%! %%BoundingBox: (atend) %%Pages: (atend) %%DocumentFonts: (atend) %%EndComments % % FrameMaker PostScript Prolog 3.0, for use with FrameMaker 3.0 % Copyright (c) 1986,87,89,90,91 by Frame Technology Corporation. % All rights reserved. % % Known Problems: % Due to bugs in Transcript, the 'PS-Adobe-' is omitted from line 1 /FMversion (3.0) def % Set up Color vs. Black-and-White /FMPrintInColor systemdict /colorimage known systemdict /currentcolortransfer known or def % Uncomment this line to force b&w on color printer % /FMPrintInColor false def /FrameDict 195 dict def systemdict /errordict known not {/errordict 10 dict def errordict /rangecheck {stop} put} if % The readline in 23.0 doesn't recognize cr's as nl's on AppleTalk FrameDict /tmprangecheck errordict /rangecheck get put errordict /rangecheck {FrameDict /bug true put} put FrameDict /bug false put mark % Some PS machines read past the CR, so keep the following 3 lines together! currentfile 5 string readline 00 0000000000 cleartomark errordict /rangecheck FrameDict /tmprangecheck get put FrameDict /bug get { /readline { /gstring exch def /gfile exch def /gindex 0 def { gfile read pop dup 10 eq {exit} if dup 13 eq {exit} if gstring exch gindex exch put /gindex gindex 1 add def } loop pop gstring 0 gindex getinterval true } def } if /FMVERSION { FMversion ne { /Times-Roman findfont 18 scalefont setfont 100 100 moveto (FrameMaker version does not match postscript_prolog!) dup = show showpage } if } def /FMLOCAL { FrameDict begin 0 def end } def /gstring FMLOCAL /gfile FMLOCAL /gindex FMLOCAL /orgxfer FMLOCAL /orgproc FMLOCAL /organgle FMLOCAL /orgfreq FMLOCAL /yscale FMLOCAL /xscale FMLOCAL /manualfeed FMLOCAL /paperheight FMLOCAL /paperwidth FMLOCAL /FMDOCUMENT { array /FMfonts exch def /#copies exch def FrameDict begin 0 ne dup {setmanualfeed} if /manualfeed exch def /paperheight exch def /paperwidth exch def /yscale exch def /xscale exch def currenttransfer cvlit /orgxfer exch def currentscreen cvlit /orgproc exch def /organgle exch def /orgfreq exch def setpapername manualfeed {true} {papersize} ifelse {manualpapersize} {false} ifelse {desperatepapersize} if end } def /pagesave FMLOCAL /orgmatrix FMLOCAL /landscape FMLOCAL /FMBEGINPAGE { FrameDict begin /pagesave save def 3.86 setmiterlimit /landscape exch 0 ne def landscape { 90 rotate 0 exch neg translate pop } {pop pop} ifelse xscale yscale scale /orgmatrix matrix def gsave } def /FMENDPAGE { grestore pagesave restore end showpage } def /FMFONTDEFINE { FrameDict begin findfont ReEncode 1 index exch definefont FMfonts 3 1 roll put end } def /FMFILLS { FrameDict begin array /fillvals exch def end } def /FMFILL { FrameDict begin fillvals 3 1 roll put end } def /FMNORMALIZEGRAPHICS { newpath 0.0 0.0 moveto 1 setlinewidth 0 setlinecap 0 0 0 sethsbcolor 0 setgray } bind def /fx FMLOCAL /fy FMLOCAL /fh FMLOCAL /fw FMLOCAL /llx FMLOCAL /lly FMLOCAL /urx FMLOCAL /ury FMLOCAL /FMBEGINEPSF { end /FMEPSF save def /showpage {} def FMNORMALIZEGRAPHICS [/fy /fx /fh /fw /ury /urx /lly /llx] {exch def} forall fx fy translate rotate fw urx llx sub div fh ury lly sub div scale llx neg lly neg translate } bind def /FMENDEPSF { FMEPSF restore FrameDict begin } bind def FrameDict begin /setmanualfeed { %%BeginFeature *ManualFeed True statusdict /manualfeed true put %%EndFeature } def /max {2 copy lt {exch} if pop} bind def /min {2 copy gt {exch} if pop} bind def /inch {72 mul} def /pagedimen { paperheight sub abs 16 lt exch paperwidth sub abs 16 lt and {/papername exch def} {pop} ifelse } def /papersizedict FMLOCAL /setpapername { /papersizedict 14 dict def papersizedict begin /papername /unknown def /Letter 8.5 inch 11.0 inch pagedimen /LetterSmall 7.68 inch 10.16 inch pagedimen /Tabloid 11.0 inch 17.0 inch pagedimen /Ledger 17.0 inch 11.0 inch pagedimen /Legal 8.5 inch 14.0 inch pagedimen /Statement 5.5 inch 8.5 inch pagedimen /Executive 7.5 inch 10.0 inch pagedimen /A3 11.69 inch 16.5 inch pagedimen /A4 8.26 inch 11.69 inch pagedimen /A4Small 7.47 inch 10.85 inch pagedimen /B4 10.125 inch 14.33 inch pagedimen /B5 7.16 inch 10.125 inch pagedimen end } def /papersize { papersizedict begin /Letter {lettertray letter} def /LetterSmall {lettertray lettersmall} def /Tabloid {11x17tray 11x17} def /Ledger {ledgertray ledger} def /Legal {legaltray legal} def /Statement {statementtray statement} def /Executive {executivetray executive} def /A3 {a3tray a3} def /A4 {a4tray a4} def /A4Small {a4tray a4small} def /B4 {b4tray b4} def /B5 {b5tray b5} def /unknown {unknown} def papersizedict dup papername known {papername} {/unknown} ifelse get end /FMdicttop countdictstack 1 add def statusdict begin stopped end countdictstack -1 FMdicttop {pop end} for } def /manualpapersize { papersizedict begin /Letter {letter} def /LetterSmall {lettersmall} def /Tabloid {11x17} def /Ledger {ledger} def /Legal {legal} def /Statement {statement} def /Executive {executive} def /A3 {a3} def /A4 {a4} def /A4Small {a4small} def /B4 {b4} def /B5 {b5} def /unknown {unknown} def papersizedict dup papername known {papername} {/unknown} ifelse get end stopped } def /desperatepapersize { statusdict /setpageparams known { paperwidth paperheight 0 1 statusdict begin {setpageparams} stopped pop end } if } def /savematrix { orgmatrix currentmatrix pop } bind def /restorematrix { orgmatrix setmatrix } bind def /dmatrix matrix def /dpi 72 0 dmatrix defaultmatrix dtransform dup mul exch dup mul add sqrt def /freq dpi 18.75 div 8 div round dup 0 eq {pop 1} if 8 mul dpi exch div def /sangle 1 0 dmatrix defaultmatrix dtransform exch atan def /DiacriticEncoding [ /.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 /quotesingle /parenleft /parenright /asterisk /plus /comma /hyphen /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 /grave /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 /Adieresis /Aring /Ccedilla /Eacute /Ntilde /Odieresis /Udieresis /aacute /agrave /acircumflex /adieresis /atilde /aring /ccedilla /eacute /egrave /ecircumflex /edieresis /iacute /igrave /icircumflex /idieresis /ntilde /oacute /ograve /ocircumflex /odieresis /otilde /uacute /ugrave /ucircumflex /udieresis /dagger /.notdef /cent /sterling /section /bullet /paragraph /germandbls /registered /copyright /trademark /acute /dieresis /.notdef /AE /Oslash /.notdef /.notdef /.notdef /.notdef /yen /.notdef /.notdef /.notdef /.notdef /.notdef /.notdef /ordfeminine /ordmasculine /.notdef /ae /oslash /questiondown /exclamdown /logicalnot /.notdef /florin /.notdef /.notdef /guillemotleft /guillemotright /ellipsis /.notdef /Agrave /Atilde /Otilde /OE /oe /endash /emdash /quotedblleft /quotedblright /quoteleft /quoteright /.notdef /.notdef /ydieresis /Ydieresis /fraction /currency /guilsinglleft /guilsinglright /fi /fl /daggerdbl /periodcentered /quotesinglbase /quotedblbase /perthousand /Acircumflex /Ecircumflex /Aacute /Edieresis /Egrave /Iacute /Icircumflex /Idieresis /Igrave /Oacute /Ocircumflex /.notdef /Ograve /Uacute /Ucircumflex /Ugrave /dotlessi /circumflex /tilde /macron /breve /dotaccent /ring /cedilla /hungarumlaut /ogonek /caron ] def /ReEncode { dup length dict begin { 1 index /FID ne {def} {pop pop} ifelse } forall 0 eq {/Encoding DiacriticEncoding def} if currentdict end } bind def /graymode true def /bwidth FMLOCAL /bpside FMLOCAL /bstring FMLOCAL /onbits FMLOCAL /offbits FMLOCAL /xindex FMLOCAL /yindex FMLOCAL /x FMLOCAL /y FMLOCAL /setpattern { /bwidth exch def /bpside exch def /bstring exch def /onbits 0 def /offbits 0 def freq sangle landscape {90 add} if {/y exch def /x exch def /xindex x 1 add 2 div bpside mul cvi def /yindex y 1 add 2 div bpside mul cvi def bstring yindex bwidth mul xindex 8 idiv add get 1 7 xindex 8 mod sub bitshift and 0 ne {/onbits onbits 1 add def 1} {/offbits offbits 1 add def 0} ifelse } setscreen {} settransfer offbits offbits onbits add div FMsetgray /graymode false def } bind def /grayness { FMsetgray graymode not { /graymode true def orgxfer cvx settransfer orgfreq organgle orgproc cvx setscreen } if } bind def /HUE FMLOCAL /SAT FMLOCAL /BRIGHT FMLOCAL /Colors FMLOCAL FMPrintInColor { /HUE 0 def /SAT 0 def /BRIGHT 0 def % array of arrays Hue and Sat values for the separations [HUE BRIGHT] /Colors [[0 0 ] % black [0 0 ] % white [0.00 1.0] % red [0.37 1.0] % green [0.60 1.0] % blue [0.50 1.0] % cyan [0.83 1.0] % magenta [0.16 1.0] % comment / yellow ] def /BEGINBITMAPCOLOR { BITMAPCOLOR} def /BEGINBITMAPCOLORc { BITMAPCOLORc} def /BEGINBITMAPTRUECOLOR { BITMAPTRUECOLOR } def /BEGINBITMAPTRUECOLORc { BITMAPTRUECOLORc } def /K { Colors exch get dup 0 get /HUE exch store 1 get /BRIGHT exch store HUE 0 eq BRIGHT 0 eq and {1.0 SAT sub setgray} {HUE SAT BRIGHT sethsbcolor} ifelse } def /FMsetgray { /SAT exch 1.0 exch sub store HUE 0 eq BRIGHT 0 eq and {1.0 SAT sub setgray} {HUE SAT BRIGHT sethsbcolor} ifelse } bind def } { /BEGINBITMAPCOLOR { BITMAPGRAY} def /BEGINBITMAPCOLORc { BITMAPGRAYc} def /BEGINBITMAPTRUECOLOR { BITMAPTRUEGRAY } def /BEGINBITMAPTRUECOLORc { BITMAPTRUEGRAYc } def /FMsetgray {setgray} bind def /K { pop } def } ifelse /normalize { transform round exch round exch itransform } bind def /dnormalize { dtransform round exch round exch idtransform } bind def /lnormalize { 0 dtransform exch cvi 2 idiv 2 mul 1 add exch idtransform pop } bind def /H { lnormalize setlinewidth } bind def /Z { setlinecap } bind def /fillvals FMLOCAL /X { fillvals exch get dup type /stringtype eq {8 1 setpattern} {grayness} ifelse } bind def /V { gsave eofill grestore } bind def /N { stroke } bind def /M {newpath moveto} bind def /E {lineto} bind def /D {curveto} bind def /O {closepath} bind def /n FMLOCAL /L { /n exch def newpath normalize moveto 2 1 n {pop normalize lineto} for } bind def /Y { L closepath } bind def /x1 FMLOCAL /x2 FMLOCAL /y1 FMLOCAL /y2 FMLOCAL /rad FMLOCAL /R { /y2 exch def /x2 exch def /y1 exch def /x1 exch def x1 y1 x2 y1 x2 y2 x1 y2 4 Y } bind def /RR { /rad exch def normalize /y2 exch def /x2 exch def normalize /y1 exch def /x1 exch def newpath x1 y1 rad add moveto x1 y2 x2 y2 rad arcto x2 y2 x2 y1 rad arcto x2 y1 x1 y1 rad arcto x1 y1 x1 y2 rad arcto closepath 16 {pop} repeat } bind def /C { grestore gsave R clip } bind def /FMpointsize FMLOCAL /F { FMfonts exch get FMpointsize scalefont setfont } bind def /Q { /FMpointsize exch def F } bind def /T { moveto show } bind def /RF { rotate 0 ne {-1 1 scale} if } bind def /TF { gsave moveto RF show grestore } bind def /P { moveto 0 32 3 2 roll widthshow } bind def /PF { gsave moveto RF 0 32 3 2 roll widthshow grestore } bind def /S { moveto 0 exch ashow } bind def /SF { gsave moveto RF 0 exch ashow grestore } bind def /B { moveto 0 32 4 2 roll 0 exch awidthshow } bind def /BF { gsave moveto RF 0 32 4 2 roll 0 exch awidthshow grestore } bind def /G { gsave newpath normalize translate 0.0 0.0 moveto dnormalize scale 0.0 0.0 1.0 5 3 roll arc closepath fill grestore } bind def /A { gsave savematrix newpath 2 index 2 div add exch 3 index 2 div sub exch normalize 2 index 2 div sub exch 3 index 2 div add exch translate scale 0.0 0.0 1.0 5 3 roll arc restorematrix stroke grestore } bind def /x FMLOCAL /y FMLOCAL /w FMLOCAL /h FMLOCAL /xx FMLOCAL /yy FMLOCAL /ww FMLOCAL /hh FMLOCAL /FMsaveobject FMLOCAL /FMoptop FMLOCAL /FMdicttop FMLOCAL /BEGINPRINTCODE { /FMdicttop countdictstack 1 add def /FMoptop count 4 sub def /FMsaveobject save def userdict begin /showpage {} def FMNORMALIZEGRAPHICS 3 index neg 3 index neg translate } bind def /ENDPRINTCODE { count -1 FMoptop {pop pop} for countdictstack -1 FMdicttop {pop end} for FMsaveobject restore } bind def /gn { 0 { 46 mul cf read pop 32 sub dup 46 lt {exit} if 46 sub add } loop add } bind def /str FMLOCAL /cfs { /str sl string def 0 1 sl 1 sub {str exch val put} for str def } bind def /ic [ 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0223 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0223 0 {0 hx} {1 hx} {2 hx} {3 hx} {4 hx} {5 hx} {6 hx} {7 hx} {8 hx} {9 hx} {10 hx} {11 hx} {12 hx} {13 hx} {14 hx} {15 hx} {16 hx} {17 hx} {18 hx} {19 hx} {gn hx} {0} {1} {2} {3} {4} {5} {6} {7} {8} {9} {10} {11} {12} {13} {14} {15} {16} {17} {18} {19} {gn} {0 wh} {1 wh} {2 wh} {3 wh} {4 wh} {5 wh} {6 wh} {7 wh} {8 wh} {9 wh} {10 wh} {11 wh} {12 wh} {13 wh} {14 wh} {gn wh} {0 bl} {1 bl} {2 bl} {3 bl} {4 bl} {5 bl} {6 bl} {7 bl} {8 bl} {9 bl} {10 bl} {11 bl} {12 bl} {13 bl} {14 bl} {gn bl} {0 fl} {1 fl} {2 fl} {3 fl} {4 fl} {5 fl} {6 fl} {7 fl} {8 fl} {9 fl} {10 fl} {11 fl} {12 fl} {13 fl} {14 fl} {gn fl} ] def /sl FMLOCAL /val FMLOCAL /ws FMLOCAL /im FMLOCAL /bs FMLOCAL /cs FMLOCAL /len FMLOCAL /pos FMLOCAL /ms { /sl exch def /val 255 def /ws cfs /im cfs /val 0 def /bs cfs /cs cfs } bind def 400 ms /ip { is 0 cf cs readline pop { ic exch get exec add } forall pop } bind def /wh { /len exch def /pos exch def ws 0 len getinterval im pos len getinterval copy pop pos len } bind def /bl { /len exch def /pos exch def bs 0 len getinterval im pos len getinterval copy pop pos len } bind def /s1 1 string def /fl { /len exch def /pos exch def /val cf s1 readhexstring pop 0 get def pos 1 pos len add 1 sub {im exch val put} for pos len } bind def /hx { 3 copy getinterval cf exch readhexstring pop pop } bind def /h FMLOCAL /w FMLOCAL /d FMLOCAL /lb FMLOCAL /bitmapsave FMLOCAL /is FMLOCAL /cf FMLOCAL /wbytes { dup 8 eq {pop} {1 eq {7 add 8 idiv} {3 add 4 idiv} ifelse} ifelse } bind def /BEGINBITMAPBWc { 1 {} COMMONBITMAPc } bind def /BEGINBITMAPGRAYc { 8 {} COMMONBITMAPc } bind def /BEGINBITMAP2BITc { 2 {} COMMONBITMAPc } bind def /COMMONBITMAPc { /r exch def /d exch def gsave translate rotate scale /h exch def /w exch def /lb w d wbytes def sl lb lt {lb ms} if /bitmapsave save def r /is im 0 lb getinterval def ws 0 lb getinterval is copy pop /cf currentfile def w h d [w 0 0 h neg 0 h] {ip} image bitmapsave restore grestore } bind def /BEGINBITMAPBW { 1 {} COMMONBITMAP } bind def /BEGINBITMAPGRAY { 8 {} COMMONBITMAP } bind def /BEGINBITMAP2BIT { 2 {} COMMONBITMAP } bind def /COMMONBITMAP { /r exch def /d exch def gsave translate rotate scale /h exch def /w exch def /bitmapsave save def r /is w d wbytes string def /cf currentfile def w h d [w 0 0 h neg 0 h] {cf is readhexstring pop} image bitmapsave restore grestore } bind def /proc1 FMLOCAL /proc2 FMLOCAL /newproc FMLOCAL /Fmcc { /proc2 exch cvlit def /proc1 exch cvlit def /newproc proc1 length proc2 length add array def newproc 0 proc1 putinterval newproc proc1 length proc2 putinterval newproc cvx } bind def /ngrayt 256 array def /nredt 256 array def /nbluet 256 array def /ngreent 256 array def /gryt FMLOCAL /blut FMLOCAL /grnt FMLOCAL /redt FMLOCAL /indx FMLOCAL /cynu FMLOCAL /magu FMLOCAL /yelu FMLOCAL /k FMLOCAL /u FMLOCAL /colorsetup { currentcolortransfer /gryt exch def /blut exch def /grnt exch def /redt exch def 0 1 255 { /indx exch def /cynu 1 red indx get 255 div sub def /magu 1 green indx get 255 div sub def /yelu 1 blue indx get 255 div sub def /k cynu magu min yelu min def /u k currentundercolorremoval exec def nredt indx 1 0 cynu u sub max sub redt exec put ngreent indx 1 0 magu u sub max sub grnt exec put nbluet indx 1 0 yelu u sub max sub blut exec put ngrayt indx 1 k currentblackgeneration exec sub gryt exec put } for {255 mul cvi nredt exch get} {255 mul cvi ngreent exch get} {255 mul cvi nbluet exch get} {255 mul cvi ngrayt exch get} setcolortransfer {pop 0} setundercolorremoval {} setblackgeneration } bind def /tran FMLOCAL /fakecolorsetup { /tran 256 string def 0 1 255 {/indx exch def tran indx red indx get 77 mul green indx get 151 mul blue indx get 28 mul add add 256 idiv put} for currenttransfer {255 mul cvi tran exch get 255.0 div} exch Fmcc settransfer } bind def /BITMAPCOLOR { /d 8 def gsave translate rotate scale /h exch def /w exch def /bitmapsave save def colorsetup /is w d wbytes string def /cf currentfile def w h d [w 0 0 h neg 0 h] {cf is readhexstring pop} {is} {is} true 3 colorimage bitmapsave restore grestore } bind def /BITMAPCOLORc { /d 8 def gsave translate rotate scale /h exch def /w exch def /lb w d wbytes def sl lb lt {lb ms} if /bitmapsave save def colorsetup /is im 0 lb getinterval def ws 0 lb getinterval is copy pop /cf currentfile def w h d [w 0 0 h neg 0 h] {ip} {is} {is} true 3 colorimage bitmapsave restore grestore } bind def /BITMAPTRUECOLORc { gsave translate rotate scale /h exch def /w exch def /bitmapsave save def /is w string def ws 0 w getinterval is copy pop /cf currentfile def w h 8 [w 0 0 h neg 0 h] {ip} {gip} {bip} true 3 colorimage bitmapsave restore grestore } bind def /BITMAPTRUECOLOR { gsave translate rotate scale /h exch def /w exch def /bitmapsave save def /is w string def /gis w string def /bis w string def /cf currentfile def w h 8 [w 0 0 h neg 0 h] { cf is readhexstring pop } { cf gis readhexstring pop } { cf bis readhexstring pop } true 3 colorimage bitmapsave restore grestore } bind def /BITMAPTRUEGRAYc { gsave translate rotate scale /h exch def /w exch def /bitmapsave save def /is w string def ws 0 w getinterval is copy pop /cf currentfile def w h 8 [w 0 0 h neg 0 h] {ip gip bip w gray} image bitmapsave restore grestore } bind def /ww FMLOCAL /r FMLOCAL /g FMLOCAL /b FMLOCAL /i FMLOCAL /gray { /ww exch def /b exch def /g exch def /r exch def 0 1 ww 1 sub { /i exch def r i get .299 mul g i get .587 mul b i get .114 mul add add r i 3 -1 roll floor cvi put } for r } bind def /BITMAPTRUEGRAY { gsave translate rotate scale /h exch def /w exch def /bitmapsave save def /is w string def /gis w string def /bis w string def /cf currentfile def w h 8 [w 0 0 h neg 0 h] { cf is readhexstring pop cf gis readhexstring pop cf bis readhexstring pop w gray} image bitmapsave restore grestore } bind def /BITMAPGRAY { 8 {fakecolorsetup} COMMONBITMAP } bind def /BITMAPGRAYc { 8 {fakecolorsetup} COMMONBITMAPc } bind def /ENDBITMAP { } bind def end /ALDsave FMLOCAL /ALDmatrix matrix def ALDmatrix currentmatrix pop /StartALD { /ALDsave save def savematrix ALDmatrix setmatrix } bind def /InALD { restorematrix } bind def /DoneALD { ALDsave restore } bind def %%EndProlog %%BeginSetup (3.0) FMVERSION 1 1 595.3 841.89 0 1 16 FMDOCUMENT 0 0 /AvantGarde-Demi FMFONTDEFINE 1 0 /Times-Roman FMFONTDEFINE 2 0 /Times-Bold FMFONTDEFINE 3 0 /Times-Italic FMFONTDEFINE 4 1 /Symbol FMFONTDEFINE 32 FMFILLS 0 0 FMFILL 1 .1 FMFILL 2 .3 FMFILL 3 .5 FMFILL 4 .7 FMFILL 5 .9 FMFILL 6 .97 FMFILL 7 1 FMFILL 8 <0f1e3c78f0e1c387> FMFILL 9 <0f87c3e1f0783c1e> FMFILL 10 FMFILL 11 FMFILL 12 <8142241818244281> FMFILL 13 <03060c183060c081> FMFILL 14 <8040201008040201> FMFILL 16 1 FMFILL 17 .9 FMFILL 18 .7 FMFILL 19 .5 FMFILL 20 .3 FMFILL 21 .1 FMFILL 22 0.03 FMFILL 23 0 FMFILL 24 FMFILL 25 FMFILL 26 <3333333333333333> FMFILL 27 <0000ffff0000ffff> FMFILL 28 <7ebddbe7e7dbbd7e> FMFILL 29 FMFILL 30 <7fbfdfeff7fbfdfe> FMFILL %%EndSetup %%Page: "1" 1 %%BeginPaperSize: A4 %%EndPaperSize 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 0 13 Q 0 X (A NEURAL NETWORK THAT LEARNS TO DO HYPHENATION) 120.35 736.28 T 1 12 Q (Bernd Fritzke) 120.35 702.95 T (Institut f\237r Mathematische Maschinen und Datenverarbeitung) 120.35 688.95 T (Lehrstuhl f\237r Programmiersprachen) 120.35 674.95 T (Universit\212t Erlangen-N\237rnberg) 120.35 660.95 T (D-8520 Erlangen, Germany) 120.35 646.95 T (Christof Nasahl) 120.35 618.95 T (UB ANL A433-SI) 120.35 604.95 T (Siemens AG) 120.35 590.95 T (Gr\237ndlacher Stra\247e 248) 120.35 576.95 T (D-8510 F\237rth, Germany) 120.35 562.95 T 2 F -0.22 (Abstract:) 120.35 543.95 P 1 F -0.22 ( Hyphenation of German words is a highly irregular problem. Existing solu-) 168.97 543.95 P 1.6 (tions for automatic hyphenation are not very satisfying. We successfully applied a) 120.35 529.95 P 0.26 (\322sequential network\323 to this problem. The training algorithm was standard backprop-) 120.35 515.95 P 1 (agation. The network was trained with a collection of 1000 German words together) 120.35 501.95 P 0.27 (with their correct hyphenation. In subsequent tests with unknown words we achieved) 120.35 487.95 P -0.24 (a correctness of 96.8 percent. Analysis of the simulation results indicate, that with fur-) 120.35 473.95 P (ther increases of the training data improvements are still possible.) 120.35 459.95 T 2 14 Q (Intr) 63.65 438.61 T (oduction) 87.5 438.61 T 1 12 Q 0.82 (Hyphenation is a standard in text processing software. It is needed to prevent the occurrence of) 63.65 418.95 P 0.1 (wide gaps in the text. In the German language hyphenation is a nontrivial problem, mostly due to) 63.65 404.95 P 1.1 (the occurrence of compound words. Existing solutions are generally rule based or make use of) 63.65 390.95 P 0.36 (large dictionaries. In contrary to the English language, the results of such programs are often not) 63.65 376.95 P (satisfying.) 63.65 362.95 T 0.89 (Neural networks are able to generalize: In a domain with a certain inherent structure learning a) 63.65 343.95 P 0.13 (number of examples might suffice to make the network react properly also on unknown inputs of) 63.65 329.95 P 1.67 (the same domain. Applications of this principle range from speech to text conversion[5] over) 63.65 315.95 P (game playing[6] to truck steering[2].) 63.65 301.95 T 0.42 (The difficulties of describing German hyphenation by rules together with the nevertheless inher-) 63.65 282.95 P -0.17 (ent structure of the problem suggests, that neural networks might be appropriate here. The general) 63.65 268.95 P (proceeding can be as follows:) 63.65 254.95 T (1\051) 90.86 235.95 T -0.21 (Training a network with a limited set of words until it is sufficiently able to determine) 119.21 235.95 P (the correct hyphenation for the word corpus.) 119.21 221.95 T (2\051) 90.86 202.95 T (Applying the trained network to unknown words.) 119.21 202.95 T 1.38 (The point of interest is, to what degree unknown words can be hyphenated correctly, i.e. how) 63.65 183.95 P (much \322generalization\323 takes place.) 63.65 169.95 T 2 14 Q (1.) 63.65 142.61 T ( The Network Ar) 76.69 142.61 T (chitectur) 178.64 142.61 T (e) 232.01 142.61 T 1 12 Q 1.83 (The architecture we used has similarities with the one used by Sejnowski and Rosenberg for) 63.65 122.95 P 1.12 (NETtalk[5]. It consists of three layers of processing units: input layer, hidden layer and output) 63.65 108.95 P (layer.) 63.65 94.95 T 1.02 (The input layer receives the information of 8 adjacent letters \050a \322window\323\051 which are part of a) 63.65 75.95 P FMENDPAGE %%EndPage: "1" 2 %%Page: "2" 2 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 1 12 Q 0 X 0.21 (word. The output layer consists of one single unit, which indicates whether hyphenation is possi-) 63.65 781.95 P -0.24 (ble or not at the \322hyphenation position\323 between the fourth and the fifth letter of the input window) 63.65 767.95 P 1.73 (As) 63.65 753.95 P 2 F 1.73 (hyphenation position) 81.71 753.95 P 1 F 1.73 ( we denote in the following every position between two letters in a) 191.08 753.95 P 1.41 (word. The complete hyphenation of the word is thus available by shifting the 8-letter-window) 63.65 739.95 P (character by character over the word.) 63.65 725.95 T 0.87 (Every word in the training corpus is converted to several pairs of input/output patterns, e.g. the) 63.65 711.95 P 0.73 (word \322ROTATION\323, hyphenated \322RO-TA-TI-ON\323 in the German language, is converted to the) 63.65 697.95 P (following four patterns:) 63.65 683.95 T 0.95 (Empty positions in the input patterns are filled with blanks. Letters outside the window are not) 63.65 501.95 P 0.45 (considered. As in the German language single letters are never hyphenated, we don\325t have to re-) 63.65 487.95 P (gard \050and code\051 the first and the last hyphenation position.) 63.65 473.95 T (The complete network architecture is shown in figure 2.) 63.65 459.95 T 0.5 (Every character position in the word is coded by) 63.65 123.95 P 3 F 0.5 (n) 303.64 123.95 P 1 F 0.5 ( neurons with) 309.64 123.95 P 3 F 0.5 (n) 379.43 123.95 P 1 F 0.5 ( being the size of the alphabet) 385.43 123.95 P 0.65 (used. In the German language we have the letters a..z plus 4 extra characters. Together with the) 63.65 109.95 P 0.01 (blank as fill character we have) 63.65 95.95 P 0.01 ( different letters. To code a letter we set the corresponding) 251.47 95.95 P -0.12 (neuron to one and the other 30 to zero \0501-out-of-n coding\051. This is done for each of the eight input) 63.65 81.95 P 63.65 69.95 531.65 789.95 C 63.65 519.08 531.65 679.95 C 61.31 519.08 533.99 679.95 R 7 X 0 K V 87.15 569.48 512.67 644.75 R V 1 12 Q 0 X (pattern 1:) 101.33 636.75 T (_) 186.37 636.75 T (_) 200.52 636.75 T (R) 214.71 636.75 T (O) 228.88 636.75 T (T) 243.06 636.75 T (A) 257.23 636.75 T (T) 271.4 636.75 T (I) 285.58 636.75 T (yes) 384.79 636.75 T (pattern 2:) 101.33 622.75 T (_) 186.37 622.75 T (R) 200.52 622.75 T (O) 214.71 622.75 T (T) 228.88 622.75 T (A) 243.06 622.75 T (T) 257.23 622.75 T (I) 271.4 622.75 T (O) 285.58 622.75 T (no) 384.79 622.75 T (pattern 3:) 101.33 608.75 T (R) 186.37 608.75 T (O) 200.52 608.75 T (T) 214.71 608.75 T (A) 228.88 608.75 T (T) 243.06 608.75 T (I) 257.23 608.75 T (O) 271.4 608.75 T (N) 285.58 608.75 T (yes) 384.79 608.75 T (pattern 4:) 101.33 594.75 T (O) 186.37 594.75 T (T) 200.52 594.75 T (A) 214.71 594.75 T (T) 228.88 594.75 T (I) 243.06 594.75 T (O) 257.23 594.75 T (N) 271.4 594.75 T (_) 285.58 594.75 T (no) 384.79 594.75 T (pattern 5:) 101.33 580.75 T (T) 186.37 580.75 T (A) 200.52 580.75 T (T) 214.71 580.75 T (I) 228.88 580.75 T (O) 243.06 580.75 T (N) 257.23 580.75 T (_) 271.4 580.75 T (_) 285.58 580.75 T (yes) 384.79 580.75 T 240.68 648.92 239.65 569.48 2 L 0.5 H 2 Z N 275.07 548.28 430.07 560.78 R 7 X V 0 X (examined hyphenation position) 275.07 552.78 T 202.62 651.43 473.87 663.93 R 7 X V 0 X (input information) 202.62 655.93 T (output information) 350.41 655.93 T 251.04 557.86 240.9 564.29 252.9 564.21 251.97 561.04 4 Y V 251.98 561.04 270.07 555.95 2 L N 2 F (Fig. 1:) 89.77 527.27 T 1 F ( The input/output patterns generated from the German word \322Rotation\323) 132.29 527.27 T 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C 63.65 69.95 531.65 789.95 C 63.65 145.19 531.65 455.95 C 61.19 145.19 534.11 455.95 R 7 X 0 K V 2 12 Q 0 X (Fig. 2:) 315.35 214.6 T 1 F ( The Network Architecture) 357.87 214.6 T 7 X 90 450 5.61 5.61 147.46 352.03 G 0.5 H 2 Z 0 X 90 450 5.61 5.61 147.46 352.03 A 7 X 90 450 5.61 5.61 156.91 352.03 G 0 X 90 450 5.61 5.61 156.91 352.03 A 7 X 90 450 5.61 5.61 166.37 352.03 G 0 X 90 450 5.61 5.61 166.37 352.03 A 7 X 90 450 5.61 5.61 175.82 352.03 G 0 X 90 450 5.61 5.61 175.82 352.03 A 7 X 90 450 5.61 5.61 185.28 352.03 G 0 X 90 450 5.61 5.61 185.28 352.03 A 90 450 1.4 1.4 204.19 352.03 G 90 450 1.4 1.4 204.19 352.03 A 90 450 1.4 1.4 213.64 352.03 G 90 450 1.4 1.4 213.64 352.03 A 90 450 1.4 1.4 223.1 352.03 G 90 450 1.4 1.4 223.1 352.03 A 7 X 90 450 5.61 5.61 242.01 352.03 G 0 X 90 450 5.61 5.61 242.01 352.03 A 7 X 90 450 5.61 5.61 251.46 352.03 G 0 X 90 450 5.61 5.61 251.46 352.03 A 7 X 90 450 5.61 5.61 260.92 352.03 G 0 X 90 450 5.61 5.61 260.92 352.03 A 7 X 90 450 5.61 5.61 270.37 352.03 G 0 X 90 450 5.61 5.61 270.37 352.03 A 7 X 90 450 5.61 5.61 279.83 352.03 G 0 X 90 450 5.61 5.61 279.83 352.03 A 7 X 90 450 5.91 5.91 199.31 301.23 G 0 X 90 450 5.91 5.91 199.31 301.23 A 7 X 90 450 5.91 5.91 192.14 294.05 G 0 X 90 450 5.91 5.91 192.14 294.05 A 7 X 90 450 5.91 5.91 184.97 286.88 G 0 X 90 450 5.91 5.91 184.97 286.88 A 7 X 90 450 5.91 5.91 177.79 279.71 G 0 X 90 450 5.91 5.91 177.79 279.71 A 7 X 90 450 5.91 5.91 170.62 272.54 G 0 X 90 450 5.91 5.91 170.62 272.54 A 90 450 1.48 1.48 156.28 258.2 G 90 450 1.48 1.48 156.28 258.2 A 90 450 1.48 1.48 149.11 251.03 G 90 450 1.48 1.48 149.11 251.03 A 90 450 1.48 1.48 141.94 243.86 G 90 450 1.48 1.48 141.94 243.86 A 7 X 90 450 5.91 5.91 127.6 229.52 G 0 X 90 450 5.91 5.91 127.6 229.52 A 7 X 90 450 5.91 5.91 120.43 222.34 G 0 X 90 450 5.91 5.91 120.43 222.34 A 7 X 90 450 5.91 5.91 113.26 215.17 G 0 X 90 450 5.91 5.91 113.26 215.17 A 7 X 90 450 5.91 5.91 106.08 208 G 0 X 90 450 5.91 5.91 106.08 208 A 7 X 90 450 5.91 5.91 98.91 200.83 G 0 X 90 450 5.91 5.91 98.91 200.83 A 7 X 90 450 5.91 5.91 214.54 301.23 G 0 X 90 450 5.91 5.91 214.54 301.23 A 7 X 90 450 5.91 5.91 207.37 294.05 G 0 X 90 450 5.91 5.91 207.37 294.05 A 7 X 90 450 5.91 5.91 200.2 286.88 G 0 X 90 450 5.91 5.91 200.2 286.88 A 7 X 90 450 5.91 5.91 193.03 279.71 G 0 X 90 450 5.91 5.91 193.03 279.71 A 7 X 90 450 5.91 5.91 185.85 272.54 G 0 X 90 450 5.91 5.91 185.85 272.54 A 90 450 1.48 1.48 171.51 258.2 G 90 450 1.48 1.48 171.51 258.2 A 90 450 1.48 1.48 164.34 251.03 G 90 450 1.48 1.48 164.34 251.03 A 90 450 1.48 1.48 157.17 243.86 G 90 450 1.48 1.48 157.17 243.86 A 7 X 90 450 5.91 5.91 142.83 229.52 G 0 X 90 450 5.91 5.91 142.83 229.52 A 7 X 90 450 5.91 5.91 135.66 222.34 G 0 X 90 450 5.91 5.91 135.66 222.34 A 7 X 90 450 5.91 5.91 128.49 215.17 G 0 X 90 450 5.91 5.91 128.49 215.17 A 7 X 90 450 5.91 5.91 121.32 208 G 0 X 90 450 5.91 5.91 121.32 208 A 7 X 90 450 5.91 5.91 114.14 200.83 G 0 X 90 450 5.91 5.91 114.14 200.83 A 7 X 90 450 5.91 5.91 229.77 301.23 G 0 X 90 450 5.91 5.91 229.77 301.23 A 7 X 90 450 5.91 5.91 222.6 294.05 G 0 X 90 450 5.91 5.91 222.6 294.05 A 7 X 90 450 5.91 5.91 215.43 286.88 G 0 X 90 450 5.91 5.91 215.43 286.88 A 7 X 90 450 5.91 5.91 208.26 279.71 G 0 X 90 450 5.91 5.91 208.26 279.71 A 7 X 90 450 5.91 5.91 201.09 272.54 G 0 X 90 450 5.91 5.91 201.09 272.54 A 90 450 1.48 1.48 186.74 258.2 G 90 450 1.48 1.48 186.74 258.2 A 90 450 1.48 1.48 179.57 251.03 G 90 450 1.48 1.48 179.57 251.03 A 90 450 1.48 1.48 172.4 243.86 G 90 450 1.48 1.48 172.4 243.86 A 7 X 90 450 5.91 5.91 158.06 229.52 G 0 X 90 450 5.91 5.91 158.06 229.52 A 7 X 90 450 5.91 5.91 150.89 222.34 G 0 X 90 450 5.91 5.91 150.89 222.34 A 7 X 90 450 5.91 5.91 143.72 215.17 G 0 X 90 450 5.91 5.91 143.72 215.17 A 7 X 90 450 5.91 5.91 136.55 208 G 0 X 90 450 5.91 5.91 136.55 208 A 7 X 90 450 5.91 5.91 129.37 200.83 G 0 X 90 450 5.91 5.91 129.37 200.83 A 7 X 90 450 5.91 5.91 245 301.23 G 0 X 90 450 5.91 5.91 245 301.23 A 7 X 90 450 5.91 5.91 237.83 294.05 G 0 X 90 450 5.91 5.91 237.83 294.05 A 7 X 90 450 5.91 5.91 230.66 286.88 G 0 X 90 450 5.91 5.91 230.66 286.88 A 7 X 90 450 5.91 5.91 223.49 279.71 G 0 X 90 450 5.91 5.91 223.49 279.71 A 7 X 90 450 5.91 5.91 216.32 272.54 G 0 X 90 450 5.91 5.91 216.32 272.54 A 90 450 1.48 1.48 201.97 258.2 G 90 450 1.48 1.48 201.97 258.2 A 90 450 1.48 1.48 194.8 251.03 G 90 450 1.48 1.48 194.8 251.03 A 90 450 1.48 1.48 187.63 243.86 G 90 450 1.48 1.48 187.63 243.86 A 7 X 90 450 5.91 5.91 173.29 229.52 G 0 X 90 450 5.91 5.91 173.29 229.52 A 7 X 90 450 5.91 5.91 166.12 222.34 G 0 X 90 450 5.91 5.91 166.12 222.34 A 7 X 90 450 5.91 5.91 158.95 215.17 G 0 X 90 450 5.91 5.91 158.95 215.17 A 7 X 90 450 5.91 5.91 151.78 208 G 0 X 90 450 5.91 5.91 151.78 208 A 7 X 90 450 5.91 5.91 144.6 200.83 G 0 X 90 450 5.91 5.91 144.6 200.83 A 7 X 90 450 5.91 5.91 260.23 301.23 G 0 X 90 450 5.91 5.91 260.23 301.23 A 7 X 90 450 5.91 5.91 253.06 294.05 G 0 X 90 450 5.91 5.91 253.06 294.05 A 7 X 90 450 5.91 5.91 245.89 286.88 G 0 X 90 450 5.91 5.91 245.89 286.88 A 7 X 90 450 5.91 5.91 238.72 279.71 G 0 X 90 450 5.91 5.91 238.72 279.71 A 7 X 90 450 5.91 5.91 231.55 272.54 G 0 X 90 450 5.91 5.91 231.55 272.54 A 90 450 1.48 1.48 217.21 258.2 G 90 450 1.48 1.48 217.21 258.2 A 90 450 1.48 1.48 210.03 251.03 G 90 450 1.48 1.48 210.03 251.03 A 90 450 1.48 1.48 202.86 243.86 G 90 450 1.48 1.48 202.86 243.86 A 7 X 90 450 5.91 5.91 188.52 229.52 G 0 X 90 450 5.91 5.91 188.52 229.52 A 7 X 90 450 5.91 5.91 181.35 222.34 G 0 X 90 450 5.91 5.91 181.35 222.34 A 7 X 90 450 5.91 5.91 174.18 215.17 G 0 X 90 450 5.91 5.91 174.18 215.17 A 7 X 90 450 5.91 5.91 167.01 208 G 0 X 90 450 5.91 5.91 167.01 208 A 7 X 90 450 5.91 5.91 159.84 200.83 G 0 X 90 450 5.91 5.91 159.84 200.83 A 7 X 90 450 5.91 5.91 275.46 301.23 G 0 X 90 450 5.91 5.91 275.46 301.23 A 7 X 90 450 5.91 5.91 268.29 294.05 G 0 X 90 450 5.91 5.91 268.29 294.05 A 7 X 90 450 5.91 5.91 261.12 286.88 G 0 X 90 450 5.91 5.91 261.12 286.88 A 7 X 90 450 5.91 5.91 253.95 279.71 G 0 X 90 450 5.91 5.91 253.95 279.71 A 7 X 90 450 5.91 5.91 246.78 272.54 G 0 X 90 450 5.91 5.91 246.78 272.54 A 90 450 1.48 1.48 232.44 258.2 G 90 450 1.48 1.48 232.44 258.2 A 90 450 1.48 1.48 225.26 251.03 G 90 450 1.48 1.48 225.26 251.03 A 90 450 1.48 1.48 218.09 243.86 G 90 450 1.48 1.48 218.09 243.86 A 7 X 90 450 5.91 5.91 203.75 229.52 G 0 X 90 450 5.91 5.91 203.75 229.52 A 7 X 90 450 5.91 5.91 196.58 222.34 G 0 X 90 450 5.91 5.91 196.58 222.34 A 7 X 90 450 5.91 5.91 189.41 215.17 G 0 X 90 450 5.91 5.91 189.41 215.17 A 7 X 90 450 5.91 5.91 182.24 208 G 0 X 90 450 5.91 5.91 182.24 208 A 7 X 90 450 5.91 5.91 175.07 200.83 G 0 X 90 450 5.91 5.91 175.07 200.83 A 7 X 90 450 5.91 5.91 290.69 301.23 G 0 X 90 450 5.91 5.91 290.69 301.23 A 7 X 90 450 5.91 5.91 283.52 294.05 G 0 X 90 450 5.91 5.91 283.52 294.05 A 7 X 90 450 5.91 5.91 276.35 286.88 G 0 X 90 450 5.91 5.91 276.35 286.88 A 7 X 90 450 5.91 5.91 269.18 279.71 G 0 X 90 450 5.91 5.91 269.18 279.71 A 7 X 90 450 5.91 5.91 262.01 272.54 G 0 X 90 450 5.91 5.91 262.01 272.54 A 90 450 1.48 1.48 247.67 258.2 G 90 450 1.48 1.48 247.67 258.2 A 90 450 1.48 1.48 240.5 251.03 G 90 450 1.48 1.48 240.5 251.03 A 90 450 1.48 1.48 233.32 243.86 G 90 450 1.48 1.48 233.32 243.86 A 7 X 90 450 5.91 5.91 218.98 229.52 G 0 X 90 450 5.91 5.91 218.98 229.52 A 7 X 90 450 5.91 5.91 211.81 222.34 G 0 X 90 450 5.91 5.91 211.81 222.34 A 7 X 90 450 5.91 5.91 204.64 215.17 G 0 X 90 450 5.91 5.91 204.64 215.17 A 7 X 90 450 5.91 5.91 197.47 208 G 0 X 90 450 5.91 5.91 197.47 208 A 7 X 90 450 5.91 5.91 190.3 200.83 G 0 X 90 450 5.91 5.91 190.3 200.83 A 7 X 90 450 5.91 5.91 305.92 301.23 G 0 X 90 450 5.91 5.91 305.92 301.23 A 7 X 90 450 5.91 5.91 298.75 294.05 G 0 X 90 450 5.91 5.91 298.75 294.05 A 7 X 90 450 5.91 5.91 291.58 286.88 G 0 X 90 450 5.91 5.91 291.58 286.88 A 7 X 90 450 5.91 5.91 284.41 279.71 G 0 X 90 450 5.91 5.91 284.41 279.71 A 7 X 90 450 5.91 5.91 277.24 272.54 G 0 X 90 450 5.91 5.91 277.24 272.54 A 90 450 1.48 1.48 262.9 258.2 G 90 450 1.48 1.48 262.9 258.2 A 90 450 1.48 1.48 255.73 251.03 G 90 450 1.48 1.48 255.73 251.03 A 90 450 1.48 1.48 248.56 243.86 G 90 450 1.48 1.48 248.56 243.86 A 7 X 90 450 5.91 5.91 234.21 229.52 G 0 X 90 450 5.91 5.91 234.21 229.52 A 7 X 90 450 5.91 5.91 227.04 222.34 G 0 X 90 450 5.91 5.91 227.04 222.34 A 7 X 90 450 5.91 5.91 219.87 215.17 G 0 X 90 450 5.91 5.91 219.87 215.17 A 7 X 90 450 5.91 5.91 212.7 208 G 0 X 90 450 5.91 5.91 212.7 208 A 7 X 90 450 5.91 5.91 205.53 200.83 G 0 X 90 450 5.91 5.91 205.53 200.83 A 1 11 Q (1) 225.42 194.38 T (2) 235.15 202.35 T (3) 243.26 211.32 T (31) 326.81 296.49 T (30) 319.32 288.39 T (29) 312.47 280.28 T 7 X 90 450 5.61 5.61 213.64 436.2 G 0 X 90 450 5.61 5.61 213.64 436.2 A 7 X 90 450 5.61 5.61 268.51 199.9 G 0 X 90 450 5.61 5.61 268.51 199.9 A 206.09 320.23 212.71 343.3 219.32 320.23 212.71 320.23 4 Y V 212.71 316.49 212.71 320.23 2 L 4 H N 205.47 375.72 212.08 398.79 218.7 375.72 212.08 375.72 4 Y V 212.08 369.49 212.08 375.72 2 L N 272.98 218.76 269.44 206.14 265.75 218.72 269.36 218.74 4 Y V 1 H 3 90 52.37 240.66 217.07 206.14 A 90 180 2.65 3.12 216.13 443.53 A 84.87 407 160.48 438.78 R 7 X V 1 12 Q 0 X (output layer) 84.87 430.78 T (\0501 unit\051) 84.87 416.78 T (hidden layer) 84.87 359.86 T (\05040 units\051) 84.87 345.86 T (input layer) 84.87 274.76 T (\050249 units\051) 84.87 260.76 T (_) 83.64 178.87 T (_) 98.84 178.87 T (R) 114.04 178.87 T (O) 129.24 178.87 T (T) 144.44 178.87 T (A) 159.64 178.87 T (T) 174.84 178.87 T (I) 190.04 178.87 T 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C 212.9 92.34 251.47 105.15 C 3 12 Q 0 X 0 K (n) 213.9 95.95 T 1 F (3) 238.48 95.95 T (1) 244.47 95.95 T 4 F (=) 225.89 95.95 T -8.35 24.95 603.65 816.95 C FMENDPAGE %%EndPage: "2" 3 %%Page: "3" 3 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 1 12 Q 0 X (positions. Thus) 63.65 781.95 T (neurons are needed to code the input.) 209.1 781.95 T 0.84 (Every input neuron has a connection to every unit in the hidden layer. Every unit in the hidden) 63.65 767.95 P (layer is connected to the output neuron.) 63.65 753.95 T 0.69 (There is one single feedback connection going from the output neuron to an extra input neuron.) 63.65 739.95 P 0.67 (This connection has been introduced to give the network the chance to use the contextual infor-) 63.65 725.95 P 0.04 (mation if hyphenation has been possible at the previous position \050For every syllable consists of at) 63.65 711.95 P 0.72 (least two letters in the German language, adjacent hyphenation positions are not possible!\051. The) 63.65 697.95 P -0.27 (feedback connection is realized by copying the previous value of the output neuron to the extra in-) 63.65 683.95 P 0.13 (put neuron before every pattern presentation. This type of network is called \322sequential network\323) 63.65 669.95 P (due to Jordan [1].) 63.65 655.95 T 0.03 (We made experiments with and without the feedback connection. Thereby the nets with feedback) 63.65 641.95 P 0.4 (had significantly \050with a significance of 0.06\051 better results. For this reason we choosed this net-) 63.65 627.95 P (work architecture for all following investigations.) 63.65 613.95 T (As learning rule we used standard backpropagation[4] with a momentum. The activation function) 63.65 594.95 T (was the logistic function) 63.65 580.95 T (In the output pattern \322Yes\323 was coded as 1 and \322No\323 as 0.) 63.65 510.95 T 2 14 Q (2.) 63.65 483.61 T ( Choosing W) 76.69 483.61 T (ord Bases) 152.89 483.61 T 1 12 Q 0.69 (One important aspect was the choice of the word bases for training and test of the network. We) 63.65 463.95 P (built a word base out of three groups:) 63.65 449.95 T (\245) 82.66 430.95 T (the 325 most frequently used words in the German language \050according to [3]\051.) 100.66 430.95 T (\245) 82.66 411.95 T (850 words selected from articles in the \322Frankfurter Allgemeine Zeitung\323, a widespread) 100.66 411.95 T (German newspaper) 100.66 397.95 T (.) 192.91 397.95 T (\245) 82.66 378.95 T (25 words with rare letters or letter combinations.) 100.66 378.95 T -0.15 (This makes 1200 words at the whole. Out of these we constituted different training bases W) 63.65 359.95 P 3 10 Q -0.12 (n) 502.48 356.95 P 1 12 Q -0.15 ( with) 507.48 359.95 P 3 F 0.11 (n) 63.65 345.95 P 4 14 Q 0.13 ( \316) 69.65 345.95 P 1 12 Q 0.11 ({5, 10, 20, 40, 80, 160, 320, 640, 1000} by selecting) 83.25 345.95 P 3 F 0.11 (n) 339.48 345.95 P 1 F 0.11 ( words at random. Thereby the smaller) 345.48 345.95 P (training bases are subsets of the lar) 63.65 331.95 T (ger ones.) 231.31 331.95 T (As test base W) 63.65 317.95 T 1 10 Q (t) 135.26 314.95 T 1 12 Q ( we took the 200 words which were not incorporated in W) 138.04 317.95 T 1 10 Q (1000) 417.5 314.95 T 1 12 Q (.) 437.49 317.95 T (The words in the word bases were transformed to input/output patterns as described above.) 63.65 303.95 T 2 14 Q (3.) 63.65 276.61 T ( Learning to Hyphenate the T) 76.69 276.61 T (raining W) 253.29 276.61 T (ords) 313.55 276.61 T 1 12 Q (First we trained a network with the standard backpropagation to hyphenate the words in the train-) 63.65 256.95 T 0.21 (ing bases W) 63.65 242.95 P 3 10 Q 0.18 (n) 122.7 239.95 P 1 12 Q 0.21 (,) 127.69 242.95 P 3 F 0.21 (n) 133.9 242.95 P 4 F 0.21 (\316) 143.11 242.95 P 1 F 0.21 ({5, 10, ..., 1000}. One learning cycle consisted of the presentation of all words) 151.66 242.95 P 0.31 (to the network. The presentation order of the words was random. But for each word the different) 63.65 228.95 P -0.07 (input/output patterns for that word \050) 63.65 214.95 P 3 F -0.07 (m) 235.55 214.95 P 1 F -0.07 (-3 for a word of length) 244.21 214.95 P 3 F -0.07 (m) 355.72 214.95 P 1 F -0.07 (\051 were presented sequentially. This) 364.38 214.95 P (was necessary to take advantage of the feedback connection.) 63.65 200.95 T -0.26 (We performed up to 10 runs for each word base W) 63.65 186.95 P 3 10 Q -0.21 (n) 304.88 183.95 P 1 12 Q -0.26 (. Fig. 3 shows the convergence behavior of the) 309.87 186.95 P (network.) 63.65 172.95 T 0.81 (For each word base is given the number of input/output patterns and the total square error after) 63.65 158.95 P (convergence. The total square error is defined as follows:) 63.65 144.95 T 139.95 778.34 209.1 791.12 C 1 12 Q 0 X 0 K (8) 140.95 781.95 T (3) 159.54 781.95 T (1) 165.53 781.95 T 4 F (\264) 149.95 781.95 T 1 F (2) 190.11 781.95 T (4) 196.11 781.95 T (8) 202.11 781.95 T 4 F (=) 177.53 781.95 T -8.35 24.95 603.65 816.95 C 63.65 69.95 531.65 789.95 C 69.65 534.39 525.65 576.95 C 3 9 Q 0 X 0 K (f) 268.73 554.57 T (x) 277.14 554.57 T 4 F (\050) 273.25 554.57 T (\051) 281.52 554.57 T 1 F (1) 311.02 560.47 T (1) 300.7 545.26 T 3 F (e) 314.63 545.26 T 3 7 Q (x) 322.72 549.73 T 4 F (-) 318.62 549.73 T 4 9 Q (+) 307.44 545.26 T (=) 290.52 554.57 T 300.7 556.52 325.58 556.52 2 L 0.33 H 0 Z N 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C 63.65 69.95 531.65 789.95 C 69.65 70.21 525.65 126.95 C 1 12 Q 0 X 0 K (t) 92.78 94.39 T (o) 96.11 94.39 T (t) 102.11 94.39 T (a) 105.44 94.39 T (l) 110.77 94.39 T (\021) 114.1 94.39 T (s) 117.1 94.39 T (q) 121.77 94.39 T (u) 127.76 94.39 T (a) 133.76 94.39 T (r) 139.09 94.39 T (e) 143.08 94.39 T (\021) 148.4 94.39 T (e) 151.4 94.39 T (r) 156.73 94.39 T (r) 160.72 94.39 T (o) 164.71 94.39 T (r) 170.71 94.39 T (d) 271.28 94.39 T (e) 277.28 94.39 T (s) 282.6 94.39 T (i) 287.27 94.39 T (r) 290.6 94.39 T (e) 294.6 94.39 T (d) 299.92 94.39 T (\021) 305.92 94.39 T (o) 308.92 94.39 T (u) 314.91 94.39 T (t) 320.91 94.39 T (p) 324.24 94.39 T (u) 330.24 94.39 T (t) 336.24 94.39 T (\021) 339.57 94.39 T (f) 342.57 94.39 T (o) 346.56 94.39 T (r) 352.56 94.39 T (\021) 356.55 94.39 T (p) 359.55 94.39 T (a) 365.55 94.39 T (t) 370.88 94.39 T (t) 374.21 94.39 T (e) 377.54 94.39 T (r) 382.87 94.39 T (n) 386.86 94.39 T 3 F (i) 394.56 94.39 T 1 F (a) 410.48 94.39 T (c) 415.81 94.39 T (t) 421.13 94.39 T (u) 424.47 94.39 T (a) 430.46 94.39 T (l) 435.79 94.39 T (\021) 439.12 94.39 T (n) 442.12 94.39 T (e) 448.12 94.39 T (t) 453.44 94.39 T (\021) 456.78 94.39 T (o) 459.78 94.39 T (u) 465.77 94.39 T (t) 471.77 94.39 T (p) 475.1 94.39 T (u) 481.1 94.39 T (t) 487.1 94.39 T 4 F (-) 400.9 94.39 T (\050) 266.18 94.39 T (\051) 492.04 94.39 T 1 9 Q (2) 498.03 100.51 T 3 F (i) 219.77 80.1 T 1 F (1) 233.2 80.1 T 4 F (=) 225.27 80.1 T 1 F (n) 194.29 109.17 T (u) 198.78 109.17 T (m) 203.28 109.17 T (b) 210.27 109.17 T (e) 214.76 109.17 T (r) 218.75 109.17 T (\021) 221.75 109.17 T (o) 223.99 109.17 T (f) 228.49 109.17 T (\021) 231.48 109.17 T (p) 233.73 109.17 T (a) 238.22 109.17 T (t) 242.21 109.17 T (t) 244.71 109.17 T (e) 247.21 109.17 T (r) 251.2 109.17 T (n) 254.19 109.17 T (s) 258.68 109.17 T 4 18 Q (\345) 222.32 91.19 T 4 12 Q (=) 181.71 94.39 T 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C FMENDPAGE %%EndPage: "3" 4 %%Page: "4" 4 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 1 12 Q 0 X (The relative error is computed as) 63.65 653.37 T 1.31 (The \322maximal possible error\323 is equal to the \322number of input/output patterns\323 because in the) 63.65 583.37 P 0.45 (worst case the network would map every input-pattern for which hyphenation should occur on 0) 63.65 569.37 P (\050instead of 1\051 and vice versa. The error would be 1 or -1 in either case, the square error always 1.) 63.65 555.37 T 0.43 (As the table indicates the network is not able to learn the hyphenation perfectly if the size of the) 63.65 527.37 P 0.19 (training base exceeds 20. In all cases the remaining relative error was below one percent. The er-) 63.65 513.37 P -0.15 (ror could probably be eliminated by further increase of the hidden layer. However since one train-) 63.65 499.37 P 1.35 (ing cycle with W) 63.65 485.37 P 1 10 Q 1.13 (1000) 149.99 482.37 P 1 12 Q 1.35 ( took already about 44 cpu minutes on our HP-9000/330 workstation we) 169.98 485.37 P 0.01 (didn\325) 63.65 471.37 P 0.01 (t check that out.) 88.75 471.37 P 0.01 ( Nevertheless the effect we want to show is obvious also with these \322errone-) 165.41 471.37 P (ous\323 networks as the next chapter shows.) 63.65 457.37 T 2 14 Q (4.) 63.65 430.03 T ( T) 76.69 430.03 T (esting the Network with Unknown W) 88.24 430.03 T (ords) 309.4 430.03 T 1 12 Q 0.83 (Learning the hyphenation of a fixed number of words is not very exciting. This could easily be) 63.65 410.37 P 0.42 (done with a table. Instead we are interested if the network would acquire some general ability to) 63.65 396.37 P 0.9 (hyphenate through learning the training words. For this reason we performed various tests with) 63.65 382.37 P -0.17 (the word base W) 63.65 74.37 P 1 10 Q -0.14 (t) 144.06 71.37 P 1 12 Q -0.17 ( which consists of 200 words which had not been used during the training. Fig. 4) 146.84 74.37 P 64.23 661.37 531.07 789.95 C 64.23 661.37 531.07 789.95 R 7 X 0 K V 80.9 696.85 506.42 769.44 R V 1 12 Q 0 X (base is W) 95.06 761.44 T 3 10 Q (n) 141.69 758.44 T 1 12 Q ( with) 146.69 761.44 T 3 F (n) 174.01 761.44 T 1 F (=) 180.01 761.44 T (5) 236.79 761.44 T (10) 265.14 761.44 T (20) 293.49 761.44 T (40) 321.83 761.44 T (80) 350.18 761.44 T (160) 378.52 761.44 T (320) 406.87 761.44 T (640) 435.22 761.44 T (1000) 465.65 761.44 T (number of I/O patterns) 95.06 733.44 T (34) 236.79 733.44 T (60) 265.14 733.44 T (101) 293.49 733.44 T (215) 321.83 733.44 T (468) 350.18 733.44 T (908) 378.52 733.44 T (1917) 406.87 733.44 T (3744) 435.22 733.44 T (5793) 465.65 733.44 T (\336nal square error) 95.06 719.44 T (0) 236.79 719.44 T (0) 265.14 719.44 T (0) 293.49 719.44 T (0.1) 321.83 719.44 T (1.6) 350.18 719.44 T (4.7) 378.52 719.44 T (9) 406.87 719.44 T (32) 435.22 719.44 T (39) 465.65 719.44 T (relative error in %) 95.06 705.44 T (0) 236.79 705.44 T (0) 265.14 705.44 T (0) 293.49 705.44 T (.046) 321.83 705.44 T (.342) 350.18 705.44 T (.518) 378.52 705.44 T (.470) 406.87 705.44 T (.855) 435.22 705.44 T (.673) 465.65 705.44 T 81.73 750.24 499.65 750.24 2 L 0.5 H 2 Z N 222.89 774.95 222.89 689.84 2 L N 64.23 675.15 531.95 689.33 R 7 X V 2 F 0 X (Fig. 3:) 91.45 681.33 T 1 F ( Error after learning the different training bases W) 133.96 681.33 T 3 10 Q (n) 375.09 678.33 T -8.35 24.95 603.65 816.95 C 63.65 69.95 531.65 789.95 C 69.65 607.16 525.66 649.37 C 1 12 Q 0 X 0 K (r) 114.72 622.28 T (e) 118.71 622.28 T (l) 124.04 622.28 T (a) 127.37 622.28 T (t) 132.69 622.28 T (i) 136.03 622.28 T (v) 139.36 622.28 T (e) 145.36 622.28 T (\021) 150.68 622.28 T (e) 153.68 622.28 T (r) 159.01 622.28 T (r) 163 622.28 T (o) 166.99 622.28 T (r) 172.99 622.28 T (\021) 176.99 622.28 T (=) 179.98 622.28 T (t) 203.79 629.47 T (o) 207.12 629.47 T (t) 213.12 629.47 T (a) 216.45 629.47 T (l) 221.78 629.47 T (\021) 225.11 629.47 T (s) 228.11 629.47 T (q) 232.77 629.47 T (u) 238.77 629.47 T (a) 244.77 629.47 T (r) 250.09 629.47 T (e) 254.09 629.47 T (\021) 259.41 629.47 T (e) 262.41 629.47 T (r) 267.73 629.47 T (r) 271.73 629.47 T (o) 275.72 629.47 T (r) 281.72 629.47 T (m) 189.45 614.69 T (a) 198.79 614.69 T (x) 204.11 614.69 T (i) 210.11 614.69 T (m) 213.44 614.69 T (a) 222.77 614.69 T (l) 228.1 614.69 T (\021) 231.43 614.69 T (p) 234.43 614.69 T (o) 240.43 614.69 T (s) 246.42 614.69 T (s) 251.09 614.69 T (i) 255.75 614.69 T (b) 259.09 614.69 T (l) 265.08 614.69 T (e) 268.42 614.69 T (\021) 273.74 614.69 T (e) 276.74 614.69 T (r) 282.07 614.69 T (r) 286.06 614.69 T (o) 290.05 614.69 T (r) 296.05 614.69 T (t) 356.79 629.47 T (o) 360.12 629.47 T (t) 366.12 629.47 T (a) 369.45 629.47 T (l) 374.78 629.47 T (\021) 378.11 629.47 T (s) 381.11 629.47 T (q) 385.78 629.47 T (u) 391.77 629.47 T (a) 397.77 629.47 T (r) 403.1 629.47 T (e) 407.09 629.47 T (\021) 412.41 629.47 T (e) 415.41 629.47 T (r) 420.74 629.47 T (r) 424.73 629.47 T (o) 428.73 629.47 T (r) 434.72 629.47 T (n) 321.63 614.69 T (u) 327.63 614.69 T (m) 333.62 614.69 T (b) 342.95 614.69 T (e) 348.95 614.69 T (r) 354.28 614.69 T (\021) 358.27 614.69 T (o) 361.27 614.69 T (f) 367.27 614.69 T (\021) 371.26 614.69 T (i) 374.26 614.69 T (n) 377.59 614.69 T (p) 383.59 614.69 T (u) 389.58 614.69 T (t) 395.58 614.69 T (/) 398.92 614.69 T (o) 402.25 614.69 T (u) 408.25 614.69 T (t) 414.24 614.69 T (p) 417.58 614.69 T (u) 423.57 614.69 T (t) 429.57 614.69 T (\021) 432.9 614.69 T (p) 435.9 614.69 T (a) 441.9 614.69 T (t) 447.23 614.69 T (t) 450.56 614.69 T (e) 453.89 614.69 T (r) 459.22 614.69 T (n) 463.21 614.69 T (s) 469.21 614.69 T (.) 476.58 622.28 T 4 F (=) 308.05 622.28 T 189.45 624.86 300.8 624.86 2 L 0.33 H 0 Z N 321.63 624.86 474.62 624.86 2 L N 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C 64.73 88.58 530.58 378.37 C 64.73 88.58 530.58 378.37 R 7 X 0 K V 66.73 81.13 520.62 109.5 R V 2 12 Q 0 X (Fig. 4:) 79.15 101.5 T 1 F ( Hyphenation error for the test base W) 113.32 101.5 T 1 10 Q (t) 297.5 98.5 T 1 12 Q ( after learning the different training bases W) 300.28 101.5 T 3 10 Q (n) 513.1 98.5 T 108.61 339.49 207.62 357.49 R 7 X V 1 12 Q 0 X (error in percentage) 108.61 349.49 T 106.01 245.64 107.98 211.12 109.98 198.1 113.94 188.65 121.87 186.47 137.75 182.13 169.5 175.24 232.99 167.33 359.97 162.8 502.83 161.22 10 L 1 H 2 Z N 106.01 153.34 106.01 360.46 2 L N 106.01 153.34 512.76 153.34 2 L N 106.01 153.34 106.01 149.39 2 L N (0) 103.01 139.19 T 125.85 153.34 125.85 149.39 2 L N 145.68 153.34 145.68 149.39 2 L N (100) 136.69 139.19 T 165.52 153.34 165.52 149.39 2 L N 185.38 153.34 185.38 149.39 2 L N (200) 176.39 139.19 T 205.22 153.34 205.22 149.39 2 L N 225.06 153.34 225.06 149.39 2 L N (300) 216.06 139.19 T 244.9 153.34 244.9 149.39 2 L N 264.73 153.34 264.73 149.39 2 L N (400) 255.74 139.19 T 284.57 153.34 284.57 149.39 2 L N 304.41 153.34 304.41 149.39 2 L N (500) 295.41 139.19 T 324.27 153.34 324.27 149.39 2 L N 344.11 153.34 344.11 149.39 2 L N (600) 335.11 139.19 T 363.95 153.34 363.95 149.39 2 L N 383.78 153.34 383.78 149.39 2 L N (700) 374.79 139.19 T 403.62 153.34 403.62 149.39 2 L N 423.46 153.34 423.46 149.39 2 L N (800) 414.46 139.19 T 443.3 153.34 443.3 149.39 2 L N 463.16 153.34 463.16 149.39 2 L N (900) 454.16 139.19 T 483 153.34 483 149.39 2 L N 502.83 153.34 502.83 149.39 2 L N (1000) 490.84 139.19 T 106.01 153.34 102.03 153.34 2 L N (0) 94.6 152.99 T 106.01 173.06 102.03 173.06 2 L N (10) 89.14 172.72 T 106.01 192.78 102.03 192.78 2 L N (20) 89.14 192.46 T 106.01 212.51 102.03 212.51 2 L N (30) 89.14 212.18 T 106.01 232.23 102.03 232.23 2 L N (40) 89.14 231.91 T 106.01 251.95 102.03 251.95 2 L N (50) 89.14 251.63 T 106.01 271.7 102.03 271.7 2 L N (60) 89.14 271.35 T 106.01 291.42 102.03 291.42 2 L N (70) 89.14 291.07 T 106.01 311.14 102.03 311.14 2 L N (80) 89.14 310.79 T 106.01 330.86 102.03 330.86 2 L N (90) 89.14 330.54 T 106.01 350.58 102.03 350.58 2 L N (100) 83.68 350.26 T 7 X 90 450 1.86 1.85 106.01 245.64 G 0.5 H 0 Z 0 X 90 450 1.86 1.85 106.01 245.64 A 7 X 90 450 1.86 1.85 107.98 211.12 G 0 X 90 450 1.86 1.85 107.98 211.12 A 7 X 90 450 1.86 1.86 109.98 198.1 G 0 X 90 450 1.86 1.86 109.98 198.1 A 7 X 90 450 1.86 1.85 113.94 188.65 G 0 X 90 450 1.86 1.85 113.94 188.65 A 7 X 90 450 1.86 1.85 121.87 186.47 G 0 X 90 450 1.86 1.85 121.87 186.47 A 7 X 90 450 1.86 1.85 137.75 182.13 G 0 X 90 450 1.86 1.85 137.75 182.13 A 7 X 90 450 1.86 1.85 169.5 175.24 G 0 X 90 450 1.86 1.85 169.5 175.24 A 7 X 90 450 1.86 1.85 232.99 167.33 G 0 X 90 450 1.86 1.85 232.99 167.33 A 7 X 90 450 1.86 1.85 359.97 162.8 G 0 X 90 450 1.86 1.85 359.97 162.8 A 7 X 90 450 1.86 1.85 502.83 161.22 G 0 X 90 450 1.86 1.85 502.83 161.22 A 374.62 122.49 515.62 134.49 R 7 X V 0 X (number of training patterns) 374.62 126.49 T (W) 491.98 170.03 T 1 10 Q (1000) 503.31 167.03 T 1 12 Q (W) 352.65 172.86 T 1 10 Q (640) 363.97 169.86 T 1 12 Q (W) 223.15 175.36 T 1 10 Q (320) 234.47 172.36 T 1 12 Q (W) 164.98 183.03 T 1 10 Q (160) 176.3 180.03 T 239.62 296.49 427.62 309.49 R 7 X V 1 12 Q 0 X (incorrectly hyphenated positions) 239.62 301.49 T -8.35 24.95 603.65 816.95 C FMENDPAGE %%EndPage: "4" 5 %%Page: "5" 5 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 1 12 Q 0 X (shows the hyphenation error for these new words after training with different word bases W) 63.65 781.95 T 3 10 Q (n) 504.64 778.95 T 1 12 Q (.) 509.64 781.95 T 0.69 (There is a clear correlation between the size of the training base and the hyphenation ability for) 63.65 767.95 P 0.28 (new words. With a training base of only 10 words 77.8% of the test patterns are hyphenated cor-) 63.65 753.95 P (rectly. This improves to 96% with a training base of 1000 words.) 63.65 739.95 T 0.41 (It is interesting to remark, that improvements in the hyphenation of the training words do not al-) 63.65 725.95 P 0.19 (ways lead to better hyphenation of the test words. In Fig.5 the hyphenation error for W) 63.65 711.95 P 1 10 Q 0.16 (t) 483.18 708.95 P 1 12 Q 0.19 ( is shown) 485.96 711.95 P 0.3 (in dif) 63.65 697.95 P 0.3 (ferent stages of the training with W) 89.39 697.95 P 1 10 Q 0.25 (1000) 261.04 694.95 P 1 12 Q 0.3 (, the lar) 281.03 697.95 P 0.3 (gest training base. While until cycle 25 both) 317.71 697.95 P -0.22 (the error for W) 63.65 683.95 P 1 10 Q -0.18 (1000) 135.26 680.95 P 1 12 Q -0.22 ( and the error for W) 155.24 683.95 P 1 10 Q -0.18 (t) 249.73 680.95 P 1 12 Q -0.22 ( decrease, further training only diminishes the error for the) 252.51 683.95 P (training words.) 63.65 669.95 T 2 14 Q (5.) 63.65 401.48 T ( Effects of the Hidden Layer Size) 76.69 401.48 T 1 12 Q 0.17 (To determine the influence of the number of hidden units on the performance of our network, we) 63.65 381.82 P 0.35 (made simulations also with 10 and 80 hidden units. In each case the training base was W) 63.65 367.82 P 1 10 Q 0.29 (1000) 495.32 364.82 P 1 12 Q 0.35 (. In) 515.31 367.82 P 0.02 (Fig.6 the square error for the two new networks and for the one with 40 hidden units is shown for) 63.65 129.82 P 0.35 (a number of training cycles. The conver) 63.65 115.82 P 0.35 (gence seems to be better if the number of hidden units is) 257.03 115.82 P (increased.) 63.65 101.82 T 0.66 (The hyphenation ability for our test base W) 63.65 87.82 P 1 10 Q 0.55 (t) 277.11 84.82 P 1 12 Q 0.66 ( improved too. After 35 cycles the 80-hidden-units-) 279.89 87.82 P 63.65 69.95 531.65 789.95 C 63.65 424.82 531.65 665.95 C 61.69 424.82 533.62 665.95 R 7 X 0 K V 120.54 656.14 131.74 659.91 R V 3 H 2 Z N 359.34 547.99 373.74 562.7 R V N 93.64 646.53 119.64 657.13 R V 1 12 Q 0 X (error) 93.64 649.13 T 339.47 466.28 374.97 476.22 R 7 X V 0 X (cycles) 339.47 468.22 T 2 F (Fig. 5:) 389.77 503.16 T 1 F (Hyphenation error) 432.29 503.16 T (for the test base W) 389.77 489.16 T 1 10 Q (t) 479.69 486.16 T 1 12 Q ( after dif-) 482.47 489.16 T -0.22 (ferent numbers of training cy-) 389.77 475.16 P (cles with W) 389.77 461.16 T 1 10 Q (1000) 447.06 458.16 T 90.28 462.06 90.28 653.66 2 L 1 H N 90.28 462.06 373.74 462.06 2 L N 90.28 462.06 90.28 457.55 2 L N 1 12 Q (0) 87.29 445.71 T 112.96 462.06 112.96 457.55 2 L N (10) 106.96 445.71 T 135.64 462.06 135.64 457.55 2 L N (20) 129.64 445.71 T 158.31 462.06 158.31 457.55 2 L N (30) 152.32 445.71 T 180.99 462.06 180.99 457.55 2 L N (40) 174.99 445.71 T 203.67 462.06 203.67 457.55 2 L N (50) 197.67 445.71 T 226.34 462.06 226.34 457.55 2 L N (60) 220.35 445.71 T 249.02 462.06 249.02 457.55 2 L N (70) 243.02 445.71 T 271.7 462.06 271.7 457.55 2 L N (80) 265.7 445.71 T 294.38 462.06 294.38 457.55 2 L N (90) 288.38 445.71 T 317.05 462.06 317.05 457.55 2 L N (100) 308.06 445.71 T 339.73 462.06 339.73 457.55 2 L N (110) 330.73 445.71 T 362.4 462.06 362.4 457.55 2 L N (120) 353.41 445.71 T 90.28 462.06 85.75 462.06 2 L N (0) 77.48 461.49 T 90.28 484.6 85.75 484.6 2 L N (10) 71.49 484.03 T 90.28 507.14 85.75 507.14 2 L N (20) 71.49 506.57 T 90.28 529.68 85.75 529.68 2 L N (30) 71.49 529.11 T 90.28 552.22 85.75 552.22 2 L N (40) 71.49 551.65 T 90.28 574.76 85.75 574.76 2 L N (50) 71.49 574.19 T 90.28 597.3 85.75 597.3 2 L N (60) 71.49 596.73 T 90.28 619.84 85.75 619.84 2 L N (70) 71.49 619.27 T 90.28 642.38 85.75 642.38 2 L N (80) 71.49 641.81 T 101.62 619.84 112.96 606.32 124.3 558.98 135.64 545.46 146.98 549.97 158.31 552.22 169.65 547.71 180.99 552.22 192.33 558.98 203.67 556.73 215.01 552.22 226.34 554.47 237.68 556.73 249.02 556.73 260.36 552.22 271.7 549.97 283.04 552.22 294.38 556.73 305.71 554.47 317.05 556.73 328.39 554.47 339.73 554.47 351.07 556.73 362.4 561.24 24 L 3 H 2 X N 131.1 655.68 M 133.37 648.02 E 135.64 626.38 E 137.9 621.42 E 140.17 611.28 E 142.44 610.15 E 144.71 603.39 E 146.98 599.56 E 149.24 597.53 E 151.51 597.98 E 153.78 593.24 E 156.05 595.5 E 158.31 597.75 E 160.58 593.24 E 162.85 589.86 E 165.12 587.39 E 167.38 588.06 E 169.65 585.81 E 171.92 584.68 E 174.19 580.85 E 176.46 580.62 E 178.72 582.88 E 180.99 578.59 E 183.26 577.47 E 185.53 584 E 187.79 584.91 E 190.06 577.47 E 192.33 573.41 E 194.6 577.92 E 196.87 574.76 E 199.13 572.51 E 201.4 569.8 E 203.67 569.58 E 205.93 570.93 E 208.2 568.22 E 210.47 568.9 E 212.74 567.55 E 215.01 565.97 E 217.27 569.35 E 219.54 569.58 E 221.81 564.84 E 224.08 570.48 E 226.34 564.62 E 228.61 565.97 E 230.88 567.32 E 233.15 565.97 E 235.41 565.07 E 237.68 565.3 E 239.95 564.62 E 242.22 564.17 E 244.49 564.17 E 246.75 564.62 E 249.02 565.74 E 251.29 565.3 E 253.56 563.94 E 255.82 565.3 E 258.09 564.84 E 260.36 564.39 E 262.63 565.3 E 264.89 563.72 E 267.16 564.62 E 269.43 565.3 E 271.7 564.39 E 273.97 565.52 E 276.23 565.52 E 278.5 563.94 E 280.77 563.94 E 283.04 565.3 E 285.3 563.49 E 287.57 563.94 E 289.84 564.39 E 292.11 564.17 E 294.38 564.62 E 296.64 564.17 E 298.91 563.49 E 301.18 564.39 E 303.45 563.72 E 305.71 564.39 E 307.98 563.72 E 310.25 563.72 E 312.52 563.49 E 314.78 561.69 E 317.05 562.14 E 319.32 560.34 E 321.59 559.89 E 323.86 559.89 E 326.12 558.98 E 328.39 558.98 E 330.66 558.31 E 332.93 556.95 E 335.19 557.63 E 337.46 557.18 E 339.73 557.63 E 342 557.63 E 344.26 553.35 E 346.53 555.15 E 348.8 553.35 E 351.07 554.02 E 353.33 552.45 E 355.6 552.9 E 357.87 554.25 E 360.14 552.67 E 362.4 552.22 E 364.67 552.67 E 366.94 550.64 E 369.21 550.64 E 371.48 549.52 E 373.74 550.64 E 376.01 550.19 E 378.28 551.77 E 1 H 0 X N 264.51 608.15 437.1 641.12 R 7 X V 0 X (total square error of W) 264.51 633.12 T 1 10 Q (1000) 373.75 630.12 T 1 12 Q (incorrectly hyphenated positions) 264.51 619.12 T 237.65 635.53 258.95 635.53 2 L 0.5 H N 237.65 621.7 258.95 621.7 2 L 3 H 2 X N 364.15 546.2 382.65 556.45 R 7 X V N 63.65 69.95 531.65 789.95 C -8.35 24.95 603.65 816.95 C 61.68 142.91 533.62 363.82 C 61.68 142.91 533.62 363.82 R 7 X 0 K V 359.55 150.24 529.75 210.9 R V 2 12 Q 0 X (Fig. 6:) 386.76 202.9 T 1 F (Effect of the hidden) 429.28 202.9 T -0.11 (layer size on the hyphenation) 386.76 188.9 P (error for the test base W) 386.76 174.9 T 1 10 Q (t) 502.99 171.9 T 110.08 165.65 110.08 347.71 2 L 1 H 2 Z N 110.08 165.65 322.67 165.65 2 L N 110.08 165.65 110.08 161.37 2 L N 1 12 Q (0) 107.08 149.98 T 138.42 165.65 138.42 161.37 2 L N (10) 132.43 149.98 T 166.77 165.65 166.77 161.37 2 L N (20) 160.77 149.98 T 195.12 165.65 195.12 161.37 2 L N (30) 189.12 149.98 T 223.46 165.65 223.46 161.37 2 L N (40) 217.46 149.98 T 251.81 165.65 251.81 161.37 2 L N (50) 245.81 149.98 T 280.15 165.65 280.15 161.37 2 L N (60) 274.16 149.98 T 308.5 165.65 308.5 161.37 2 L N (70) 302.5 149.98 T 110.08 165.65 104.41 165.65 2 L N (0) 95.57 164.97 T 110.08 187.07 104.41 187.07 2 L N (10) 89.58 186.39 T 110.08 208.49 104.41 208.49 2 L N (20) 89.58 207.81 T 110.08 229.91 104.41 229.91 2 L N (30) 89.58 229.23 T 110.08 251.32 104.41 251.32 2 L N (40) 89.58 250.64 T 110.08 272.74 104.41 272.74 2 L N (50) 89.58 272.06 T 110.08 294.16 104.41 294.16 2 L N (60) 89.58 293.48 T 110.08 315.58 104.41 315.58 2 L N (70) 89.58 314.9 T 110.08 337 104.41 337 2 L N (80) 89.58 336.32 T 180.94 346.85 183.78 339.78 186.61 331.86 189.45 327.15 192.28 331.43 195.12 332.29 197.95 324.15 200.78 325.43 203.62 322.22 206.45 318.58 209.29 313.44 212.12 310.44 214.96 316.86 217.79 310.01 220.63 299.52 223.46 305.08 226.3 302.51 229.13 297.8 231.96 296.52 234.8 303.16 237.63 295.23 240.47 298.44 243.3 295.02 246.14 289.88 248.97 292.02 251.81 288.38 254.64 290.09 257.48 290.3 260.31 287.09 263.15 286.45 265.98 285.81 268.82 286.66 271.65 284.95 274.48 285.59 277.32 287.73 280.15 285.59 282.99 284.95 285.82 288.38 288.66 286.45 291.49 286.66 294.33 285.59 297.16 286.24 300 286.88 302.83 285.59 305.67 285.81 308.5 286.66 311.33 284.74 314.17 287.31 317.01 285.59 319.84 285.81 322.67 284.52 325.51 286.02 328.34 284.52 331.18 286.02 54 L N 161.1 349.63 163.93 342.35 166.77 321.79 169.6 317.08 172.44 307.44 175.27 306.37 178.11 299.94 180.94 296.3 183.78 294.38 186.61 294.8 189.45 290.3 192.28 292.45 195.12 294.59 197.95 290.3 200.78 287.09 203.62 284.74 206.45 285.38 209.29 283.24 212.12 282.17 214.96 278.53 217.79 278.31 220.63 280.45 223.46 276.38 226.3 275.31 229.13 281.52 231.96 282.38 234.8 275.31 237.63 271.46 240.47 275.74 243.3 272.74 246.14 270.6 248.97 268.03 251.81 267.82 254.64 269.1 257.48 266.53 260.31 267.17 263.15 265.89 265.98 264.39 268.82 267.6 271.65 267.82 274.48 263.32 277.32 268.67 280.15 263.1 282.99 264.39 285.82 265.67 288.66 264.39 291.49 263.53 294.33 263.75 297.16 263.1 300 262.68 302.83 262.68 305.67 263.1 308.5 264.17 311.33 263.75 314.17 262.46 317.01 263.75 319.84 263.32 322.67 262.89 325.51 263.75 328.34 262.25 331.18 263.1 61 L N 163.93 349.42 166.77 333.78 169.6 325 172.44 308.94 175.27 304.66 178.11 300.59 180.94 289.45 183.78 284.95 186.61 286.88 189.45 283.67 192.28 278.1 195.12 275.53 197.95 274.46 200.78 276.38 203.62 272.96 206.45 269.96 209.29 270.82 212.12 267.82 214.96 269.1 217.79 269.96 220.63 266.96 223.46 266.53 226.3 268.46 229.13 266.53 231.96 268.24 234.8 266.1 237.63 266.53 240.47 264.82 243.3 263.53 246.14 264.82 248.97 264.82 251.81 263.53 254.64 262.25 257.48 261.82 260.31 260.53 263.15 259.68 265.98 261.39 268.82 257.75 271.65 258.39 274.48 257.96 277.32 256.68 280.15 257.11 282.99 258.61 285.82 257.75 288.66 257.54 291.49 258.61 294.33 257.54 297.16 256.47 300 257.11 302.83 257.11 305.67 257.96 308.5 256.89 311.33 257.54 314.17 257.75 317.01 257.54 319.84 256.25 322.67 256.25 325.51 254.54 328.34 254.97 331.18 255.18 60 L N 286.74 167.22 322.74 176.29 R 7 X V 0 X (cycles) 286.74 168.29 T 112.74 337.23 139.74 347.05 R 7 X V 0 X (error) 112.74 339.05 T (10 Hidden Units) 197.74 327.67 T (40 Hidden Units) 248.57 271.21 T (80 Hidden Units) 236.07 244.71 T 342.78 309.88 515.37 327.85 R 7 X V 0 X (incorrectly hyphenated positions) 342.78 319.85 T -8.35 24.95 603.65 816.95 C FMENDPAGE %%EndPage: "5" 6 %%Page: "6" 6 595.3 841.89 0 FMBEGINPAGE 63.65 797.95 531.65 807.95 R 7 X 0 K V 63.65 57.61 531.65 67.61 R V 63.65 69.95 531.65 789.95 R V 1 12 Q 0 X 0.34 (network hyphenated 1024 of 1060 test patterns correctly) 63.65 781.95 P 0.34 (. This is 96.6% in contrary to 96.04% of) 336.05 781.95 P (the network with 40 hidden units. So the use of lar) 63.65 767.95 T (ger networks might still improve our results.) 305.62 767.95 T (On the other hand too many hidden units usually lead to bad generalization:) 63.65 753.95 T -0.14 (The training base is learned \322by heart\323 and there is no need for finding efficient representations of) 63.65 739.95 P 0.04 (the problem. So the increase of the hidden layer beyond a certain point is likely to worsen the hy-) 63.65 725.95 P (phenation ability for new words.) 63.65 711.95 T 2 14 Q (6.) 63.65 684.61 T ( Conclusions) 76.69 684.61 T 1 12 Q -0.1 (We applied a sequential neural network with 40 hidden units to the problem of hyphenation in the) 63.65 664.95 P (German language.) 63.65 650.95 T 1.29 (In the training phase the network was able to learn 99.3% of the 5793 training patterns which) 63.65 636.95 P (were formed from 1000 German words.) 63.65 622.95 T 0.25 (In the test phase the network was able to determine correctly the hyphenation of 96.04% of 1060) 63.65 608.95 P -0.18 (patterns formed from 200 German words. During the simulations we made the following observa-) 63.65 594.95 P (tions:) 63.65 580.95 T (\245) 82.66 561.95 T (Increase of the training base leads to better hyphenation of new words.) 100.66 561.95 T (\245) 82.66 542.95 T (Increase of the training time leads to better learning of the training patterns, but after a) 100.66 542.95 T (certain point the hyphenation of new words does not improve anymore.) 100.66 528.95 T (\245) 82.66 509.95 T -0.38 (Increase of the hidden layer size can improve both the learning of training patterns and the) 100.66 509.95 P -0.28 (hyphenation ability for new words. \05096.8% correct hyphenation of positions in new words) 100.66 495.95 P (with 80 units instead of 96.04% with 40\051) 100.66 481.95 T 1.14 (So hyphenation with neural networks seems possible. In order to improve our results the word) 63.65 462.95 P (bases used for training should still be larger than the one we used.) 63.65 448.95 T 0.05 (One interesting aspect of this approach is the fact that hyphenation for some other languages than) 63.65 434.95 P (German can be done by simply exchanging the training base.) 63.65 420.95 T 2 14 Q (Refer) 63.65 399.61 T (ences) 96.8 399.61 T 1 12 Q ([1]) 76.69 379.95 T (Jordan, M. I., \322) 105.04 379.95 T 3 F (Attractor dynamics and parallelism in a connectionist sequential ma-) 177.98 379.95 T (chine) 105.04 365.95 T 1 F (,\323 Proceedings of the Eighth Annual Meeting of the Cognitive Science Society,) 131.01 365.95 T (Hillsdale, NJ,1986) 105.04 351.95 T ([2]) 76.69 332.95 T (Nguyen, D. and B.Widrow, \323) 105.04 332.95 T 3 F (The Truck Backer-Upper: An example of self-learning in) 245.6 332.95 T (neural networks) 105.04 318.95 T 1 F (\323 Proceedings of the IJCNN, pp.II-357, Washington D.C., 1989) 182.66 318.95 T ([3]) 76.69 299.95 T (Plickat, H.,\323) 105.04 299.95 T 3 F (Deutscher Grundwortschatz) 164.33 299.95 T 1 F (\323, Beltz Verlag, Weinheim, 1986) 299.92 299.95 T ([4]) 76.69 280.95 T (Rumelhart, D.E., G.E. Hinton, and R.J. Williams, \322) 105.04 280.95 T 3 F (Learning internal representation by) 351.23 280.95 T -0.31 (error propagation) 105.04 266.95 P 1 F -0.31 (\323 in Parallel Distributed Processing, vol.1, ed. Rumelhart, McClelland,) 192.35 266.95 P (pp. 675-695, MIT Press, San Diego, 1986) 105.04 252.95 T ([5]) 76.69 233.95 T (Sejnowski, T. J. and C. R. Rosenberg, \322) 105.04 233.95 T 3 F (Parallel networks that learn to pronounce En-) 295.92 233.95 T (glish text) 105.04 219.95 T 1 F (\323 Complex Systems, vol. 1, pp. 145-168, 1987) 148.68 219.95 T ([6]) 76.69 200.95 T (Tesauro, G. and T. Sejnowski, \322) 105.04 200.95 T 3 F (A \324neural\325 network that learns to play backgammon) 259.26 200.95 T 1 F (\323) 508.1 200.95 T (First IEEE Conf. on Neural Information Processing Systems, pp. 794-803, Denver CO,) 105.04 186.95 T (1988) 105.04 172.95 T FMENDPAGE %%EndPage: "6" 7 %%Trailer %%BoundingBox: 0 0 595.3 841.89 %%Pages: 6 1 %%DocumentFonts: AvantGarde-Demi %%+ Times-Roman %%+ Times-Bold %%+ Times-Italic %%+ Symbol