen)o(tal)h(evidence)h(that)e (cotraining)h(can)f(b)q(e)h(useful)g(in)g(at)e(least)i(one)f(case,)h (it)f(is)h(useful)g(to)-59 469 y(c)o(haracterize)d(the)f(general)h (problem)g(setting)g(in)g(whic)o(h)g(this)g(kind)g(of)f(algorithm)h (can)f(b)q(ene\014t)i(from)d(unlab)q(eled)-59 526 y(data.)j(Here)d(w)o (e)f(de\014ne)h(a)f(problem)h(setting)f(in)h(whic)o(h)g(w)o(e)f(can)g (pro)o(v)o(e)g(that)f(unlab)q(eled)k(data)d(will)h(b)q(e)g(of)f(help.) 12 582 y(T)l(o)d(de\014ne)i(the)f(general)g(problem)h(setting)f(that)f (captures)h(the)g(essen)o(tial)h(structure)e(of)h(our)f(w)o(eb)h (classi\014cation)-59 639 y(example,)h(recall)g(that)e(earlier)h(w)o(e) g(de\014ned)h(a)e(sup)q(ervised)j(learning)f(problem)f(as)g(one)f (where)h(w)o(e)g(ha)o(v)o(e)f(some)h(set)f(of)-59 695 y(instances)f Fe(X)t Fh(,)f(a)g(target)g(function)h Fe(f)18 b Fh(:)12 b Fe(X)k Fd(!)d Fe(Y)d Fh(,)h(and)g(lab)q(eled)h(examples)f Fd(fh)p Fe(x)1261 702 y Fc(i)1275 695 y Fe(;)d(f)d Fh(\()p Fe(x)1367 702 y Fc(i)1380 695 y Fh(\))p Fd(ig)p Fh(.)17 b(In)11 b(the)g Ff(c)n(otr)n(aining)e Fh(setting,)-59 752 y(w)o(e)18 b(assume)g(suc)o(h)h(a)f(sup)q(ervised)i(learning)f (problem)g(in)g(whic)o(h)g(the)g(instances)g(are)f(dra)o(wn)f(from)h Fe(X)j Fh(according)-59 808 y(to)15 b(some)g(\014xed)h(\(p)q(ossibly)g (unkno)o(wn\))g(probabilit)o(y)g(distribution)h Fd(D)q Fh(.)k(W)l(e)15 b(further)g(assume)h(that)e Fe(X)i Fh(=)e Fe(X)1804 815 y Fb(1)1833 808 y Fd(\002)d Fe(X)1917 815 y Fb(2)1936 808 y Fh(;)-59 864 y(that)i(is,)h(the)g(instances)g(in)h Fe(X)i Fh(can)d(b)q(e)g(factored)f(in)o(to)h(t)o(w)o(o)f(parts)g (\(e.g,.)18 b(the)c(h)o(yp)q(erlink)h(w)o(ords)e(and)h(the)g(w)o(eb)g (page)-59 921 y(w)o(ords\).)23 b(W)l(e)16 b(require)h(also)g(that)f Fe(X)575 928 y Fb(1)610 921 y Fh(and)h Fe(X)738 928 y Fb(2)774 921 y Fh(eac)o(h)f(con)o(tain)h(information)f(su\016cien)o(t)i (to)d(classify)j(the)e(example.)-59 977 y(In)e(other)e(w)o(ords,)h(w)o (e)g(require)h(that)e(there)h(exist)h(some)f(function)h Fe(g)1096 984 y Fb(1)1128 977 y Fh(:)e Fe(X)1191 984 y Fb(1)1223 977 y Fd(!)h Fe(Y)24 b Fh(and)13 b(some)g(function)h Fe(g)1727 984 y Fb(2)1759 977 y Fh(:)e Fe(X)1822 984 y Fb(2)1854 977 y Fd(!)h Fe(Y)-59 1034 y Fh(suc)o(h)j(that)f(for)f(all) j Fe(x)c Fd(2)g Fe(X)q(;)8 b(g)440 1041 y Fb(1)458 1034 y Fh(\()p Fe(x)502 1041 y Fb(1)522 1034 y Fh(\))13 b(=)g Fe(g)623 1041 y Fb(2)642 1034 y Fh(\()p Fe(x)686 1041 y Fb(2)706 1034 y Fh(\))g(=)g Fe(f)5 b Fh(\()p Fe(x)p Fh(\).)20 b(Note)15 b(the)h(learner)g(is)g(not)f(exp)q(ected)i(to)d (kno)o(w)h Fe(f)5 b Fh(,)16 b Fe(g)1817 1041 y Fb(1)1851 1034 y Fh(or)f Fe(g)1929 1041 y Fb(2)-59 1090 y Fh(in)h(adv)m(ance.)k (W)l(e)15 b(simply)h(require)g(that)e(it)h(b)q(e)h(p)q(ossible)g(to)f (express)g Fe(f)20 b Fh(in)15 b(terms)g(of)f Fe(x)1447 1097 y Fb(1)1481 1090 y Fh(and)i(in)f(terms)g(of)f Fe(x)1826 1097 y Fb(2)1860 1090 y Fh(\(i.e.,)-59 1147 y(that)h Fe(X)78 1154 y Fb(1)113 1147 y Fh(and)i Fe(X)241 1154 y Fb(2)276 1147 y Fh(b)q(oth)f(b)q(e)h(su\016cien)o(t)f(to)g(classify)g Fe(X)k Fh(according)c(to)f Fe(f)5 b Fh(\).)22 b(If)16 b(the)g(ab)q(o)o(v)o(e)g(constrain)o(ts)f(on)h Fe(X)1840 1154 y Fb(1)1876 1147 y Fh(and)-59 1203 y Fe(X)-21 1210 y Fb(2)12 1203 y Fh(are)e(satis\014ed,)g(then)g(w)o(e)g(will)h(sa)o(y)e (that)h Fe(X)738 1210 y Fb(1)771 1203 y Fh(and)g Fe(X)896 1210 y Fb(2)929 1203 y Fh(are)f Ff(r)n(e)n(dundantly)i(su\016cient)e Fh(to)g(classify)i Fe(X)i Fh(with)d(resp)q(ect)-59 1260 y(to)h Fe(f)5 b Fh(.)12 1316 y(In)20 b([Blum)g(and)g(Mitc)o(hell)i (1998])c(w)o(e)h(sho)o(w)g(that)h(if)g Fe(f)25 b Fh(:)20 b Fe(X)k Fd(!)c Fe(Y)30 b Fh(is)21 b(P)l(A)o(C)e(learnable)i(from)e (noisy)i(lab)q(eled)-59 1373 y(data,)c Fe(X)98 1380 y Fb(1)135 1373 y Fh(and)g Fe(X)263 1380 y Fb(2)300 1373 y Fh(are)g(redundan)o(tly)i(su\016cien)o(t)f(to)f(classify)h Fe(X)j Fh(with)d(resp)q(ect)g(to)e Fe(f)5 b Fh(,)18 b(and)g(if)g Fe(x)1645 1380 y Fc(B)1672 1385 y Fa(1)1708 1373 y Fh(and)g Fe(x)1825 1380 y Fc(B)1852 1385 y Fa(2)1888 1373 y Fh(are)-59 1429 y(conditionally)h(indep)q(enden)o(t)g(giv)o(en)e Fe(f)5 b Fh(\()p Fe(x)p Fh(\),)16 b(then)h Fe(f)k Fh(can)c(b)q(e)g(P)l (A)o(C)f(learned)i(giv)o(en)f(a)f(w)o(eak)g(initial)j(classi\014er)f (plus)-59 1485 y(only)f Ff(unlab)n(ele)n(d)d Fh(data.)22 b(What)16 b(this)g(means)g(is)h(that)e(if)i(the)f(lab)q(eled)i(data)e (are)g(su\016cien)o(t)g(to)g(train)g(a)g(b)q(etter)g(than)-59 1542 y(random)e(initial)j(classi\014er,)f(if)f Fe(X)512 1549 y Fb(1)546 1542 y Fh(and)g Fe(X)672 1549 y Fb(2)706 1542 y Fh(are)f(distributed)j(indep)q(enden)o(tly)g(for)d(eac)o(h)h (target)f(v)m(alue)i(of)e Fe(f)5 b Fh(,)14 b(and)h(if)-59 1598 y Fe(f)21 b Fh(can)15 b(b)q(e)i(learned)f(accurately)g(giv)o(en)h (an)e(arbitrary)g(n)o(um)o(b)q(er)h(of)f Ff(lab)n(ele)n(d)g Fh(examples,)h(then)g Fe(f)21 b Fh(can)15 b(b)q(e)i(learned)f(to)-59 1655 y(similar)g(accuracy)f(giv)o(en)h(just)e(this)i(w)o(eak)e(initial) j(classi\014er)f(and)g(an)f(arbitrary)f(n)o(um)o(b)q(er)h(of)g Ff(unlab)n(ele)n(d)e Fh(examples.)-59 1711 y(The)f(signi\014cance)i(is) f(that)e(this)i(pro)o(v)o(es)e(that)h(unlab)q(eled)j(examples)d(can)h (substitute)f(for)g(lab)q(eled)i(examples,)f(under)-59 1768 y(the)i(conditions)i(of)d(the)i(theorem.)12 1824 y(In)e([Blum)g(and)f(Mitc)o(hell)i(1998])d(w)o(e)h(also)h(sho)o(w)e (that)h(unlab)q(eled)j(data)d(is)h(of)f(p)q(oten)o(tial)h(v)m(alue)h (under)f(the)f(more)-59 1881 y(general)20 b(setting)f(in)h(whic)o(h)f Fe(X)483 1888 y Fb(1)522 1881 y Fh(and)g Fe(X)652 1888 y Fb(2)691 1881 y Fh(are)f(not)h(indep)q(enden)o(tly)j(distributed.)33 b(There)19 b(w)o(e)g(sho)o(w)g(that)f(if)i(the)-59 1937 y(training)d(data)e(is)i(noise-free,)g(then)g(unless)g Fe(X)758 1944 y Fb(1)794 1937 y Fh(and)f Fe(X)921 1944 y Fb(2)957 1937 y Fh(are)g(deterministically)j(related)e(\(e.g.,)d(one) j(v)m(alue)g(of)f Fe(x)1929 1944 y Fb(1)-59 1994 y Fh(is)g(alw)o(a)o (ys)e(paired)i(with)g(the)f(same)g Fe(x)593 2001 y Fb(2)613 1994 y Fh(\),)f(the)h(unlab)q(eled)j(data)d(can)g(b)q(e)h(of)f(use.)-59 2137 y Fg(5)69 b(Other)22 b(Learning)i(T)-6 b(asks)24 b(with)d(Redundan)n(tly)i(Su\016cien)n(t)e(F)-6 b(eatures)-59 2238 y Fh(Giv)o(en)18 b(the)f(success)i(of)e(cotraining)h(as)f(a)g (metho)q(d)h(for)f(using)h(unlab)q(eled)i(data)d(when)h(learning)h(to)d (classify)j(w)o(eb)-59 2295 y(pages,)12 b(and)g(giv)o(en)g(the)f (formal)h(c)o(haracterization)f(of)h(the)f(class)h(of)g(problems)g(for) f(whic)o(h)h(cotraining)h(can)e(b)q(e)i(of)e(use,)-59 2351 y(it)k(is)g(in)o(teresting)g(to)f(ask)g(what)g(other)g(natural)h (learning)g(problems)g(allo)o(w)g(this)g(approac)o(h.)k(Here)c(w)o(e)f (summarize)-59 2408 y(a)h(n)o(um)o(b)q(er)g(of)g(tasks:)934 2581 y(5)p eop %%Page: 6 6 6 5 bop 9 311 a Fd(\017)23 b Fh(Learning)15 b(to)f(classify)i(noun)f (phrases)f(in)o(to)h(seman)o(tic)g(classes.)20 b(In)15 b([Rilo\013)g(and)g(Jones)g(1999],)e(an)h(approac)o(h)55 368 y(similar)f(to)e(cotraining)i(w)o(as)e(applied)j(to)e(the)g (problem)h(of)f(learning)h(a)f(seman)o(tic)g(lexicon)i(o)o(v)o(er)d (noun)i(phrases)55 424 y(in)j(English.)22 b(In)16 b(this)g(pap)q(er,)g (one)f(task)g(w)o(as)g(to)f(learn)j(to)d(classify)j(noun)f(phrases)f (as)g(p)q(ositiv)o(e)i(or)e(negativ)o(e)55 481 y(examples)j(of)g(lo)q (cations)g(\(e.g.,)f(\\San)h(Sebastian")g(is)g(p)q(ositiv)o(e,)i (\\blue")e(is)g(negativ)o(e\).)28 b(In)19 b(this)f(case)g(eac)o(h)55 537 y(example,)e Fe(x)p Fh(,)f(is)h(a)f(sen)o(tence)h(suc)o(h)g(as)f (\\W)l(e)h(are)f(lo)q(cated)h(in)h(lo)o(v)o(ely)f(Pittsburgh.")21 b(The)15 b(factorization)h(of)f Fe(x)55 594 y Fh(in)o(to)g(t)o(w)o(o)f (redundan)o(tly)i(su\016cien)o(t)g(feature)f(sets)f(is)i(done)g(as)f (follo)o(ws:)k Fe(x)1323 601 y Fb(1)1358 594 y Fh(is)d(the)f(noun)h (phrase)f(itself)h(\(e.g.,)55 650 y(\\Pittsburgh")f(in)i(the)g(ab)q(o)o (v)o(e)f(sen)o(tence\),)g(and)g Fe(x)920 657 y Fb(2)956 650 y Fh(is)h(the)f(linguistic)j(con)o(text)d(in)h(whic)o(h)g(the)f (noun)h(phrase)55 707 y(app)q(ears)c(\(e.g.,)g(\\W)l(e)g(are)h(lo)q (cated)g(in)g(lo)o(v)o(ely)h(|"\).)k(Note)13 b(that)g(in)i(most)e (cases)g(it)h(is)g(p)q(ossible)i(to)d(determine)55 763 y(whether)h(the)h(noun)f(phrase)h(is)g(a)f(lo)q(cation)h(giv)o(en)g (either)g Fe(x)1088 770 y Fb(1)1122 763 y Fh(or)f Fe(x)1203 770 y Fb(2)1223 763 y Fh(.)19 b(Note)14 b(also)h(that)f(the)g(v)m (alues)i(of)e Fe(x)1842 770 y Fb(1)1876 763 y Fh(and)55 819 y Fe(x)81 826 y Fb(2)118 819 y Fh(are)j(distributed)j(fairly)e (indep)q(enden)o(tly)j(\(i.e.,)c(with)h(migh)o(t)g(see)g(the)g(same)f Fe(x)1494 826 y Fb(2)1531 819 y Fh(as)h(ab)q(o)o(v)o(e,)f(but)h(with)g (a)55 876 y(di\013eren)o(t)13 b Fe(x)258 883 y Fb(1)278 876 y Fh(\).)19 b(Starting)13 b(with)g(a)g(list)h(of)f(just)g(10)g(kno) o(wn)g(noun)h(phrases,)f(their)h(system)f(used)h(an)f(algorithm)55 932 y(similar)j(to)f(cotraining)g(to)g(learn)h(dozens)f(of)g (additional)i(lo)q(cations.)9 1026 y Fd(\017)23 b Fh(Learning)12 b(to)f(select)h(w)o(ord)f(sense.)19 b(In)12 b([Y)l(aro)o(wsky)l(,)f (1995],)f(an)i(approac)o(h)f(similar)h(to)f(cotraining)h(w)o(as)f (applied)55 1083 y(to)k(learning)i(to)f(disam)o(biguate)g(w)o(ord)g (senses)g(\(e.g.,)f(to)g(determine)i(whether)f(the)h(w)o(ord)e(\\plan)o (t")h(refers)f(to)55 1139 y(a)g(man)o(ufacturing)g(plan)o(t)h(or)f(to)g (a)g(b)q(otanical)i(plan)o(t\).)k(Here,)15 b(eac)o(h)h(instance)g Fe(x)f Fh(corresp)q(onds)h(to)f(a)g(con)o(text)55 1196 y(con)o(taining)h(the)g(w)o(ord)f(in)i(question)f(\(similar)h(to)e(the) h(Rilo\013)h(and)f(Jones)g(example)h(ab)q(o)o(v)o(e\).)k(Ho)o(w)o(ev)o (er,)15 b(the)55 1252 y(factorization)k(of)h Fe(x)g Fh(is)g(in)o(to)g (more)f(than)h(t)o(w)o(o)e(comp)q(onen)o(ts,)j(with)f(eac)o(h)g(w)o (ord)g(serving)g(as)f(a)h(candidate)55 1308 y(comp)q(onen)o(t,)13 b(and)f(this)i(algorithm)e(seems)h(to)f(not)h(map)f(directly)i(in)o(to) f(cotraining.)20 b(But)12 b(it)h(is)h(quite)f(similar,)55 1365 y(and)i(the)g(problem)h(could)g(b)q(e)g(attac)o(k)o(ed)e(using)i (a)f(cotraining)h(approac)o(h.)9 1459 y Fd(\017)23 b Fh(Learning)18 b(to)f(recognize)h(phonemes)g(in)g(sp)q(eec)o(h.)27 b(In)18 b([de)f(Sa)h(and)f(Ballard,)i(1998],)d(an)h(approac)o(h)g(to)f (fully)55 1515 y(unsup)q(ervised)i(learning)f(w)o(as)e(applied)j(to)e (a)f(problem)i(with)g(nearly)f(redundan)o(tly)h(su\016cien)o(t)g (features.)22 b(In)55 1572 y(this)12 b(case,)h(the)f(task)f(is)i(to)f (learn)g(to)g(classify)h(sp)q(eec)o(h)g(phonemes,)g(based)g(on)f(b)q (oth)g(the)h(audio)f(signal)h(and)g(the)55 1628 y(video)k(signal)h(w)o (atc)o(hing)e(the)h(sp)q(eak)o(er's)g(lips.)26 b(Here)17 b(eac)o(h)f(instance)i Fe(x)f Fh(corresp)q(onds)g(to)f(the)h(full)h (data,)e Fe(x)1929 1635 y Fb(1)55 1685 y Fh(corresp)q(onds)e(to)g(the)g (audio)g(signal,)h(and)g Fe(x)812 1692 y Fb(2)845 1685 y Fh(corresp)q(onds)g(to)e(the)i(video)f(signal.)21 b(Again,)14 b(w)o(e)g(ma)o(y)f(roughly)55 1741 y(assume)k(that)f(either)i(of)f Fe(x)525 1748 y Fb(1)562 1741 y Fh(or)g Fe(x)646 1748 y Fb(2)683 1741 y Fh(is)h(su\016cien)o(t)g(to)f(determine)h(the)f (phoneme,)i(and)e(therefore)g(can)h(train)55 1797 y(t)o(w)o(o)e (classi\014ers)i(on)g(limited)h(lab)q(eled)h(data,)d(and)g(let)h(them)g (train)f(eac)o(h)h(other)f(o)o(v)o(er)f(the)i(unlab)q(eled)i(data.)55 1854 y(In)c(fact,)e(de)i(Sa)f(and)h(Ballard)g(use)g(no)f(lab)q(eled)j (data)c(at)h(all,)h(instead)g(p)q(erforming)g(a)f(kind)h(of)f (clustering)i(in)55 1910 y(whic)o(h)g(the)g(audio)g(signal)h(m)o(ust)e (predict)i(the)f(video,)g(and)g(vice)h(v)o(ersa.)24 b(Their)17 b(system)g(w)o(as)f(able)h(to)f(learn)55 1967 y(clusters)f(corresp)q (onding)i(to)d(the)h(sp)q(ok)o(en)h(phonemes)g(in)g(the)f(data.)9 2061 y Fd(\017)23 b Fh(Ob)s(ject)13 b(recognition)h(in)h(m)o(ultimedia) g(data.)j(Consider)c(the)g(problem)g(of)f(learning)i(to)e(recognize)h (ob)s(jects)f(in)55 2117 y(m)o(ultimedia)g(data,)e(suc)o(h)h(as)f(the)h (con)o(tin)o(uous)f(stream)g(of)g(audio,)i(video,)f(and)g(other)f (sensory)h(input)g(a)f(p)q(erson)55 2174 y(receiv)o(es.)30 b(A)19 b(v)o(ery)f(similar)h(task)f(is)h(to)f(learn)h(to)f(classify)h (television)i(segmen)o(ts,)d(based)h(on)f(the)h(ongoing)55 2230 y(stream)e(of)i(video,)g(audio,)h(and)f(close)g(captioned)g(text.) 30 b(In)19 b(this)g(latter)g(case,)g(consider)g(the)g(problem)g(of)55 2286 y(learning)14 b(to)e(sp)q(ot)h(television)i(segmen)o(ts)d(in)i (whic)o(h)g(Boris)f(Y)l(eltsin)i(app)q(ears.)k(Here)13 b(w)o(e)g(could)h(consider)g(eac)o(h)55 2343 y(instance)20 b Fe(x)f Fh(to)g(b)q(e)h(a)f(snapshot)g(con)o(taining)h(the)f(audio,)i (video)f(and)f(text)g(at)g(a)g(particular)h(time.)33 b(Here)55 2399 y Fe(x)81 2406 y Fb(1)119 2399 y Fh(could)20 b(b)q(e)g(the)f(audio,)h Fe(x)559 2406 y Fb(2)597 2399 y Fh(the)f(video,)h(and)f Fe(x)934 2406 y Fb(3)973 2399 y Fh(the)g(text.)30 b(Note)19 b(that)f(in)i(some)e(cases)h(w)o(e)g (migh)o(t)f(see)55 2456 y(\\easy")c(example)h(of)g(video)g(\(a)f(full)j (face)d(of)h(Y)l(eltsin\),)h(or)e(an)h(easy)f(audio)i(\(his)f(v)o(oice) g(without)g(bac)o(kground)934 2581 y(6)p eop %%Page: 7 7 7 6 bop 55 311 a Fh(noise\),)15 b(or)f(an)h(easy)g(text)f(\(the)h(w)o (ord)f(Y)l(eltsin\).)21 b(Th)o(us,)15 b(w)o(e)f(could)i(exp)q(ect)g (cotraining)f(to)g(pro)o(vide)g(a)g(useful)55 368 y(approac)o(h)e(to)g (using)i(unlab)q(eled)i(data)c(to)g(train)h(classi\014ers.)21 b(T)l(o)13 b(m)o(y)h(kno)o(wledge,)g(this)g(exp)q(erimen)o(t)h(has)f (not)55 424 y(b)q(een)i(attempted.)-59 567 y Fg(6)69 b(Discussion)22 b(and)h(Conclusions)-59 669 y Fh(W)l(e)15 b(ha)o(v)o(e)f(describ)q(ed)j(a)d(class)h(of)g(sup)q(ervised)h (learning)g(problems)f(for)g(whic)o(h)g(unlab)q(eled)j(data)c(can)h(b)q (e)g(pro)o(v)o(en)f(to)-59 725 y(impro)o(v)o(e)i(learning)h(accuracy)l (.)22 b(The)16 b(k)o(ey)f(de\014ning)j(features)d(of)h(this)g(problem)g (class)h(are)e(that)g(\(1\))g(the)h(instances)-59 782 y Fe(X)g Fh(can)c(b)q(e)h(factored)f(in)o(to)h(t)o(w)o(o)e(or)h(more)g (comp)q(onen)o(ts)h Fe(X)930 789 y Fb(1)949 782 y Fh(,)g Fe(X)1013 789 y Fb(2)1040 782 y Fe(:)8 b(:)g(:)n Fh(,)k(whic)o(h)i(are) e Ff(r)n(e)n(dundantly)h(su\016cient)f Fh(to)g(classify)-59 838 y(the)j(example,)h(and)g(\(2\))f(these)g(comp)q(onen)o(ts)h(co)o(v) m(ary)l(,)f(so)g(that)f(a)h(particular)i Fe(x)1321 845 y Fc(i)1350 838 y Fh(do)q(es)f(not)f(alw)o(a)o(ys)g(co-o)q(ccur)g(with) -59 895 y(the)g(same)g Fe(x)159 902 y Fc(j)177 895 y Fh(..)12 951 y(W)l(e)f(ha)o(v)o(e)g(describ)q(ed)j(the)d(cotraining)i (algorithm)e(that)g(uses)h(unlab)q(eled)i(data)d(for)g(suc)o(h)h (problems,)f(and)h(ha)o(v)o(e)-59 1008 y(presen)o(ted)21 b(exp)q(erimen)o(tal)g(and)g(theoretical)g(results)g(sho)o(wing)f(that) f(in)j(these)e(problems)h(unlab)q(eled)i(data)d(can)-59 1064 y(indeed)d(b)q(e)f(useful)g(to)f(impro)o(v)o(e)g(accuracy)g(for)g (sup)q(ervised)h(learning.)12 1121 y(The)e(connection)h(to)e(h)o(uman)h (and)g(animal)h(learning)g(app)q(ears)f(to)f(b)q(e)h(a)g(ric)o(h)g (area)g(for)f(future)h(study)l(.)20 b(Humans)-59 1177 y(and)12 b(other)g(animals)h(ha)o(v)o(e)f(a)f(ric)o(h)i(set)f(of)g (sensory)g(input,)h(whic)o(h)g(includes)h(redundan)o(tly)f(su\016cien)o (t)g(data)f(for)f(man)o(y)-59 1233 y(tasks.)19 b(F)l(or)14 b(example,)h(smell,)h(vision,)g(sound)f(can)g(all)h(b)q(e)f(useful)h (in)g(trying)f(to)f(classify)i(whether)f(fo)q(o)q(d)g(is)g(nearb)o(y)l (,)-59 1290 y(and)h(in)h(man)o(y)f(cases)g(the)g(data)f(from)g(just)h (one)g(of)g(these)g(c)o(hannels)h(is)g(su\016cien)o(t.)23 b(de)16 b(Sa)g(and)h(Ballard's)f(success)-59 1346 y(with)h(a)g (cotraining-lik)o(e)i(algorithm)e(for)f(learning)j(to)d(classify)i (phonemes)g(from)e(video)i(and)f(audio)g({)g(similar)h(to)-59 1403 y(that)g(observ)o(ed)g(b)o(y)g(h)o(umans)g({)g(is)h(also)f (suggestiv)o(e.)30 b(P)o(erhaps)18 b(p)q(eople)h(rely)g(less)g(than)f (w)o(e)g(susp)q(ect)h(on)g(lab)q(eled)-59 1459 y(data)14 b(to)h(ac)o(hiev)o(e)h(successful)g(learning)h(for)d(v)m(arious)i (classi\014cation)h(tasks.)-59 1603 y Fg(7)69 b(Ac)n(kno)n(wledgemen)n (ts)-59 1704 y Fh(Thanks)19 b(to)g(m)o(y)g(collab)q(orator)g(on)g(this) h(w)o(ork,)e(Avrim)i(Blum,)h(for)d(man)o(y)h(in)o(teresting)h (discussions)h(and)e(ideas.)-59 1760 y(Sebastian)i(Thrun,)h(Rosie)f (Jones,)h(Andrew)f(McCallum,)h(and)e(Kamal)h(Nigam)g(ha)o(v)o(e)f(also) g(con)o(tributed)h(ideas)-59 1817 y(and)g(suggestions)g(along)f(the)h (w)o(a)o(y)l(.)36 b(This)21 b(w)o(ork)f(has)h(b)q(een)h(supp)q(orted)f (b)o(y)g(Darpa)f(under)h(researc)o(h)g(con)o(tract)-59 1873 y(F30602-97-1-0215.)-59 2017 y Fg(8)69 b(References)-59 2178 y Fh([Anderson)18 b(et)f(al.)27 b(1995])16 b(Pro)q(duction)j (system)e(mo)q(dels)h(of)f(complex)i(cognition.)28 b(In)18 b(Pro)q(ceedings)g(of)g(the)f(Sev-)-59 2234 y(en)o(teen)o(th)j(Ann)o (ual)i(Conference)f(of)f(the)g(Cognitiv)o(e)h(Science)h(So)q(ciet)o(y)f ([pp.)36 b(9-12].)e(Hillsdale,)24 b(NJ:)c(La)o(wrence)-59 2291 y(Erlbaum)c(Asso)q(ciates.)12 2377 y([Blum)i(and)g(Mitc)o(hell,)i (1998])d(Com)o(bining)i(Lab)q(eled)h(and)e(Unlab)q(eled)j(Data)c(with)h (Co-T)l(raining,)h(COL)l(T98.)934 2581 y(7)p eop %%Page: 8 8 8 7 bop -59 311 a Fh(Av)m(ailable)17 b(at)e(h)o(ttp://www.cs.cm)o (u.edu/)p Fd(\030)p Fh(w)o(ebkb.)12 398 y([Cra)o(v)o(en)c(et)i(al.,)g (1998])f(Learning)i(to)e(extract)g(sym)o(b)q(olic)i(kno)o(wledge)g (from)e(the)h(w)o(orld)g(wide)h(w)o(eb.)19 b(in)14 b Ff(Pr)n(o)n(c)n(e)n(e)n(d-)-59 454 y(ings)d(of)g(the)h(15th)h(National) e(Confer)n(enc)n(e)f(on)h(A)o(rti\014cial)g(Intel)r(ligenc)n(e)e (\(AAAI-98\))p Fh(.)17 b(Av)m(ailable)12 b(at)e(h)o(ttp://www.cs.cm)o (u.edu/)p Fd(\030)p Fh(w)o(ebkb.)12 541 y([de)15 b(Sa,)g(1994])e (Learning)j(classi\014cation)h(with)f(unlab)q(eled)h(data,)e(NIPS-6,)g (1994.)12 627 y([de)f(Sa)h(and)g(Ballard,)g(1998])e(Category)g (learning)j(through)e(m)o(ulti-mo)q(dalit)o(y)i(sensing,)g Ff(Neur)n(al)f(Computation)-59 684 y Fh(10\(5\),)e(1998.)12 770 y([Rilo\013)i(and)g(Jones,)g(1999])f(Learning)i(dictionaries)h(for) d(information)h(extraction)g(b)o(y)g(m)o(ulti-lev)o(el)i(b)q(o)q (otstrap-)-59 827 y(ping,)f Ff(AAAI99)p Fh(,)e(to)g(app)q(ear.)20 b(Av)m(ailable)e(at)c(h)o(ttp://www.cs.cm)o(u.edu/)p Fd(\030)p Fh(w)o(ebkb.)12 913 y([Mitc)o(hell)d(1997].)17 b Ff(Machine)11 b(le)n(arning.)17 b Fh(New)10 b(Y)l(ork:)17 b(McGra)o(w)9 b(Hill,)j(1997.)18 b(See)10 b(h)o(ttp://www.cs.cm)o (u.edu/)p Fd(\030)p Fh(tom/mlb)q(o)q(ok.h)o(tml.)12 999 y([New)o(ell)16 b(1990].)i Ff(Uni\014e)n(d)d(the)n(ories)h(of)g(c)n(o)n (gnition.)j Fh(Cam)o(bridge,)c(MA:)g(Harv)m(ard)g(Univ)o(ersit)o(y)g (Press,)g(1990.)12 1086 y([Y)l(aro)o(wsky)l(,)f(1995])h(Unsup)q (ervised)j(w)o(ord)e(sense)h(disam)o(biguation)g(riv)m(aling)h(sup)q (ervised)g(metho)q(ds.)23 b Ff(Pr)n(o)n(c)n(e)n(e)n(d-)-59 1142 y(ings)15 b(of)i(the)f(33r)n(d)h(A)o(nnual)e(Me)n(eting)g(of)i (the)f(A)o(CL)p Fh(,)e(pp.)20 b(189-196.)934 2581 y(8)p eop %%Trailer end userdict /end-hook known{end-hook}if %%EOF