1: %!PS-Adobe-3.0 XDRT

2: %%XRXbegin

3: %%OID_ATT_JOB_TYPE OID_VAL_JOB_TYPE_PRINT;

4: %%OID_ATT_START_SHEET OID_VAL_JOB_SHEET_FULL;

5: %%OID_ATT_PRINT_QUALITY OID_VAL_PRINT_QUALITY_NORMAL;

6: %%XRXend

7: %!PS-Adobe-3.0

8: %%Title: Microsoft Word - ICASSP-submission-8.doc

9: %%Creator: Pscript.dll Version 5.0

10: %%CreationDate: 11/15/2000 13:27:38

11: %%BoundingBox: (atend)

12: %%Pages: (atend)

13: %%Orientation: Portrait

14: %%PageOrder: Special

15: %%DocumentNeededResources: (atend)

16: %%DocumentSuppliedResources: (atend)

17: %%DocumentData: Clean7Bit

18: %%TargetDevice: (Xerox DC 240/255/265 PS2) (2016.108) 13

19: %%LanguageLevel: 2

20: %%EndComments

21: 

22: %%BeginDefaults

23: %%PageBoundingBox: 12 12 600 780

24: %%EndDefaults

25: 

26: 

27: %%BeginProlog

28: %%BeginResource: file Pscript_WinNT_ErrorHandler 5.0 0

29: /currentpacking where{pop/oldpack currentpacking def/setpacking where{pop false

30: setpacking}if}if/$brkpage 64 dict def $brkpage begin/prnt{dup type/stringtype

31: ne{=string cvs}if dup length 6 mul/tx exch def/ty 10 def currentpoint/toy exch

32: def/tox exch def 1 setgray newpath tox toy 2 sub moveto 0 ty rlineto tx 0

33: rlineto 0 ty neg rlineto closepath fill tox toy moveto 0 setgray show}bind def

34: /nl{currentpoint exch pop lmargin exch moveto 0 -10 rmoveto}def/=={/cp 0 def

35: typeprint nl}def/typeprint{dup type exec}readonly def/lmargin 72 def/rmargin 72

36: def/tprint{dup length cp add rmargin gt{nl/cp 0 def}if dup length cp add/cp

37: exch def prnt}readonly def/cvsprint{=string cvs tprint( )tprint}readonly def

38: /integertype{cvsprint}readonly def/realtype{cvsprint}readonly def/booleantype

39: {cvsprint}readonly def/operatortype{(--)tprint =string cvs tprint(-- )tprint}

40: readonly def/marktype{pop(-mark- )tprint}readonly def/dicttype{pop

41: (-dictionary- )tprint}readonly def/nulltype{pop(-null- )tprint}readonly def

42: /filetype{pop(-filestream- )tprint}readonly def/savetype{pop(-savelevel- )

43: tprint}readonly def/fonttype{pop(-fontid- )tprint}readonly def/nametype{dup

44: xcheck not{(/)tprint}if cvsprint}readonly def/stringtype{dup rcheck{(\()tprint

45: tprint(\))tprint}{pop(-string- )tprint}ifelse}readonly def/arraytype{dup rcheck

46: {dup xcheck{({)tprint{typeprint}forall(})tprint}{([)tprint{typeprint}forall(])

47: tprint}ifelse}{pop(-array- )tprint}ifelse}readonly def/packedarraytype{dup

48: rcheck{dup xcheck{({)tprint{typeprint}forall(})tprint}{([)tprint{typeprint}

49: forall(])tprint}ifelse}{pop(-packedarray- )tprint}ifelse}readonly def/courier

50: /Courier findfont 10 scalefont def end errordict/handleerror{systemdict begin

51: $error begin $brkpage begin newerror{/newerror false store vmstatus pop pop 0

52: ne{grestoreall}if showpage initgraphics courier setfont lmargin 720 moveto

53: errorname(VMError)eq{PrtVMMsg}{(ERROR: )prnt errorname prnt nl

54: (OFFENDING COMMAND: )prnt/command load prnt $error/ostack known{nl nl(STACK:)

55: prnt nl nl $error/ostack get aload length{==}repeat}if}ifelse systemdict

56: /showpage get exec(%%[ Error: )print errorname =print(; OffendingCommand: )

57: print/command load =print( ]%%)= flush}if end end end}dup 0 systemdict put dup

58: 4 $brkpage put bind readonly put/currentpacking where{pop/setpacking where{pop

59: oldpack setpacking}if}if

60: %%EndResource

61: userdict /Pscript_WinNT_Incr 230 dict dup begin put

62: %%BeginResource: file Pscript_FatalError 5.0 0

63: /FatalErrorIf{{initgraphics findfont 1 index 0 eq{exch pop}{dup length dict

64: begin{1 index/FID ne{def}{pop pop}ifelse}forall/Encoding{ISOLatin1Encoding}

65: stopped{StandardEncoding}if def currentdict end/ErrFont-Latin1 exch definefont}

66: ifelse exch scalefont setfont counttomark 3 div cvi{moveto show}repeat showpage

67: quit}{cleartomark}ifelse}bind def

68: %%EndResource

69: userdict begin/PrtVMMsg{vmstatus exch sub exch pop gt{[

70: (This job requires more memory than is available in this printer.)100 500

71: (Try one or more of the following, and then print again:)100 485

72: (For the output format, choose Optimize For Portability.)115 470

73: (In the Device Settings page, make sure the Available PostScript Memory is accurate.)

74: 115 455(Reduce the number of fonts in the document.)115 440

75: (Print the document in parts.)115 425 12/Times-Roman showpage

76: (%%[ PrinterError: Low Printer VM ]%%)= true FatalErrorIf}if}bind def end

77: version cvi 2016 ge{/VM?{pop}bind def}{/VM? userdict/PrtVMMsg get def}ifelse

78: 105000 VM?

79: %%BeginResource: file Pscript_Win_Basic 5.0 0

80: /d/def load def/,/load load d/~/exch , d/?/ifelse , d/!/pop , d/`/begin , d/^

81: /index , d/@/dup , d/+/translate , d/$/roll , d/U/userdict , d/M/moveto , d/-

82: /rlineto , d/&/currentdict , d/:/gsave , d/;/grestore , d/F/false , d/T/true ,

83: d/N/newpath , d/E/end , d/Ac/arc , d/An/arcn , d/A/ashow , d/D/awidthshow , d/C

84: /closepath , d/V/div , d/O/eofill , d/L/fill , d/I/lineto , d/-c/curveto , d/-M

85: /rmoveto , d/+S/scale , d/Ji/setfont , d/Lc/setlinecap , d/Lj/setlinejoin , d

86: /Lw/setlinewidth , d/Lm/setmiterlimit , d/sd/setdash , d/S/show , d/LH/showpage

87: , d/K/stroke , d/W/widthshow , d/R/rotate , d/L2? false/languagelevel where{pop

88: languagelevel 2 ge{pop true}if}if d L2?{/xS/xshow , d/yS/yshow , d/zS/xyshow ,

89: d}if/b{bind d}bind d/bd{bind d}bind d/xd{~ d}bd/ld{, d}bd/bn/bind ld/lw/Lw ld

90: /lc/Lc ld/lj/Lj ld/sg/setgray ld/ADO_mxRot null d/self & d/OrgMx matrix

91: currentmatrix d/reinitialize{: OrgMx setmatrix[/TextInit/GraphInit/UtilsInit

92: counttomark{@ where{self eq}{F}?{cvx exec}{!}?}repeat cleartomark ;}b

93: /initialize{`{/ADO_mxRot ~ d/TextInitialised? F d reinitialize E}{U

94: /Pscript_Win_Data 230 dict @ ` put/ADO_mxRot ~ d/TextInitialised? F d

95: reinitialize}?}b/terminate{!{& self eq{exit}{E}?}loop E}b/suspend/terminate , d

96: /resume{` Pscript_Win_Data `}b U `/lucas 21690 d/featurebegin{countdictstack

97: lucas[}b/featurecleanup{stopped{cleartomark @ lucas eq{! exit}if}loop

98: countdictstack ~ sub @ 0 gt{{E}repeat}{!}?}b E/snap{transform 0.25 sub round

99: 0.25 add ~ 0.25 sub round 0.25 add ~ itransform}b/dsnap{dtransform round ~

100: round ~ idtransform}b/nonzero_round{@ 0.5 ge{round}{@ -0.5 lt{round}{0 ge{1}

101: {-1}?}?}?}b/nonzero_dsnap{dtransform nonzero_round ~ nonzero_round ~

102: idtransform}b U<04>cvn{}put/rr{1 ^ 0 - 0 ~ - neg 0 - C}b/irp{4 -2 $ + +S fx 4 2

103: $ M 1 ^ 0 - 0 ~ - neg 0 -}b/rp{4 2 $ M 1 ^ 0 - 0 ~ - neg 0 -}b/solid{[]0 sd}b/g

104: {@ not{U/DefIf_save save put}if U/DefIf_bool 2 ^ put}b/DefIf_El{if U/DefIf_bool

105: get not @{U/DefIf_save get restore}if}b/e{DefIf_El !}b/UDF{L2?{undefinefont}{!}

106: ?}b/UDR{L2?{undefineresource}{! !}?}b/freeVM{/Courier findfont[40 0 0 -40 0 0]

107: makefont Ji 2 vmreclaim}b

108: %%EndResource

109: %%BeginResource: file Pscript_Win_Utils_L2 5.0 0

110: /rf/rectfill , d/fx{1 1 dtransform @ 0 ge{1 sub 0.5}{1 add -0.5}? 3 -1 $ @ 0 ge

111: {1 sub 0.5}{1 add -0.5}? 3 1 $ 4 1 $ idtransform 4 -2 $ idtransform}b/BZ{4 -2 $

112: snap + +S fx rf}b/rs/rectstroke , d/rc/rectclip , d/UtilsInit{currentglobal{F

113: setglobal}if}b/scol{! setcolor}b/colspA/DeviceGray d/colspABC/DeviceRGB d

114: /colspRefresh{colspABC setcolorspace}b/SetColSpace{colspABC setcolorspace}b

115: /resourcestatus where{!/ColorRendering/ProcSet resourcestatus{! ! T}{F}?}{F}?

116: not{/ColorRendering<</GetHalftoneName{currenthalftone @/HalftoneName known{

117: /HalftoneName get}{!/none}?}bn/GetPageDeviceName{currentpagedevice @

118: /PageDeviceName known{/PageDeviceName get @ null eq{!/none}if}{!/none}?}bn

119: /GetSubstituteCRD{!/DefaultColorRendering/ColorRendering resourcestatus{! !

120: /DefaultColorRendering}{(DefaultColorRendering*){cvn exit}127 string

121: /ColorRendering resourceforall}?}bn>>/defineresource where{!/ProcSet

122: defineresource !}{! !}?}if/buildcrdname{/ColorRendering/ProcSet findresource `

123: mark GetHalftoneName @ type @/nametype ne ~/stringtype ne and{!/none}if(.)

124: GetPageDeviceName @ type @/nametype ne ~/stringtype ne and{!/none}if(.)5 ^ 0 5

125: -1 1{^ length add}for string 6 1 $ 5 ^ 5{~ 1 ^ cvs length 1 ^ length 1 ^ sub

126: getinterval}repeat ! cvn 3 1 $ ! ! E}b/definecolorrendering{~ buildcrdname ~

127: /ColorRendering defineresource !}b/findcolorrendering where{!}{

128: /findcolorrendering{buildcrdname @/ColorRendering resourcestatus{! ! T}{

129: /ColorRendering/ProcSet findresource ` GetSubstituteCRD E F}?}b}?

130: /selectcolorrendering{findcolorrendering !/ColorRendering findresource

131: setcolorrendering}b

132: %%EndResource

133: end

134: %%EndProlog

135: 

136: %%BeginSetup

137: statusdict begin (%%[ ProductName: ) print product print ( ]%%)= flush end

138: [ 1 0 0 1 0 0 ] false Pscript_WinNT_Incr dup /initialize get exec

139: featurebegin{

140: %%BeginNonPPDFeature: JobTimeout 0

141: 0 /languagelevel where{pop languagelevel}{1}ifelse 2 ge{1 dict dup/JobTimeout  4 -1 roll put setuserparams}{statusdict/setjobtimeout get exec}ifelse

142: %%EndNonPPDFeature

143: }featurecleanup

144: 

145: featurebegin{

146: %%BeginNonPPDFeature: WaitTimeout 300

147: 300 /languagelevel where{pop languagelevel}{1}ifelse 2 ge{1 dict dup/WaitTimeout 4 -1 roll put setuserparams}{statusdict/waittimeout 3 -1 roll put}ifelse

148: %%EndNonPPDFeature

149: }featurecleanup

150: 

151: featurebegin{

152: %%BeginFeature: *Resolution 600x600dpi

153: 

154:         (<<) cvx exec /DeviceRenderingInfo (<<) cvx exec /Type 4 /ValuesPerColorComponent 2(>>) cvx exec /HWResolution [600 600] (>>) cvx exec setpagedevice

155: %%EndFeature

156: }featurecleanup

157: featurebegin{

158: %%BeginFeature: *StapleLocation None

159: 

160:         (<<) cvx exec /Staple 0(>>) cvx exec setpagedevice

161: %%EndFeature

162: }featurecleanup

163: featurebegin{

164: %%BeginFeature: *Jog EndOfSet

165: 

166:       (<<) cvx exec /Jog 3 (>>) cvx exec setpagedevice

167: %%EndFeature

168: }featurecleanup

169: featurebegin{

170: %%BeginFeature: *XRImageQuality 4

171: 

172:         (<<) cvx exec /PostRenderingEnhance false (>>) cvx exec setpagedevice

173: %%EndFeature

174: }featurecleanup

175: featurebegin{

176: %%BeginFeature: *MediaType SystemDefault

177: 

178:         (<<) cvx exec /MediaType (System-Default) (>>) cvx exec setpagedevice

179: %%EndFeature

180: }featurecleanup

181: featurebegin{

182: %%BeginFeature: *MediaColor White

183: 

184:         (<<) cvx exec /MediaColor (White) (>>) cvx exec setpagedevice

185: %%EndFeature

186: }featurecleanup

187: featurebegin{

188: %%BeginFeature: *PageSize Letter

189: 

190:     (<<) cvx exec /PageSize [612 792] /ImagingBBox null (>>) cvx exec setpagedevice

191: %%EndFeature

192: }featurecleanup

193: featurebegin{

194: %%BeginFeature: *Duplex DuplexNoTumble

195: 

196:         (<<) cvx exec /Duplex true /Tumble false (>>) cvx exec setpagedevice

197: %%EndFeature

198: }featurecleanup

199: featurebegin{

200: %%BeginFeature: *XrxBanner True

201: 

202: %%EndFeature

203: }featurecleanup

204: featurebegin{

205: %%BeginFeature: *XrxDraft False

206: 

207: %%EndFeature

208: }featurecleanup

209: featurebegin{

210: %%BeginFeature: *Collate True

211: 

212:       << /Collate true >> setpagedevice

213: %%EndFeature

214: }featurecleanup

215: featurebegin{

216: %%BeginFeature: *Option1 True

217:  

218: %%EndFeature

219: }featurecleanup

220: featurebegin{

221: %%BeginFeature: *Option2 True

222:  

223: %%EndFeature

224: }featurecleanup

225: featurebegin{

226: %%BeginFeature: *Option3 True

227:  

228: %%EndFeature

229: }featurecleanup

230: featurebegin{

231: %%BeginFeature: *InstalledMemory None

232: 

233: %%EndFeature

234: }featurecleanup

235: 1 setlinecap 1 setlinejoin

236: /mysetup [ 0.11999 0 0 -0.11999 11.99905 780.00091 ] def 

237: %%EndSetup

238: 

239: %%Page: 1 1

240: %%PageBoundingBox: 12 12 600 780

241: %%EndPageComments

242: %%BeginPageSetup

243: gsave %matrix defaultmatrix setmatrix

244: 90 rotate 4 72 mul .55 -72 mul moveto /Times-Roman findfont

245: 20 scalefont setfont 0.3 setgray (arXiv:cs.CL/0108006   9 Aug 2001) show grestore

246: /DeviceRGB dup setcolorspace /colspABC exch def

247: mysetup concat colspRefresh

248: %%EndPageSetup

249: 

250: /DeviceGray dup setcolorspace /colspABC exch def

251: 0 0 scol Pscript_WinNT_Incr begin

252: %%BeginResource: file Pscript_Text 5.0 0

253: /TextInit{TextInitialised? not{/Pscript_Windows_Font & d/TextInitialised? T d

254: /fM[1 0 0 1 0 0]d/mFM matrix d/iMat[1 0 0.212557 1 0 0]d}if}b/copyfont{1 ^

255: length add dict `{1 ^/FID ne{d}{! !}?}forall & E}b/EncodeDict 11 dict d/bullets

256: {{/bullet}repeat}b/rF{3 copyfont @ ` ~ EncodeDict ~ get/Encoding ~ 3 ^/0 eq{&

257: /CharStrings known{CharStrings/Eth known not{! EncodeDict/ANSIEncodingOld get}

258: if}if}if d E}b/mF{@ 7 1 $ findfont ~{@/Encoding get @ StandardEncoding eq{! T}{

259: {ISOLatin1Encoding}stopped{! F}{eq}?{T}{@ ` T 32 1 127{Encoding 1 ^ get

260: StandardEncoding 3 -1 $ get eq and}for E}?}?}{F}?{1 ^ ~ rF}{0 copyfont}? 6 -2 $

261: ! ! ~ !/pd_charset @ where{~ get 128 eq{@ FDV 2 copy get @ length array copy

262: put pd_CoverFCRange}if}{!}? 2 ^ ~ definefont fM 5 4 -1 $ put fM 4 0 put fM

263: makefont Pscript_Windows_Font 3 1 $ put}b/sLT{: Lw -M currentpoint snap M 0 - 0

264: Lc K ;}b/xUP null d/yUP null d/uW null d/xSP null d/ySP null d/sW null d/sSU{N

265: /uW ~ d/yUP ~ d/xUP ~ d}b/sU{xUP yUP uW sLT}b/sST{N/sW ~ d/ySP ~ d/xSP ~ d}b/sT

266: {xSP ySP sW sLT}b/sR{: + R 0 0 M}b/sRxy{: matrix astore concat 0 0 M}b/eR/; , d

267: /AddOrigFP{{&/FontInfo known{&/FontInfo get length 6 add}{6}? dict `

268: /WinPitchAndFamily ~ d/WinCharSet ~ d/OrigFontType ~ d/OrigFontStyle ~ d

269: /OrigFontName ~ d & E/FontInfo ~ d}{! ! ! ! !}?}b/mFS{makefont

270: Pscript_Windows_Font 3 1 $ put}b/mF42D{0 copyfont `/FontName ~ d 2 copy ~ sub 1

271: add dict `/.notdef 0 d 2 copy 1 ~{@ 3 ^ sub Encoding ~ get ~ d}for & E

272: /CharStrings ~ d ! ! & @ E/FontName get ~ definefont}b/mF42{15 dict ` @ 4 1 $

273: FontName ~ d/FontType 0 d/FMapType 2 d/FontMatrix[1 0 0 1 0 0]d 1 ^ 254 add 255

274: idiv @ array/Encoding ~ d 0 1 3 -1 $ 1 sub{@ Encoding 3 1 $ put}for/FDepVector

275: Encoding length array d/CharStrings 2 dict `/.notdef 0 d & E d 0 1 Encoding

276: length 1 sub{@ @ 10 lt{! FontName length 1 add string}{100 lt{FontName length 2

277: add string}{FontName length 3 add string}?}? @ 0 FontName @ length string cvs

278: putinterval @ 3 -1 $ @ 4 1 $ 3 string cvs FontName length ~ putinterval cvn 1 ^

279: 256 mul @ 255 add 3 -1 $ 4 ^ findfont mF42D FDepVector 3 1 $ put}for & @ E

280: /FontName get ~ definefont ! ! ! mF}b/mF_OTF_V{3{~ !}repeat 3 -1 $ ! findfont 1

281: ^ ~ definefont fM @ @ 5 0 put 4 0 put 90 matrix R matrix concatmatrix makefont

282: Pscript_Windows_Font 3 1 $ put}b/UmF{L2?{Pscript_Windows_Font ~ undef}{!}?}b

283: /UmF42{@ findfont/FDepVector get{/FontName get undefinefont}forall

284: undefinefont}b

285: %%EndResource

286: end reinitialize

287: Pscript_WinNT_Incr begin

288: %%BeginResource: file Pscript_Encoding256 5.0 0

289: /CharCol256Encoding[/.notdef/breve/caron/dotaccent/dotlessi/fi/fl/fraction

290: /hungarumlaut/Lslash/lslash/minus/ogonek/ring/Zcaron/zcaron/.notdef/.notdef

291: /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef

292: /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/exclam/quotedbl/numbersign

293: /dollar/percent/ampersand/quotesingle/parenleft/parenright/asterisk/plus/comma

294: /hyphen/period/slash/zero/one/two/three/four/five/six/seven/eight/nine/colon

295: /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

296: /T/U/V/W/X/Y/Z/bracketleft/backslash/bracketright/asciicircum/underscore/grave

297: /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

298: /asciitilde/.notdef/Euro/.notdef/quotesinglbase/florin/quotedblbase/ellipsis

299: /dagger/daggerdbl/circumflex/perthousand/Scaron/guilsinglleft/OE/.notdef

300: /.notdef/.notdef/.notdef/quoteleft/quoteright/quotedblleft/quotedblright/bullet

301: /endash/emdash/tilde/trademark/scaron/guilsinglright/oe/.notdef/.notdef

302: /Ydieresis/.notdef/exclamdown/cent/sterling/currency/yen/brokenbar/section

303: /dieresis/copyright/ordfeminine/guillemotleft/logicalnot/.notdef/registered

304: /macron/degree/plusminus/twosuperior/threesuperior/acute/mu/paragraph

305: /periodcentered/cedilla/onesuperior/ordmasculine/guillemotright/onequarter

306: /onehalf/threequarters/questiondown/Agrave/Aacute/Acircumflex/Atilde/Adieresis

307: /Aring/AE/Ccedilla/Egrave/Eacute/Ecircumflex/Edieresis/Igrave/Iacute

308: /Icircumflex/Idieresis/Eth/Ntilde/Ograve/Oacute/Ocircumflex/Otilde/Odieresis

309: /multiply/Oslash/Ugrave/Uacute/Ucircumflex/Udieresis/Yacute/Thorn/germandbls

310: /agrave/aacute/acircumflex/atilde/adieresis/aring/ae/ccedilla/egrave/eacute

311: /ecircumflex/edieresis/igrave/iacute/icircumflex/idieresis/eth/ntilde/ograve

312: /oacute/ocircumflex/otilde/odieresis/divide/oslash/ugrave/uacute/ucircumflex

313: /udieresis/yacute/thorn/ydieresis]def EncodeDict/256 CharCol256Encoding put

314: %%EndResource

315: end reinitialize

316: 

317: %%IncludeResource: font Palatino-Bold

318: F /F0 0 /256 T /Palatino-Bold mF 

319: /F0S75 F0 [117 0 0 -117 0 0 ] mFS

320: F0S75 Ji 

321: 868 609 M (CLASSES FOR FAST MAX)[84 71 91 71 71 71 71 29 65 97 84 29 65 91 71 78 29 117 91  0]xS 

322: 2322 609 M (IMUM ENTROPY TRAININ)[46 117 91 117 29 71 97 78 84 97 71 78 29 78 84 91 46 97 46  0]xS 

323: 3866 609 M (G)S 

324: 3963 609 M ( )S 

325: 

326: %%IncludeResource: font Times-Italic

327: F /F1 0 /256 T /Times-Italic mF 

328: /F1S5C F1 [92 0 0 -92 0 0 ] mFS

329: F1S5C Ji 

330: 2016 799 M (Joshua Goodman)[41 46 36 46 46 46 23 66 46 46 46 66 46  0]xS 

331: 2662 799 M ( )S 

332: 

333: %%IncludeResource: font Palatino-Roman

334: F /F2 0 /256 T /Palatino-Roman mF 

335: /F2S53 F2 [83 0 0 -83 0 0 ] mFS

336: F2S53 Ji 

337: 1987 951 M (Microsoft Research)[79 24 37 33 45 35 45 28 27 21 55 40 35 40 42 33 37  0]xS 

338: 2691 951 M ( )S 

339: 

340: %%IncludeResource: font Times-Roman

341: F /F3 0 /256 T /Times-Roman mF 

342: /F3S53 F3 [83 0 0 -83 0 0 ] mFS

343: F3S53 Ji 

344: 1738 1050 M (Redmond, Washington 98052, USA)[55 37 42 63 42 41 42 21 21 79 37 32 41 23 41 41 23 42 41 21 42 42 42 42 42 21 21 60 46  0]xS 

345: 2940 1050 M ( )S 

346: 0.105 0 scol F2S53 Ji 

347: 1881 1150 M (joshuago@microsoft.com)[19 45 35 48 50 42 46 45 62 73 24 37 33 45 35 45 28 27 21 37 45  0]xS 

348: 0 0 scol 2796 1150 M ( )S 

349: 1561 1254 M (http://research.microsoft.com/~joshuago)[48 27 27 50 21 50 50 33 40 35 40 42 33 37 48 21 73 24 37 33 45 35 45 28 27 21 37 45 73 50 50 19

350: 45 35 48 50 42 46  0]xS 

351: 3116 1254 M ( )S 

352: 

353: %%IncludeResource: font Times-Bold

354: F /F4 0 /256 T /Times-Bold mF 

355: /F4S64 F4 [100 0 0 -100 0 0 ] mFS

356: F4S64 Ji 

357: 1052 1520 M (ABSTRACT)[72 67 56 67 72 72 72  0]xS 

358: 1597 1520 M ( )S 

359: /F3S4B F3 [75 0 0 -75 0 0 ] mFS

360: F3S4B Ji 

361: 463 1661 M (Maximum entropy models are co)[67 33 37 21 57 38 57 24 33 38 21 25 38 38 36 24 57 38 38 33 21 29 24 33 25 33 24 33  0]xS 

362: 1476 1661 M (nsidered by many to be one )[38 29 21 38 33 25 33 38 24 38 36 23 57 33 38 36 23 21 38 23 38 33 23 38 38 33  0]xS 

363: 350 1747 M (of the most promising avenues of language modeling research.  )

364: [38 24 31 21 38 33 31 57 38 29 21 31 38 25 38 57 21 29 21 38 37 31 33 37 33 38 38 33 29 31 38 24

365: 31 21 33 38 37 38 33 37 33 31 57 38 38 33 21 21 38 37 30 25 33 29 33 33 25 33 38 19 19  0]xS 

366: 350 1833 M (Unfortunately, long training times make maximum entropy )[54 38 24 38 25 21 38 38 33 21 33 21 36 19 53 21 38 38 37 53 21 25 33 21 38 21 38 37 53 21 21 57

367: 33 29 53 57 33 37 33 53 57 33 37 21 57 38 57 52 33 38 21 25 38 38 36  0]xS 

368: 350 1919 M (research difficult.  We present a novel speedup technique: we )

369: [25 33 29 33 33 25 33 38 34 38 21 24 24 21 33 38 21 21 19 33 33 70 33 33 38 25 33 29 33 38 21 33

370: 33 33 38 38 37 33 21 33 29 38 33 33 38 38 38 33 21 33 33 38 38 21 38 38 33 21 33 53 33  0]xS 

371: 350 2006 M (change the form of the model to use classes.  Our)[33 38 33 38 37 33 22 21 38 33 22 24 38 25 57 21 38 24 21 21 38 33 21 57 38 38 33 21 21 21 38 21

372: 38 29 33 21 33 21 33 29 29 33 29 19 21 21 54 38  0]xS 

373: 1853 2006 M ( speedup works )[21 29 38 33 33 38 38 38 21 53 38 25 37 29  0]xS 

374: 350 2092 M (by creating two maximum entropy models, the first of which )[38 36 37 33 25 33 33 21 21 38 37 37 21 53 38 36 57 33 37 21 57 38 57 36 33 38 21 25 38 38 36 36

375: 57 38 38 33 21 29 19 36 21 38 33 36 24 21 25 29 21 36 38 24 36 53 38 21 33 38  0]xS 

376: 350 2178 M (predicts the class of each word, and the second of which predicts )

377: [38 25 33 38 21 33 21 29 22 21 38 33 22 33 21 33 29 29 22 38 24 21 33 33 33 38 21 53 38 25 38 19

378: 21 33 38 38 21 21 38 33 21 29 33 33 38 38 38 21 38 24 21 53 38 21 33 38 21 38 25 33 38 21 33 21

379: 29  0]xS 

380: 350 2264 M (the word itself.  This factoring of the model leads to fewer non)

381: [21 38 33 25 53 38 25 38 25 21 21 29 33 21 24 19 25 25 45 38 21 29 25 24 33 33 21 38 25 21 38 37

382: 25 38 24 25 21 38 33 25 57 38 38 33 21 24 21 33 33 38 29 24 21 38 24 24 33 53 33 25 24 38 38  0]xS 

383: 2299 2264 M (-)S 

384: 350 2350 M (zero indicator functions, and faster normalization,)[33 33 25 38 47 21 38 38 21 33 33 21 38 25 46 24 38 38 33 21 21 38 38 29 19 46 33 38 38 46 24 33

385: 29 21 33 25 46 38 38 25 57 33 21 21 33 33 21 21 38 38  0]xS 

386: 1987 2350 M ( achieving )[46 33 33 38 21 33 37 21 38 37  0]xS 

387: 350 2437 M (speedups of up to a factor of 35 over one of the best previous )

388: [29 38 33 33 38 38 38 29 30 38 24 30 38 38 30 21 38 29 33 29 24 33 33 21 38 25 29 38 24 29 38 38

389: 29 38 37 33 25 29 38 38 33 29 38 24 29 21 38 33 29 38 33 29 21 29 38 25 33 37 21 38 38 29  0]xS 

390: 350 2523 M (techniques.  It also results in typically slightly lower perplexities.  )

391: [21 33 33 38 38 21 38 38 33 29 19 21 21 25 21 21 33 21 29 38 21 25 33 29 38 21 21 29 21 21 38 21

392: 21 36 38 21 33 33 21 21 36 21 29 21 21 37 38 21 21 36 21 21 38 53 33 25 20 38 33 25 38 21 33 37

393: 21 21 21 33 29 19 19  0]xS 

394: 350 2609 M (The same trick can be used to speed training of other machine )

395: [45 38 33 30 29 33 57 33 30 21 25 21 33 37 30 33 33 38 29 38 33 29 38 29 33 38 29 21 38 29 29 38

396: 33 33 38 29 21 25 33 21 38 21 38 37 29 38 24 29 38 21 38 33 25 29 57 33 33 38 21 38 33  0]xS 

397: 350 2695 M (learning techniques, e.g. neural networks, appli)[21 33 33 25 38 21 38 37 20 21 33 33 38 38 21 38 38 33 29 19 20 33 19 37 19 20 38 33 38 25 33 21

398: 20 38 33 21 53 38 25 37 29 19 20 33 38 38 21  0]xS 

399: 1777 2695 M (ed to any problem )[33 38 20 21 38 20 33 38 36 20 38 25 38 38 21 33 57  0]xS 

400: 350 2782 M (with a large number of outputs, such as language modeling.)[53 21 21 38 19 33 19 21 33 25 37 33 19 38 38 57 38 33 25 19 38 24 19 38 38 21 38 38 21 29 19 19

401: 29 38 33 38 19 33 29 19 21 33 38 37 38 33 37 33 19 57 38 38 33 21 21 38 37  0]xS 

402: 2141 2782 M ( )S 

403: /F0S64 F0 [100 0 0 -100 0 0 ] mFS

404: F0S64 Ji 

405: 833 2944 M (1.)[50  0]xS 

406: 908 2944 M ( )S 

407: 968 2944 M (INTRODUCTION)[39 83 67 72 83 83 78 72 67 39 83  0]xS 

408: 1817 2944 M ( )S 

409: F3S4B Ji 

410: 350 3088 M (Maximum entropy models)[67 33 37 21 57 38 57 44 33 38 21 25 38 38 36 44 57 38 38 33 21  0]xS 

411: 1193 3088 M ( )S 

412: 1237 3088 M ([1])[26 38  0]xS 

413: 1327 3088 M ( are perhaps one of the most )[44 33 25 33 44 38 33 25 38 33 38 29 44 38 38 33 44 38 24 44 21 38 33 44 57 38 29 21  0]xS 

414: 350 3175 M (promising techniques for language model research.  )[38 25 38 57 21 29 21 38 37 52 21 33 33 38 38 21 38 38 33 29 51 24 38 25 51 21 33 38 37 38 33 37

415: 33 51 57 38 38 33 21 51 25 33 29 33 33 25 33 38 19 51  0]xS 

416: 2146 3175 M (The)[45 38  0]xS 

417: 2262 3175 M (se )[29 33  0]xS 

418: 350 3261 M (techniques)[21 33 33 38 38 21 38 38 33  0]xS 

419: 672 3261 M ( )S 

420: 695 3261 M (allow diverse sources)[33 21 21 38 53 23 38 21 37 33 25 29 33 23 29 38 38 25 33 33  0]xS 

421: 1348 3261 M ( of information to be combined.  )[23 38 24 23 21 38 24 38 25 57 33 21 21 38 38 23 21 38 23 38 33 23 33 38 57 38 21 38 33 38 19 19

422:  0]xS 

423: 350 3347 M (For each source of information, a set of constraints )[42 38 25 22 33 33 33 38 22 29 38 38 25 33 33 22 38 24 22 21 38 24 38 25 57 33 21 21 38 38 19 22

424: 33 22 29 33 21 22 38 24 22 33 38 38 29 21 25 33 21 38 21 29  0]xS 

425: 1925 3347 M (on the model )[38 38 22 21 38 33 22 57 38 38 33 21  0]xS 

426: 350 3433 M (can be determined)[33 33 38 52 38 33 51 38 33 21 33 25 57 21 38 33  0]xS 

427: 965 3433 M (, a)[19 51  0]xS 

428: 1068 3433 M (nd )[38 38  0]xS 

429: 1195 3433 M (then)[21 38 33  0]xS 

430: 1325 3433 M (, using an algorithm such as )[19 51 38 29 21 38 37 51 33 38 51 33 21 37 38 25 21 21 38 57 51 29 38 33 38 51 33 29  0]xS 

431: 350 3520 M (Generalized Iterative Scaling \(GIS\))[53 33 38 33 25 33 21 21 33 33 38 30 25 21 33 25 33 21 21 37 33 30 42 33 33 21 21 38 37 30 25 53

432: 25 42  0]xS 

433: 1445 3520 M (,)S 

434: 1464 3520 M ( )S 

435: 1494 3520 M (a model can be found that )[33 30 57 38 38 33 21 29 33 33 38 29 38 33 29 24 38 38 38 38 29 21 38 33 21  0]xS 

436: 350 3606 M (satisfies all of the constraints, while being as smooth as possible.)

437: [29 33 21 21 29 24 21 33 29 22 33 21 21 22 38 24 22 21 38 33 22 33 38 38 29 21 25 33 21 38 21 29

438: 19 22 53 38 21 21 33 22 38 33 21 38 37 21 33 29 21 29 57 38 38 21 38 21 33 29 21 38 38 29 29 21

439: 38 21 33  0]xS 

440: 2324 3606 M ( )S 

441: 350 3692 M (However, training maximum entropy models can be extremely )[54 38 53 33 37 33 25 19 34 21 25 33 21 38 21 38 37 33 57 33 37 21 57 38 57 33 33 38 21 25 38 38

442: 36 33 57 38 38 33 21 29 33 33 33 38 33 38 33 33 33 37 21 25 33 57 33 21 36  0]xS 

443: 350 3778 M (time consuming, taking weeks, months, or more.  We show that )

444: [21 21 57 33 26 33 38 38 29 38 57 21 38 37 19 26 21 33 37 21 38 37 26 53 33 33 37 29 19 26 57 38

445: 38 21 38 29 19 25 38 25 25 57 38 25 33 19 25 25 70 33 25 29 38 38 53 25 21 38 33 21  0]xS 

446: 350 3865 M (by using word classing, the training time can be significantly )

447: [38 36 36 38 29 21 38 37 36 53 38 25 38 35 33 21 33 29 29 21 38 37 19 35 21 38 33 35 21 25 33 21

448: 38 21 38 37 35 21 21 57 33 35 33 33 38 35 38 33 35 29 21 37 38 21 24 21 33 33 38 21 21 36  0]xS 

449: 350 3951 M (reduced, by up to a factor of 35.  In par)[25 33 38 38 33 33 38 19 30 38 36 30 38 38 30 21 38 30 33 30 24 33 33 21 38 25 30 38 24 30 38 38

450: 19 29 29 25 38 29 38 33  0]xS 

451: 1636 3951 M (ticular, we change the )[21 21 33 38 21 33 25 19 29 53 33 29 33 38 33 38 37 33 29 21 38 33  0]xS 

452: 350 4037 M (form of t)[24 38 25 57 23 38 24 23  0]xS 

453: 623 4037 M (he model.  Instead of predicting words directly, we first )[38 33 23 57 38 38 33 21 19 23 23 25 38 29 21 33 33 38 23 38 24 23 38 25 33 38 21 33 21 21 38 37

454: 23 53 38 25 38 29 22 38 21 25 33 33 21 21 36 19 22 53 33 22 24 21 25 29 21  0]xS 

455: 350 4123 M (predict the class that the next word belongs to, and then predict )

456: [38 25 33 38 21 33 21 26 21 38 33 26 33 21 33 29 29 26 21 38 33 21 26 21 38 33 26 38 33 37 21 25

457: 53 38 25 38 25 38 33 21 38 38 37 29 25 21 38 19 25 33 38 38 25 21 38 33 38 25 38 25 33 38 21 33

458: 21  0]xS 

459: 350 4210 M (the word itself, conditioned on its class.   The technique used is )

460: [21 38 33 25 53 38 25 38 25 21 21 29 33 21 24 19 25 33 38 38 38 21 21 21 38 38 33 38 25 38 38 25

461: 21 21 29 25 33 21 33 29 29 19 25 24 24 45 38 33 24 21 33 33 38 38 21 38 38 33 24 38 29 33 38 24

462: 21 29  0]xS 

463: 350 4296 M (actually more general: it can be applied to any problem where )

464: [33 33 21 38 33 21 21 36 31 57 38 25 33 31 37 33 38 33 25 33 21 21 31 21 21 31 33 33 38 31 38 33

465: 31 33 38 38 21 21 33 38 31 21 38 30 33 38 36 30 38 25 38 38 21 33 57 30 53 38 33 25 33  0]xS 

466: 350 4382 M (the)[21 38  0]xS 

467: 442 4382 M (re are a large number of outputs to predict, with language )[25 33 35 33 25 33 35 33 35 21 33 25 37 33 35 38 38 57 38 33 25 35 38 24 35 38 38 21 38 38 21 29

468: 35 21 38 34 38 25 33 38 21 33 21 19 34 53 21 21 38 34 21 33 38 37 38 33 37 33  0]xS 

469: 350 4468 M (modeling being just one example.  Furthermore, the technique )

470: [57 38 38 33 21 21 38 37 32 38 33 21 38 37 32 21 38 29 21 32 38 38 33 32 33 37 33 57 38 21 33 19

471: 32 32 42 38 25 21 38 33 25 57 38 25 33 19 32 21 38 33 32 21 33 33 38 38 21 38 38 33  0]xS 

472: 350 4555 M (applies not only to maximum entropy mod)[33 38 38 21 21 33 29 25 38 38 21 25 38 38 21 36 25 21 38 25 57 33 37 21 57 38 57 25 33 38 21 25

473: 38 38 36 25 57 38  0]xS 

474: 1664 4555 M (els, but to almost any )[33 21 29 19 24 38 38 21 24 21 38 24 33 21 57 38 29 21 24 33 38 36  0]xS 

475: 350 4641 M (machine learning technique for pre)[57 33 33 38 21 38 33 36 21 33 33 25 38 21 38 37 36 21 33 33 38 38 21 38 38 33 36 24 38 25 36 38

476: 25  0]xS 

477: 1469 4641 M (dict)[38 21 33  0]xS 

478: 1582 4641 M (ing probabilities that is )[21 38 37 36 38 25 38 38 33 38 21 21 21 21 21 33 29 35 21 38 33 21 35 21 29  0]xS 

479: 350 4727 M (slowed )[29 21 38 53 33 38  0]xS 

480: 596 4727 M (by a large number of outputs, including many uses of )[38 36 34 33 34 21 33 25 37 33 33 38 38 57 38 33 25 33 38 24 33 38 38 21 38 38 21 29 19 33 21 38

481: 33 21 38 38 21 38 37 33 57 33 38 36 33 38 29 33 29 33 38 24  0]xS 

482: 350 4813 M (decision trees and neural networks.)[38 33 33 21 29 21 38 38 19 21 25 33 33 29 19 33 38 38 19 38 33 38 25 33 21 19 38 33 21 53 38 25

483: 37 29  0]xS 

484: 1408 4813 M ( )S 

485: 463 4900 M ( In this paper, we first give a very brief introduction to )[19 25 38 42 21 38 21 29 42 38 33 38 33 25 19 42 53 33 42 24 21 25 29 21 42 37 21 37 33 42 33 41

486: 37 33 25 36 41 38 25 21 33 24 41 21 38 21 25 38 38 38 33 21 21 38 38 41 21 38  0]xS 

487: 350 4986 M (maximum entropy techniques for language modeling.  We then )[57 33 37 21 57 38 57 29 33 38 21 25 38 38 36 29 21 33 33 38 38 21 38 38 33 29 29 24 38 25 29 21

488: 33 38 37 38 33 37 33 29 57 38 38 33 21 21 38 37 19 29 29 70 33 28 21 38 33 38  0]xS 

489: 350 5072 M (go on to describe our speedup.  After this, we de)[37 38 25 38 38 25 21 38 25 38 33 29 33 25 21 38 33 25 38 38 25 24 29 38 33 33 38 38 38 19 24 24

490: 53 24 21 33 25 24 21 38 21 29 19 24 53 33 24 38  0]xS 

491: 1862 5072 M (scribe previous )[29 33 25 21 38 33 24 38 25 33 37 21 38 38 29  0]xS 

492: 350 5158 M (research in speeding up maximum entropy training, and compare )

493: [25 33 29 33 33 25 33 38 22 21 38 22 29 38 33 33 38 21 38 37 22 38 38 22 57 33 37 21 57 38 57 22

494: 33 38 21 25 38 38 36 22 21 25 33 21 38 21 38 37 19 22 33 38 38 21 33 38 57 38 33 25 33  0]xS 

495: 350 5245 M (it to our technique.  Next, we give experimental results, showing )

496: [21 21 23 21 38 23 38 38 25 23 21 33 33 38 38 21 38 38 33 19 22 22 54 33 37 21 19 22 53 33 22 37

497: 21 37 33 22 33 37 38 33 25 21 57 33 38 21 33 21 22 25 33 29 38 21 21 29 19 22 29 38 38 53 21 38

498: 37  0]xS 

499: 350 5331 M (both the increased speed of training, and a slight reduction in )

500: [38 38 21 38 33 21 38 33 33 21 38 33 25 33 33 29 33 38 33 29 38 33 33 38 33 38 24 33 21 25 33 21

501: 38 21 38 37 19 32 33 38 38 32 33 32 29 21 21 37 38 21 32 25 33 38 38 33 21 21 38 38 32 21 38  0]xS 

502: 350 5417 M (perplexity of the resulting models.  Finally, we)[38 33 25 38 21 33 37 21 21 36 27 38 24 27 21 38 33 27 25 33 29 38 21 21 21 38 37 27 57 38 38 33

503: 21 29 19 26 26 42 21 38 33 21 21 36 19 26 53  0]xS 

504: 1808 5417 M ( conclude with a )[26 33 38 38 33 21 38 38 33 26 53 21 21 38 26 33  0]xS 

505: 350 5503 M (short discussion of how these results can be applied to other )

506: [29 38 38 25 21 36 38 21 29 33 38 29 29 21 38 38 36 38 24 36 38 38 53 36 21 38 33 29 33 36 25 33

507: 29 38 21 21 29 36 33 33 38 36 38 33 35 33 38 38 21 21 33 38 35 21 38 35 38 21 38 33 25  0]xS 

508: 350 5590 M (machine learning techniques and to other problems.)[57 33 33 38 21 38 33 19 21 33 33 25 38 21 38 37 19 21 33 33 38 38 21 38 38 33 29 19 33 38 38 19

509: 21 38 19 38 21 38 33 25 19 38 25 38 38 21 33 57 29  0]xS 

510: 1906 5590 M ( )S 

511: 2638 1496 M (We begin with a very quick introduction to language models )[70 33 24 38 33 37 21 38 24 53 21 21 38 24 33 24 37 33 25 36 24 38 38 21 33 37 23 21 38 21 25 38

512: 38 38 33 21 21 38 38 23 21 38 23 21 33 38 37 38 33 37 33 23 57 38 38 33 21 29  0]xS 

513: 2525 1582 M (in general, and maximum entropy)[21 38 43 37 33 38 33 25 33 21 19 43 33 38 38 43 57 33 37 21 57 38 57 42 33 38 21 25 38 38  0]xS 

514: 3632 1582 M (-)S 

515: 3657 1582 M (based language models in )[38 33 29 33 38 42 21 33 38 37 38 33 37 33 42 57 38 38 33 21 29 42 21 38  0]xS 

516: 2525 1668 M (particular)[38 33 25 21 21 33 38 21 33  0]xS 

517: 2813 1668 M (.  Language models assign probabilities to word )[19 55 55 45 33 38 37 38 33 37 33 54 57 38 38 33 21 29 54 33 29 29 21 37 38 54 38 25 38 38 33 38

518: 21 21 21 21 21 33 29 54 21 38 54 53 38 25 38  0]xS 

519: 2525 1776 M (sequences )[29 33 38 38 33 38 33 33 29  0]xS 

520: /F1S4B F1 [75 0 0 -75 0 0 ] mFS

521: F1S4B Ji 

522: 2856 1776 M (P\(w)[46 24  0]xS 

523: /F3S43 F3 [67 0 0 -67 0 0 ] mFS

524: F3S43 Ji 

525: 2976 1789 M (1)S 

526: F1S4B Ji 

527: 3010 1776 M <8577>[67  0]xS 

528: /F1S43 F1 [67 0 0 -67 0 0 ] mFS

529: F1S43 Ji 

530: 3127 1789 M (n)S 

531: F1S4B Ji 

532: 3161 1776 M (\))S 

533: F3S4B Ji 

534: 3186 1776 M (.  Typically, this is done using the trigram )[19 27 27 45 36 38 21 33 33 21 21 36 19 27 21 38 21 29 26 21 29 26 38 38 38 33 26 38 29 21 38 37

535: 26 21 38 33 26 21 25 21 37 25 33 57  0]xS 

536: 2525 1869 M (approximation:)[33 38 38 25 38 37 21 57 33 21 21 38 38  0]xS 

537: 

538: %%IncludeResource: font Symbol

539: F /F5 0 /2 F /Symbol mF 

540: /F5S7D F5 [125.906 0 0 -125.906 0 0 ] mFS

541: F5S7D Ji 

542: 3615 2068 M <D5>S 

543: 2941 2068 M <D5>S 

544: /F5S43 F5 [67.148 0 0 -67.148 0 0 ] mFS

545: F5S43 Ji 

546: 3645 2147 M (=)S 

547: 4129 2069 M (-)S 

548: 3990 2069 M (-)S 

549: 2972 2147 M (=)S 

550: 3443 2069 M (-)S 

551: /F5S53 F5 [83.938 0 0 -83.938 0 0 ] mFS

552: F5S53 Ji 

553: 3554 2048 M <BB>S 

554: 2880 2048 M (=)S 

555: /F1S43 F1 [67.148 0 0 -67.148 0 0 ] mFS

556: F1S43 Ji 

557: 3648 1957 M (n)S 

558: 3624 2147 M (i)S 

559: 4107 2069 M (i)S 

560: 3968 2069 M (i)S 

561: 3844 2069 M (i)S 

562: 2974 1957 M (n)S 

563: 2950 2147 M (i)S 

564: 3421 2069 M (i)S 

565: 3171 2069 M (i)S 

566: 2794 2069 M (n)S 

567: /F1S53 F1 [83.938 0 0 -83.938 0 0 ] mFS

568: F1S53 Ji 

569: 4061 2048 M (w)S 

570: 3922 2048 M (w)S 

571: 3798 2048 M (w)S 

572: 3721 2048 M (P)S 

573: 3375 2048 M (w)S 

574: 3249 2048 M (w)S 

575: 3125 2048 M (w)S 

576: 3047 2048 M (P)S 

577: 2746 2048 M (w)S 

578: 2619 2048 M (w)S 

579: 2542 2048 M (P)S 

580: /F3S43 F3 [67.148 0 0 -67.148 0 0 ] mFS

581: F3S43 Ji 

582: 3676 2147 M (1)S 

583: 4159 2069 M (1)S 

584: 4027 2069 M (2)S 

585: 3002 2147 M (1)S 

586: 3473 2069 M (1)S 

587: 3290 2069 M (1)S 

588: 2660 2069 M (1)S 

589: /F3S53 F3 [83.938 0 0 -83.938 0 0 ] mFS

590: F3S53 Ji 

591: 4192 2048 M (\))S 

592: 3887 2048 M (|)S 

593: 3770 2048 M (\()S 

594: 3506 2048 M (\))S 

595: 3319 2048 M (...)[21 21  0]xS 

596: 3213 2048 M (|)S 

597: 3096 2048 M (\()S 

598: 2833 2048 M (\))S 

599: 2690 2048 M (...)[21 21  0]xS 

600: 2591 2048 M (\()S 

601: F3S4B Ji 

602: 4232 2048 M ( )S 

603: 2525 2232 M (Maximum entropy models for language modeling )[67 33 37 21 57 38 57 40 33 38 21 25 38 38 36 40 57 38 38 33 21 29 40 24 38 25 40 21 33 38 37 38

604: 33 37 33 39 57 38 38 33 21 21 38 37  0]xS 

605: 4158 2232 M ([2])[26 38  0]xS 

606: 4248 2232 M ( do not )[39 38 38 39 38 38 21  0]xS 

607: 2525 2318 M (necessarily use the n)[38 33 33 33 29 29 33 25 21 21 36 33 38 29 33 33 21 38 33 32  0]xS 

608: 3184 2318 M (-)S 

609: 3209 2318 M (gra)[37 25  0]xS 

610: 3304 2318 M (m approximation and can in principle )[57 32 33 38 38 25 38 37 21 57 33 21 21 38 38 32 33 38 38 32 33 33 38 32 21 38 32 38 25 21 38 33

611: 21 38 21 33  0]xS 

612: 2525 2405 M (condition on arbitrary length contexts.  The general form of a )

613: [33 38 38 38 21 21 21 38 38 33 38 38 33 33 25 38 21 21 25 33 25 36 33 21 33 38 37 21 38 32 33 38

614: 38 21 33 37 21 29 19 32 32 45 38 33 32 37 33 38 33 25 33 21 32 24 38 25 57 32 38 24 32 33  0]xS 

615: 2525 2491 M (conditional maximum entropy model is as follows:)[33 38 38 38 21 21 21 38 38 33 21 19 57 33 37 21 57 38 57 19 33 38 21 25 38 38 36 19 57 38 38 33

616: 21 19 21 29 19 33 29 19 24 38 21 21 38 53 29  0]xS 

617: 4052 2491 M ( )S 

618: Pscript_WinNT_Incr begin

619: %%BeginResource: file Pscript_Win_GdiObject 5.0 0

620: /SavedCTM null d/CTMsave{/SavedCTM SavedCTM currentmatrix d}b/CTMrestore

621: {SavedCTM setmatrix}b/mp null d/ADO_mxRot null d/GDIHMatrix null d

622: /GDIHPatternDict 22 dict d GDIHPatternDict `/PatternType 1 d/PaintType 2 d/Reps

623: L2?{1}{5}? d/XStep 8 Reps mul d/YStep XStep d/BBox[0 0 XStep YStep]d/TilingType

624: 1 d/PaintProc{` 1 Lw[]0 sd PaintData , exec E}b/FGnd null d/BGnd null d

625: /HS_Horizontal{horiz}b/HS_Vertical{vert}b/HS_FDiagonal{fdiag}b/HS_BDiagonal

626: {biag}b/HS_Cross{horiz vert}b/HS_DiagCross{fdiag biag}b/MaxXYStep XStep YStep

627: gt{XStep}{YStep}? d/horiz{Reps{0 4 M XStep 0 - 0 8 +}repeat 0 -8 Reps mul + K}b

628: /vert{Reps{4 0 M 0 YStep - 8 0 +}repeat 0 -8 Reps mul + K}b/biag{Reps{0 0 M

629: MaxXYStep @ - 0 YStep neg M MaxXYStep @ - 0 8 +}repeat 0 -8 Reps mul + 0 YStep

630: M 8 8 - K}b/fdiag{Reps{0 0 M MaxXYStep @ neg - 0 YStep M MaxXYStep @ neg - 0 8

631: +}repeat 0 -8 Reps mul + MaxXYStep @ M 8 -8 - K}b E/makehatch{4 -2 $/yOrg ~ d

632: /xOrg ~ d GDIHPatternDict/PaintData 3 -1 $ put CTMsave GDIHMatrix setmatrix

633: GDIHPatternDict matrix xOrg yOrg + mp CTMrestore ~ U ~ 2 ^ put}b/h0{/h0

634: /HS_Horizontal makehatch}b/h1{/h1/HS_Vertical makehatch}b/h2{/h2/HS_FDiagonal

635: makehatch}b/h3{/h3/HS_BDiagonal makehatch}b/h4{/h4/HS_Cross makehatch}b/h5{/h5

636: /HS_DiagCross makehatch}b/GDIBWPatternMx null d/pfprep{save 8 1 $

637: /PatternOfTheDay 8 1 $ GDIBWPatternDict `/yOrg ~ d/xOrg ~ d/PaintData ~ d/yExt

638: ~ d/Width ~ d/BGnd ~ d/FGnd ~ d/Height yExt RepsV mul d/mx[Width 0 0 Height 0

639: 0]d E build_pattern ~ !}b/pfbf{/fEOFill ~ d pfprep hbf fEOFill{O}{L}? restore}b

640: /GraphInit{GDIHMatrix null eq{/SavedCTM matrix d : ADO_mxRot concat 0 0 snap +

641: : 0.48 @ GDIHPatternDict ` YStep mul ~ XStep mul ~ nonzero_dsnap YStep V ~

642: XStep V ~ E +S/GDIHMatrix matrix currentmatrix readonly d ; : 0.24 -0.24 +S

643: GDIBWPatternDict ` Width Height E nonzero_dsnap +S/GDIBWPatternMx matrix

644: currentmatrix readonly d ; ;}if}b

645: %%EndResource

646: %%BeginResource: file Pscript_Win_GdiObject_L2 5.0 0

647: /GDIBWPatternDict 25 dict @ `/PatternType 1 d/PaintType 1 d/RepsV 1 d/RepsH 1 d

648: /BBox[0 0 RepsH 1]d/TilingType 1 d/XStep 1 d/YStep 1 d/Height 8 RepsV mul d

649: /Width 8 d/mx[Width 0 0 Height neg 0 Height]d/FGnd null d/BGnd null d

650: /SetBGndFGnd{BGnd null ne{BGnd aload ! scol BBox aload ! 2 ^ sub ~ 3 ^ sub ~

651: rf}if FGnd null ne{FGnd aload ! scol}if}b/PaintProc{` SetBGndFGnd RepsH{Width

652: Height F mx PaintData imagemask Width 0 +}repeat E}b E d/mp/makepattern , d

653: /build_pattern{CTMsave GDIBWPatternMx setmatrix/nupangle where{! nupangle -90

654: eq{nupangle R}if}if GDIBWPatternDict @ ` Width Height ne{Width Height gt{Width

655: Height V 1}{1 Height Width V}? +S}if xOrg yOrg E matrix + mp CTMrestore}b/hbf

656: {setpattern}b/hf{:/fEOFill ~ d ~ ! setpattern fEOFill{O}{L}? ;}b/pbf{: !

657: /fEOFill ~ d GDIBWPatternDict `/yOrg ~ d/xOrg ~ d/PaintData ~ d/OutputBPP ~ d

658: /Height ~ d/Width ~ d/PaintType 1 d/PatternType 1 d/TilingType 1 d/BBox[0 0

659: Width Height]d/XStep Width d/YStep Height d/mx xOrg yOrg matrix + d 20 dict @ `

660: /ImageType 1 d/Width Width d/Height Height d/ImageMatrix[1 0 0 1 0 0]d

661: /BitsPerComponent 8 d OutputBPP 24 eq{/Decode[0 1 0 1 0 1]d}{OutputBPP 8 eq{

662: /Decode[0 1]d}{/Decode[0 1 0 1 0 1 0 1]d}?}?/DataSource{PaintData}d E/ImageDict

663: ~ d/PaintProc{` ImageDict image E}b & mx makepattern setpattern E fEOFill{O}{L}

664: ? ;}b/mask_pbf{:/fEOFill ~ d 20 dict `/yOrg ~ d/xOrg ~ d/PaintData ~ d/Height ~

665: d/Width ~ d/PatternType 1 d/PaintType 2 d/TilingType 1 d/BBox[0 0 Width Height]

666: d/XStep Width d/YStep Height d/mx xOrg yOrg matrix + d/PaintProc{` Width Height

667: T[1 0 0 1 0 0]{PaintData}imagemask E}b & mx makepattern setpattern E fEOFill{O}

668: {L}? ;}b

669: %%EndResource

670: end reinitialize

671: 0 Lj 2 Lc 16 Lw solid : 3007 2733 824 85 rc N 3049 2775 M 3816 2775 I : 0.219 0.219 +S K 

672: ; ; /F3S46 F3 [70.941 0 0 -70.941 0 0 ] mFS

673: F3S46 Ji 

674: 3598 2849 M (\))S 

675: 3424 2849 M (...)[18 18  0]xS 

676: 3331 2849 M (\()S 

677: 3756 2663 M (\))S 

678: 3582 2663 M (...)[18 18  0]xS 

679: 3488 2663 M (,)S 

680: 3417 2663 M (\()S 

681: 3054 2663 M (exp)[31 35  0]xS 

682: 2955 2793 M (\))S 

683: 2781 2793 M (...)[18 18  0]xS 

684: 2687 2793 M (|)S 

685: 2604 2793 M (\()S 

686: 2585 2793 M (')S 

687: /F3S38 F3 [56.754 0 0 -56.754 0 0 ] mFS

688: F3S38 Ji 

689: 3567 2867 M (1)S 

690: 3396 2867 M (1)S 

691: 3725 2681 M (1)S 

692: 3554 2681 M (1)S 

693: 2924 2811 M (1)S 

694: 2753 2811 M (1)S 

695: /F5S38 F5 [56.754 0 0 -56.754 0 0 ] mFS

696: F5S38 Ji 

697: 3539 2867 M (-)S 

698: 3696 2681 M (-)S 

699: 2896 2811 M (-)S 

700: 11585 VM?

701: Pscript_WinNT_Incr begin

702: %%BeginResource: file Pscript_T42Hdr 5.0 0

703: /asc42 0.0 d/sF42{/asc42 ~ d Ji}bind d/bS42{0 asc42 -M}bind d/eS42{0 asc42 neg

704: -M}b/Is2015?{version cvi 2015 ge}bind d/AllocGlyphStorage{Is2015?{!}{{string}

705: forall}?}bind d/Type42DictBegin{25 dict `/FontName ~ d/Encoding ~ d 4 array

706: astore cvx/FontBBox ~ d/PaintType 0 d/FontType 42 d/FontMatrix[1 0 0 1 0 0]d

707: /CharStrings 256 dict `/.notdef 0 d & E d/sfnts}bind d/Type42DictEnd{& @

708: /FontName get ~ definefont ! E}bind d/RDS{string currentfile ~ readstring !}

709: executeonly d/PrepFor2015{Is2015?{/GlyphDirectory 16 dict d sfnts 0 get @ 2 ^

710: (glyx)putinterval 2 ^(locx)putinterval ! !}{! !}?}bind d/AddT42Char{Is2015?

711: {findfont/GlyphDirectory get ` d E ! !}{findfont/sfnts get 4 ^ get 3 ^ 2 ^

712: putinterval ! ! ! !}?}bind d/IDStrNull{1 add 2 mul @ string 0 1 3 ^ 1 sub{1 ^ ~

713: 0 put}for ~ !}bind d/IDStr{@ 1 add 2 mul string 0 1 3 ^{1 ^ ~ @ 2 mul ~ 3 copy

714: 256 idiv put ~ 1 add ~ 256 mod put}for ~ !}bind d/IDStr2{~ @ 1 add 2 mul string

715: 0 1 3 ^{1 ^ ~ @ 2 mul ~ 5 ^ add 3 copy 256 idiv put ~ 1 add ~ 256 mod put}for ~

716: ! ~ !}bind d/CIDT42Begin{25 dict ` @/WMode ~ d 0 gt{/Metrics2 16 dict d

717: /FontMatrix[0 1 -1 0 0 0]d}{/FontMatrix[1 0 0 1 0 0]d}?/CIDMap ~ d/CIDCount ~ d

718: /CIDSystemInfo 3 dict @ ` 3 -1 $/Supplement ~ d 3 -1 $/Ordering ~ d 3 -1 $

719: /Registry ~ d E d/CIDFontName ~ d/Encoding ~ d 4 array astore cvx/FontBBox ~ d

720: /CharStrings 2 dict @ `/.notdef 0 d E d/GDBytes 2 d/CIDFontType 2 d/FontType 42

721: d/PaintType 0 d/sfnts}bind d/CIDT42End{CIDFontName & E/CIDFont defineresource

722: !}bind d/T42CIDCP32K{/CIDFont findresource @ length dict copy @/FID undef `

723: /CIDFontName ~ d/CIDMap ~ d/CIDCount ~ d/Metrics2 8 dict d CIDFontName & E

724: /CIDFont defineresource !}bind d/T42CIDCPR{/CIDFont findresource @ length dict

725: copy @/FID undef `/CDevProc{!}d/CIDFontName ~ d CIDFontName & E/CIDFont

726: defineresource !}bind d/T0AddT42Char{/CIDFont findresource/GlyphDirectory get `

727: d E ! !}bind d/T0AddT42Mtx2{/CIDFont findresource/Metrics2 get ` d E}bind d

728: /UpdateCIDMap{/CIDFont findresource/CIDMap get 3 1 $ putinterval}d/AddXUID

729: {version cvi 3011 ge{/XUID ~ d}{!}?}bind d/AddFontInfoBegin{/FontInfo 16 dict @

730: `}bind d/AddFontInfo{/GlyphNames2Unicode 16 dict d}bind d/AddFontInfoEnd{E d}

731: bind d/G2UBegin{findresource/FontInfo get/GlyphNames2Unicode get `}bind d

732: /G2UEnd{E}bind d

733: %%EndResource

734: end reinitialize

735: 0 -0.219 1.113 1.004

736:  256 array 0 1 255 {1 index exch /.notdef put} for  /TTA2292A68t00

737: Type42DictBegin

738: [<00010000000a000a000a000a637674202ac4fb85000000ac000005046670

739: 676df530c471000005b00000050f676c79663810e3e8000013300000ae12

740: 68656164c5ad5c6500000ac000000036686865610f2d038100000af80000

741: 0024686d74787750493e00000b1c000002fc6c6f6361393b0de500000e18

742: 000001826d617870020305a700000f9c00000020707265707e0fc1650000

743: 0fbc0000037167646972000000000000000000000000055d0031054c001f

744: 054c001c0394001b0000ffe10000ffe10000ffe4fe59ffed056b0023fe6a

745: ffdc270f000000b4000000b4000000000000000000610025005e0029004a

746: 00c200f000be012a005000c0ffd0006300580277002d004a008f00b700c0

747: 00de004400640002001a017c0020ff9c008c0026008300a00039041affb5

748: 004c00e8080f00160024003a016f000600150067009e01dc0049004b002d

749: 003b00470089007200470051006700f7013801dc0001000a005d00610094

750: 009b00be00ce00df00190026002e00d6010701f7ffbb004300530063007b

751: 00b900de0542058b0013009100a700a900c100e5012e01a401ea03eeffb5

752: ffccffeb0026002900310036005b00620062009200a100a700a700cb0223

753: 03a3ff1aff7bffa8ffc8001c004400ab00b400c100c500ce00f601200156

754: 016302a60379fefaff6effab0000000c00150024003a003a0041004c004e

755: 007400a500ce00d600fb0109010901820220023f0263fe11ff170007001f

756: 0020002b00530055006d007e0088008d0094009a009f00af00e8013f0168

757: 01cd01ee020d028002bf0550074e074e09bffdd5ff47fff6000c000c0021

758: 00290057006b0078007e009200930096009a00a100a500c400c400f000fd

759: 0136013b013e01490152018f018601a501f8022b02db02fb030e030f032c

760: 03590360037c03e203f90805fd4ffe97fed0ff50ff78000100010006000a

761: 0025004300440054005c007000730077007c00800088008e008e00900091

762: 00ab00ab00b000b900bf00bf00c300dd00e300e300e600e800ec01020105

763: 0109010a012a012a01300130013b0140014a0155015f01670173019701b6

764: 01b801be01c501ea01f4023402360239024f0252026702700274029002b0

765: 02b102bf02e402f302fb033a03450373037f03ab03c203c803d203da0400

766: 044a046104690475049404b004d504f004f404fe050005230523052e0561

767: 05a0fda2fe4bfe5bff26ff3bff62ff8dff90ff90ffe80007001e00250026

768: 002700440050006200630066006700700070007100710079007e00850091

769: 0096009e00a100a100a500b600b800bc00c000c200c500c600d000da00dd

770: 00dd00e200f700f901000102010501150121012d01310136013d01470147

771: 015a01680176017c0180018e019a01a501b101b501b501c001dc01ea01fe

772: 023902560259025e0276027c028d0297029b02b902c002d602db02ff0306

773: 030f03180347034b0357035b035e0374039603980399039e039f03b803c0

774: 03c60402043a045004720483049004a204af04c704d304dc0530053b058a

775: 058a05b005b505bd05e60612061b0634065306b80706087908e708fe00b3

776: 00c700b100b5000000000000000000000000000000000188011001b100ad

777: 01e70304038002cc0258020b02d6018204cf01270274023c0100058e006e

778: 00f202c0021d01d2023701570363015a002500820166015b009d031c03a2

779: 034a00cc056e0153035001640053049700fc05c30061058a02b20302058a

780: 03fd01f902ab0068008e026bfda9ff59010e005301a402a600ab007d003c

781: 0029041a00d6023904b90125022902bf07ae01b8003e07060339033a080f

782: 07ae01a403e300000053080f01ea006103e2004c00e102a6035202a6080f

783: 00e601dc074f006e00fd03af0000007d003e01a4fe1100a905590260011c

784: 02a4014cfd83fdd8039bff9f011902e20000fd11fd41024301b60155ffaa

785: 02be00ef0097fef904e000be02db005501dc0063006b007800f200530077

786: 003a015600fc020d0029005300610070403f3f3e3d3c3b3a393837353433

787: 3231302f2e2d2c2b2a292827262524232221201f1e1d1c1b1a1918171615

788: 14131211100f0e0d0c0b0a090807060504030201002c4523466020b02660

789: b004262348482d2c452346236120b02661b004262348482d2c45234660b0

790: 206120b04660b004262348482d2c4523462361b0206020b02661b02061b0

791: 04262348482d2c45234660b0406120b06660b004262348482d2c45234623

792: 61b0406020b02661b04061b004262348482d2c0110203c003c2d2c204523

793: 20b0cd442320b8015a51582320b08d44235920b0ed51582320b04d442359

794: 20b0042651582320b00d44235921212d2c20204518684420b001602045b0

795: 4676688a4560442d2c01b10b0a432343650a2d2c00b10a0b4323430b2d2c

796: 00b0172370b101173e01b0172370b10217453ab10200080d2d2c45b01a23

797: 4445b01923442d2c2045b00325456164b050515845441b2121592d2cb001

798: 43632362b0002342b00f2b2d2c2045b0004360442d2c01b00643b0074365

799: 0a2d2c2069b04061b0008b20b12cc08a8cb8100062602b0c642364615c58

800: b00361592d2c45b0112bb0172344b0177ae4182d2c45b0112bb01723442d

801: 2cb01243588745b0112bb0172344b0177ae41b038a45186920b01723448a

802: 8a8720b0a05158b0112bb0172344b0177ae41b21b0177ae45959182d2c2d

803: 2cb0022546608a46b040618c482d2c4b53205c58b002855958b00185592d

804: 2c20b0032545b019234445b01a23444565234520b00325606a20b0092342

805: 23688a6a606120b01a8ab000527921b21a1a40b9ffe0001a45208a545823

806: 21b03f1b235961441cb114008a5279b31940201945208a54582321b03f1b

807: 235961442d2cb110114323430b2d2cb10e0f4323430b2d2cb10c0d432343

808: 0b2d2cb10c0d432343650b2d2cb10e0f432343650b2d2cb1101143234365

809: 0b2d2c4b525845441b2121592d2c0120b003252349b04060b0206320b000

810: 525823b002253823b002256538008a63381b212121212159012d2c4bb064

811: 51584569b00943608a103a1b212121592d2c01b005251023208af500b001

812: 6023edec2d2c01b005251023208af500b0016123edec2d2c01b0062510f5

813: 00edec2d2c20b001600110203c003c2d2c20b001610110203c003c2d2cb0

814: 2b2bb02a2a2d2c00b00743b006430b2d2c3eb02a2a2d2c352d2c76b80241

815: 23701020b802414520b0005058b00161593a2f182d2c21210c6423648bb8

816: 4000622d2c21b08051580c6423648bb82000621bb200402f2b59b002602d

817: 2c21b0c051580c6423648bb81555621bb200802f2b59b002602d2c0c6423

818: 648bb84000626023212d2cb4000100000015b00826b00826b00826b00826

819: 0f10161345683ab001162d2cb4000100000015b00826b00826b00826b008

820: 260f1016134568653ab001162d2c4b53234b515a5820458a60441b212159

821: 2d2c4b545820458a60441b2121592d2c4b53234b515a58381b2121592d2c

822: 4b5458381b2121592d2cb0134358031b02592d2cb0134358021b03592d2c

823: 4b54b012435c5a58381b2121592d2cb012435c580cb00425b00425060c64

824: 23646164b807085158b00425b00425012046b01060482046b0106048590a

825: 21211b2121592d2cb012435c580cb00425b00425060c6423646164b80708

826: 5158b00425b00425012046b8fff060482046b8fff06048590a21211b2121

827: 592d2c4b53234b515a58b03a2b1b2121592d2c4b53234b515a58b03b2b1b

828: 2121592d2c4b53234b515ab012435c5a58381b2121592d2c0c8a034b54b0

829: 0426024b545a8a8a0ab012435c5a58381b2121592d000001000000028000

830: 1e63cadc5f0f3cf5081d080000000000a348bc3b00000000b24cd4e1ffff

831: fe3e08e7080b0000000b000100000000000000010000080bfe3e00000856

832: feabfc2e08740001000000000000000000000000000000be04cd00670000

833: 000000fa00000200000002aa00e405b40061040000250464005106aa0047

834: 063900480383003a02aa005402aa002d04000091046400250200006e0464

835: 0025020000910239ffff0400004a040000f00400002c0400005304000020

836: 04000062040000590400004c0400007c04000051023900ac0239008b0464

837: 003c046400250464003c038d00580464002505c700140556002205c70012

838: 04e5001a04e30027061b004804d3002305c7002602aa0033050c004e05c7

839: 0025057d0014071d002505c7ffe805c7004c0625002305ee004a04730026

840: 04bc003804e3003b05850016038300470625003805290042065c002504e3

841: 001602aa00a706e7016a02aa004a054400250400ffef040003da050c0054

842: 046400790464004603f4004503830049042b0047034affff04d3002d02a2

843: 004304d300450464001504640031049c007c042b00280464005804640037

844: 042b00590464006904d3004e0383002b049c004705b40041057d005003f2

845: 005b057d002f03f4005b03d700d3019a00a403d700d60464003404f60013

846: 01fa00790464003c0156feab05b4003604000000060600b9060601300606

847: 00f7060600f60856002907e5004b04d3005d07e5004b04d3005d03330066

848: 04640025034a00690464003c0464007905b4006603f4004103ae006a0464

849: 0025046400250464002404640034080000e604d302390800ff8d05440020

850: 069600e4057d002d065c005e07e501820625004e0625004e069600610625

851: 004d0625004d05b4003805b4003705b4003b05b4003b05b4003a05b4010a

852: 05b4010b0625005305b4004a0652006306520063071f00320696002e0464

853: 00540200009105b4003004d3003004d300300856002907e5004c04d3005d

854: 07e5004c04d3005d03f4002f02a200410652006306520063064a002a05b4

855: 002b03120053031200530312005303120000031200000312000003f401a4

856: 03f4001c03f401a403f401a40652007502a2003e02310000057d02a6057d

857: 02a6057d00430312ff930312025d0312ff93031200230312028103120023

858: 03f4001c01a4001c00000032003200320032007600e701d3021b032f0430

859: 048d04cf051005e4061606590673069506c4075307b4082708ae0907098c

860: 0a050a4b0af10b6f0bae0c090c5e0c8c0cdf0d7c0ded0eb50f72107b10cc

861: 1164122b127c131c1370143a151315a8165c16f4175117c0188019041988

862: 19f11ab41b541c121cce1d891df91e381e891ec21ef21f0d1f2c1fdd209c

863: 2174220122a9233b23bb247324da256b262226b3274427cd284228d22945

864: 29d82a432aa52b2d2c092cb62d8b2e322edc2f582f7c2ff9304930d23108

865: 317b31a8322a32f833b233de344c34c4355035a63605366336b337143756

866: 37b8382238923923399739be3a113a933ad33b543ba53bc33bdd3c173d31

867: 3dfb3f233feb40cb4165423b429c42fe435643c7446444be453a45a44649

868: 467f46ce47bf489149a14a254a784aa14ac94b234b774c254c984d114d88

869: 4dfd4e734ec64fad507750f1517951af51ca51fc5221523c5263529352f3

870: 5326534653e1542f549354e155005554558a55a555dd560556215649567c

871: 56d9570900000001000000c0006700050000000000020010002f00400000

872: 00ed050f00000000411500120227022702260226006f01160001005f0116

873: 0001004f01160001003f01160001002f01160001022fb22f1f1fb8022eb2

874: a01f1fb8022cb21d291fb80224400be11a1f3127121f3c27291fb8021740

875: 3b6c291fe0aa191fe1aa191fab27321f883c321f833c321f2f3c321f7627

876: 321f3c27291f27aa291f6f276f6d6f866f87043700161600000012110840

877: b801e3b655140914093790b801bf40390790bd0790b30790670790640790

878: 5707904607903607902e07902b079025079024071408120810080e080c08

879: 0a080808060804080208000814b8ffe0402b000001001406100000010006

880: 04000001000410000001001002000001000200000001000002010802004a

881: 00b013034b024b5342b0382b4bb807ff52b0372b4bb007505b58b101018e

882: 59b0382bb00288b801005458b801ffb101018e851bb0124358b9000101ff

883: 858d1bb9000101ff858d5959014bb0c063004b6220b0f65323b8010a515a

884: b005234201b0124b004b5442180016763f183f123e113946443e11394644

885: 3e113946443e113946443e11394660443e11394660442b2b2b2b2b2b2b2b

886: 2b2b2b182b2b2b2b2b2b2b2b2b2b2b2b2b4b5079bc001f01750007001f01

887: 74b3071f93072b2b2b4b5379bc00900175000700900174b3079093072b2b

888: 2b181db0964b5358b0aa1d59b0324b5358b0ff1d5901b341b5407b456523

889: 456023456560234560b08b766818b080622020b140b54565234520b00326

890: 6062636820b003266165b0b5236544b040234420b1417b4565234520b003

891: 266062636820b003266165b07b236544b0412344b1007b455458b17b4065

892: 44b24140414523614459b9004e0162b12a8e456523456023456560234560

893: b089766818b080622020b9002a01624565234520b003266062636820b003

894: 266165b80162236544b02a234420b14e8e4565234520b003266062636820

895: b003266165b08e236544b04e2344b1008e455458b18e406544b24e404e45

896: 2361445945695342014b5058b108004259435c58b108004259b3020b0a12

897: 4358601b2159421610703eb0124358b93b21187e1bba040001a8000b2b59

898: b00c2342b00d2342b0124358b92d412d411bba04000400000b2b59b00e23

899: 42b00f2342b0124358b9187e3b211bba01a80400000b2b59b0102342b011

900: 2342002b752b2b2b2b2b2b2b2b2b2b2b2b2b2b2b2b737373737300184569

901: 4418456944435c58b9ffc00280b2121f322b5900000000>

902: [15961 14417 14187 ] AllocGlyphStorage

903: ]def 

904: 108 44 

905: PrepFor2015

906: AddFontInfoBegin

907: /OrigFontType /TrueType def 

908: /OrigFontName <53796D626F6C> def

909: /OrigFontStyle () def

910: /FSType 0 def

911: AddFontInfoEnd

912: [16#2c 16#85ae213a ] AddXUID

913: Type42DictEnd

914: 3 13420 184 <0001033aff5a039b07690003001d400a02420002011d0300ba05b8026bb1

915: d1182b10f63cfd3c002fed313005231133039b6161a6080f>/TTA2292A68t00 AddT42Char 

916: 1 0 0 <0002010000000500050000030007003b401e06076302010005046303000a

917: 0605630303021a09070463010019082938182b4e10f43c4dfd3c4e10f63c

918: 4d10fd3c003f3cfd3c3f3cfd3c3130211121112521112101000400fc2003

919: c0fc400500fb002004c0>/TTA2292A68t00 AddT42Char 

920: /TTA2292A68t00 findfont /CharStrings get begin

921: /parenrightex 184 def

922: /.notdef 0 def

923: end

924: /TTA2292A68t00 findfont /Encoding get

925: dup 1 /parenrightex put

926: pop

927: F /F6 0 /0 F /TTA2292A68t00 mF 

928: /F6S46 F6 [70.941 0 0 -70.941 0 0 ] mFS

929: F6S46 Ji 

930: 3784 2703 M <01>S 

931: 3784 2636 M <01>S 

932: 533 VM?

933: 3 13474 185 <0001006fff5a039b0769000f002a400c004209480800010f0e093508b802

934: 22b70e1d01ba1121d1182b10f6fdf4ed103c103c002fe4ed313001111003

935: 020207060727241336123511039b4b40e99d719317011aa684860769fdf1

936: feccfed8ff00fe7e84613d14c20105d10248f8022300>/TTA2292A68t00 AddT42Char 

937: /TTA2292A68t00 findfont /CharStrings get begin

938: /parenrightbt 185 def

939: end

940: /TTA2292A68t00 findfont /Encoding get

941: dup 2 /parenrightbt put

942: pop

943: 3784 2760 M <02>S 

944: 533 VM?

945: 3 13312 183 <0001006fff5a039b0769000c0032400f37030107480842000102000c0735

946: 08b80222b4021d0cba0eb8026ab13f182b10f6fdf4ed103c103c002ffde4

947: 3130015d05231110022726273716001211039b62a8a891e917f10177ada6

948: 0222013c0256e2c4a01565fe32fd6ffec300>/TTA2292A68t00 AddT42Char 

949: /TTA2292A68t00 findfont /CharStrings get begin

950: /parenrighttp 183 def

951: end

952: /TTA2292A68t00 findfont /Encoding get

953: dup 3 /parenrighttp put

954: pop

955: 3784 2590 M <03>S 

956: 533 VM?

957: 3 11446 168 <00010053ff5a00b407690003001d400b02420000031d0102ba0421b9015e

958: 00182b10f63cfd3c002fed313017231133b46161a6080f00>/TTA2292A68t00 AddT42Char 

959: /TTA2292A68t00 findfont /CharStrings get begin

960: /parenleftex 168 def

961: end

962: /TTA2292A68t00 findfont /Encoding get

963: dup 4 /parenleftex put

964: pop

965: 3156 2703 M <04>S 

966: 3156 2636 M <04>S 

967: 533 VM?

968: 3 11500 169 <00010053ff5a037f0769000c002a400c00420748080102000c073508b802

969: 22b7021d0cba0d218d182b10f6fdf4ed103c103c002fe4ed313013331110

970: 1217161707260002115361a9a891e917f1fe89ad0769fdddfec4fdaae2c3

971: a1146501cd0292013c00>/TTA2292A68t00 AddT42Char 

972: /TTA2292A68t00 findfont /CharStrings get begin

973: /parenleftbt 169 def

974: end

975: /TTA2292A68t00 findfont /Encoding get

976: dup 5 /parenleftbt put

977: pop

978: 3156 2760 M <05>S 

979: 533 VM?

980: 3 11338 167 <00010053ff5a037f0769000f002a400c09480842000f0e0001093508b802

981: 22b70e1d01ba102133182b10f6fdf4ed103c103c002ffde4313017111013

982: 361237363717040306021511534b40e99c729317fee6a68487a6020e0134

983: 0129ff018284613e15c2fefbd1fdb8f8fdde>/TTA2292A68t00 AddT42Char 

984: /TTA2292A68t00 findfont /CharStrings get begin

985: /parenlefttp 167 def

986: end

987: /TTA2292A68t00 findfont /Encoding get

988: dup 6 /parenlefttp put

989: pop

990: 3156 2590 M <06>S 

991: /F5S46 F5 [70.941 0 0 -70.941 0 0 ] mFS

992: F5S46 Ji 

993: 2994 2793 M (=)S 

994: 533 VM?

995: 3 11066 166 <00010000ff34056105ed001500babc0007018e000b000b0192403a083707

996: 01060400150714150e0302030402260e0d140e0e0d0c0d0d27030414030d

997: 0e03040e0d0c0302050604020e010d03040b0c264006050103b8013e4011

998: 200f0eaa0001147e1717171a154f074806b80158b304191617b8015cb321

999: 2028182b2b4ef44dfdf4f44e456544e64ded002f3cfd3c1a19ed183f3c1a

1000: fd3c391139111239011112173987082e2b057d10c487042e182b087d10c4

1001: 00113939011112391112391810edec0010f5313005213501013521132326

1002: 2623210101213237363637170521fadf029dfd6304fe1b28107b6efd1c02

1003: 32fd8b03324653394b1928cc24033f033125feaf8a76fd51fcf71b136167

1004: 0700>/TTA2292A68t00 AddT42Char 

1005: /TTA2292A68t00 findfont /CharStrings get begin

1006: /summation 166 def

1007: end

1008: /TTA2292A68t00 findfont /Encoding get

1009: dup 7 /summation put

1010: pop

1011: /F6S6A F6 [106.414 0 0 -106.414 0 0 ] mFS

1012: F6S6A Ji 

1013: 3190 2679 M <07>S 

1014: /F1S38 F1 [56.754 0 0 -56.754 0 0 ] mFS

1015: F1S38 Ji 

1016: 3518 2867 M (i)S 

1017: 3224 2746 M (j)S 

1018: 3676 2681 M (i)S 

1019: 3391 2681 M (j)S 

1020: 3323 2681 M (j)S 

1021: 2875 2811 M (i)S 

1022: /F1S46 F1 [70.941 0 0 -70.941 0 0 ] mFS

1023: F1S46 Ji 

1024: 3475 2849 M (w)S 

1025: 3358 2849 M (w)S 

1026: 3246 2849 M (Z)S 

1027: 3633 2663 M (w)S 

1028: 3516 2663 M (w)S 

1029: 3443 2663 M (w)S 

1030: 3357 2663 M (f)S 

1031: 2833 2793 M (w)S 

1032: 2715 2793 M (w)S 

1033: 2630 2793 M (w)S 

1034: 2540 2793 M (P)S 

1035: /F5S38I13Y00 F5 [56.754 0 19.41 -56.754 0 0 ] mFS

1036: F5S38I13Y00 Ji 

1037: 3289 2867 M (l)S 

1038: /F5S46I18Y00 F5 [70.941 0 24.262 -70.941 0 0 ] mFS

1039: F5S46I18Y00 Ji 

1040: 3274 2663 M (l)S 

1041: F3S4B Ji 

1042: 3831 2777 M ( )S 

1043: 2525 2975 M (The )[45 38 33  0]xS 

1044: /F5S4BI19Y00 F5 [75 0 25.648 -75 0 0 ] mFS

1045: F5S4BI19Y00 Ji 

1046: 2677 2975 M (l)S 

1047: F1S43 Ji 

1048: 2719 2988 M (j)S 

1049: F3S4B Ji 

1050: 2738 2975 M ( are real)[36 33 25 33 36 25 33 33  0]xS 

1051: 3013 2975 M (-)S 

1052: 3038 2975 M (valued constants learned in such a way as to )[37 33 21 38 33 38 36 33 38 38 29 21 33 38 21 29 36 21 33 33 25 38 33 38 36 21 38 36 29 38 33 38

1053: 36 33 36 53 33 36 36 33 29 35 21 38  0]xS 

1054: 2525 3098 M (optimize the perplexity o)[38 38 21 21 57 21 33 33 44 21 38 33 44 38 33 25 38 21 33 37 21 21 36 43  0]xS 

1055: 3351 3098 M (f training data.  )[24 43 21 25 33 21 38 21 38 37 43 38 33 21 33 19 43  0]xS 

1056: F1S4B Ji 

1057: 3925 3098 M (Z)S 

1058: /F5S43I16Y00 F5 [67 0 22.914 -67 0 0 ] mFS

1059: F5S43I16Y00 Ji 

1060: 3967 3111 M (l)S 

1061: F1S4B Ji 

1062: 4005 3098 M (\(w)[24  0]xS 

1063: F3S43 Ji 

1064: 4079 3111 M (1)S 

1065: F1S4B Ji 

1066: 4113 3098 M <8577>[67  0]xS 

1067: F1S43 Ji 

1068: 4230 3111 M (i)S 

1069: 4249 3111 M (-)S 

1070: F3S43 Ji 

1071: 4271 3111 M (1)S 

1072: F1S4B Ji 

1073: 4305 3098 M (\))S 

1074: F3S4B Ji 

1075: 4330 3098 M ( is a )[43 21 29 43 33  0]xS 

1076: 2525 3190 M (normalizing constant so that the sum of all probabilities is 1, )

1077: [38 38 25 57 33 21 21 33 21 38 37 35 33 38 38 29 21 33 38 21 35 29 38 35 21 38 33 21 35 21 38 33

1078: 34 29 38 57 34 38 24 34 33 21 21 34 38 25 38 38 33 38 21 21 21 21 21 33 29 34 21 29 34 38 19  0]xS 

1079: 2525 3390 M (simply set equal to)[29 21 57 38 21 36 35 29 33 21 35 33 38 38 33 21 35 21  0]xS 

1080: /F6S7C F6 [124.23 0 0 -124.23 0 0 ] mFS

1081: F6S7C Ji 

1082: 3153 3410 M <07>S 

1083: 3427 3410 M <07>S 

1084: /F6S52 F6 [82.82 0 0 -82.82 0 0 ] mFS

1085: F6S52 Ji 

1086: 4205 3438 M <01>S 

1087: 4205 3359 M <01>S 

1088: 4205 3505 M <02>S 

1089: 4205 3306 M <03>S 

1090: 3384 3438 M <04>S 

1091: 3384 3359 M <04>S 

1092: 3384 3505 M <05>S 

1093: 3384 3306 M <06>S 

1094: /F5S42 F5 [66.254 0 0 -66.254 0 0 ] mFS

1095: F5S42 Ji 

1096: 4092 3412 M (-)S 

1097: /F1S42 F1 [66.254 0 0 -66.254 0 0 ] mFS

1098: F1S42 Ji 

1099: 3177 3488 M (w)S 

1100: 3469 3488 M (j)S 

1101: 4068 3412 M (i)S 

1102: 3694 3412 M (j)S 

1103: 3599 3412 M (j)S 

1104: /F1S52 F1 [82.82 0 0 -82.82 0 0 ] mFS

1105: F1S52 Ji 

1106: 4011 3391 M (w)S 

1107: 3856 3391 M (w)S 

1108: 3763 3391 M (w)S 

1109: 3647 3391 M (f)S 

1110: /F3S52 F3 [82.82 0 0 -82.82 0 0 ] mFS

1111: F3S52 Ji 

1112: 4169 3391 M (\))S 

1113: 3947 3391 M (...)[21 21  0]xS 

1114: 3819 3391 M (,)S 

1115: 3731 3391 M (\()S 

1116: 3259 3391 M (exp)[37 41  0]xS 

1117: /F3S42 F3 [66.254 0 0 -66.254 0 0 ] mFS

1118: F3S42 Ji 

1119: 4126 3412 M (1)S 

1120: 3908 3412 M (1)S 

1121: /F5S52I1CY00 F5 [82.82 0 28.324 -82.82 0 0 ] mFS

1122: F5S52I1CY00 Ji 

1123: 3534 3391 M (l)S 

1124: F3S4B Ji 

1125: 4259 3390 M (. The )[19 35 45 38 33  0]xS 

1126: F1S4B Ji 

1127: 4464 3390 M (f)S 

1128: /F1S32 F1 [50 0 0 -50 0 0 ] mFS

1129: F1S32 Ji 

1130: 4485 3403 M (j)S 

1131: F3S4B Ji 

1132: 4499 3390 M ( )S 

1133: 2525 3600 M (repre)[25 33 38 25  0]xS 

1134: 2679 3600 M (sent a large set of indicator functions that always have the )

1135: [29 33 38 21 28 33 28 21 33 25 37 33 28 29 33 21 27 38 24 27 21 38 38 21 33 33 21 38 25 27 24 38

1136: 38 33 21 21 38 38 29 27 21 38 33 21 27 33 21 53 33 36 29 27 38 33 37 33 27 21 38 33  0]xS 

1137: 2525 3708 M (value 1 or 0.  For instance, we could )[37 33 21 38 33 20 38 20 38 25 19 38 19 19 19 42 38 25 19 21 38 29 21 33 38 33 33 19 19 53 33 19

1138: 33 38 38 21 38  0]xS 

1139: 3642 3708 M (use )[38 29 33  0]xS 

1140: F1S4B Ji 

1141: 3761 3708 M (f)S 

1142: F1S43 Ji 

1143: 3782 3721 M (j)S 

1144: F1S4B Ji 

1145: 3801 3708 M (\(\223Tuesday\224, w)[24 42 42 38 33 29 38 38 33 42 19 19  0]xS 

1146: F1S43 Ji 

1147: 4248 3721 M (1)S 

1148: F1S4B Ji 

1149: 4282 3708 M <8577>[67  0]xS 

1150: F1S43 Ji 

1151: 4399 3721 M (i)S 

1152: 4418 3721 M (-)S 

1153: F3S43 Ji 

1154: 4440 3721 M (1)S 

1155: F1S4B Ji 

1156: 4474 3708 M (\) )[25  0]xS 

1157: 2525 3823 M (= )[51  0]xS 

1158: F3S4B Ji 

1159: 2596 3823 M (1 if )[38 20 21 24  0]xS 

1160: F1S4B Ji 

1161: 2719 3823 M (w)S 

1162: F1S43 Ji 

1163: 2769 3836 M (i)S 

1164: 2788 3836 M (-)S 

1165: F3S43 Ji 

1166: 2810 3836 M (1)S 

1167: F1S4B Ji 

1168: 2844 3823 M ( = \223on\224 and w)[20 51 20 42 38 38 42 20 38 38 38 20  0]xS 

1169: F1S43 Ji 

1170: 3299 3836 M (i)S 

1171: 3318 3836 M (-)S 

1172: F3S43 Ji 

1173: 3340 3836 M (2)S 

1174: F1S4B Ji 

1175: 3374 3823 M ( = \223meet\224 )[20 51 19 42 54 33 33 21 42  0]xS 

1176: F3S4B Ji 

1177: 3708 3823 M (\(and, implicitly, otherwise )[25 33 38 38 19 19 21 57 38 21 21 33 21 21 21 36 19 19 38 21 38 33 25 53 21 29 33  0]xS 

1178: 2525 3938 M (0\))[38  0]xS 

1179: 2588 3938 M (.  If )[19 34 34 25 24  0]xS 

1180: F5S4BI19Y00 Ji 

1181: 2758 3938 M (l)S 

1182: F1S43 Ji 

1183: 2800 3951 M (j)S 

1184: F1S4B Ji 

1185: 2819 3938 M ( )S 

1186: F3S4B Ji 

1187: 2853 3938 M (were given a positive value, then the probability of )[53 33 25 33 34 37 21 37 33 38 34 33 34 38 38 29 21 21 21 37 33 34 37 33 21 38 33 19 34 21 38 33

1188: 38 34 21 38 33 33 38 25 38 38 33 38 21 21 21 21 36 33 38 24  0]xS 

1189: 2525 4031 M (\223Tuesday\224 in the context of \223meet on\224 would be raised.  By )[33 45 38 33 29 38 33 36 33 36 21 38 36 21 38 33 36 33 38 38 21 33 37 21 36 38 24 36 33 57 33 33

1190: 21 36 38 38 33 36 53 38 38 21 38 36 38 33 35 25 33 21 29 33 38 19 35 35 50 36  0]xS 

1191: 2525 4118 M (making many indicator functions of this t)[57 33 37 21 38 37 28 57 33 38 36 28 21 38 38 21 33 33 21 38 25 28 24 38 38 33 21 21 38 38 29 27

1192: 38 24 27 21 38 21 29 27  0]xS 

1193: 3817 4118 M (ype, we could capture )[36 38 33 19 27 53 33 27 33 38 38 21 38 27 33 33 38 21 38 25 33  0]xS 

1194: 2525 4204 M (all of the information captured by a trigram.  Similarly, we could )

1195: [33 21 21 22 38 24 22 21 38 33 22 21 38 24 38 25 57 33 21 21 38 38 22 33 33 38 21 38 25 33 38 22

1196: 38 36 22 33 21 21 25 21 37 25 33 57 19 21 21 42 21 57 21 21 33 25 21 36 19 21 53 33 21 33 38 38

1197: 21 38  0]xS 

1198: 2525 4312 M (make a bigram indicator function with )[57 33 37 33 29 33 29 38 21 37 25 33 57 29 21 38 38 21 33 33 21 38 25 28 24 38 38 33 21 21 38 38

1199: 28 53 21 21 38  0]xS 

1200: F1S4B Ji 

1201: 3752 4312 M (f)S 

1202: F1S43 Ji 

1203: 3773 4325 M (j)S 

1204: F1S4B Ji 

1205: 3792 4312 M (\(\223Tuesday\224, w)[24 42 42 38 33 29 38 38 33 42 19 28  0]xS 

1206: F3S43 Ji 

1207: 4248 4325 M (1)S 

1208: F1S4B Ji 

1209: 4282 4312 M <8577>[67  0]xS 

1210: F1S43 Ji 

1211: 4399 4325 M (i)S 

1212: 4418 4325 M (-)S 

1213: F3S43 Ji 

1214: 4440 4325 M (1)S 

1215: F1S4B Ji 

1216: 4474 4312 M (\) )[25  0]xS 

1217: 2525 4427 M (= )[51  0]xS 

1218: F3S4B Ji 

1219: 2610 4427 M (1 if )[38 34 21 24  0]xS 

1220: F1S4B Ji 

1221: 2761 4427 M (w)S 

1222: F1S43 Ji 

1223: 2811 4440 M (i)S 

1224: 2830 4440 M (-)S 

1225: F3S43 Ji 

1226: 2852 4440 M (1)S 

1227: F3S4B Ji 

1228: 2886 4427 M ( )S 

1229: F1S4B Ji 

1230: 2920 4427 M (= \223on\224)[51 34 42 38 38  0]xS 

1231: F3S4B Ji 

1232: 3165 4427 M (.  Or we could make a unigram indicator )[19 34 34 54 25 34 53 33 34 33 38 38 21 38 34 57 33 37 33 34 33 34 38 38 21 37 25 33 57 34 21 38

1233: 38 21 33 33 21 38 25  0]xS 

1234: 2525 4542 M (function with the simple )[24 38 38 33 21 21 38 38 49 53 21 21 38 49 21 38 33 48 29 21 57 38 21 33  0]xS 

1235: F1S4B Ji 

1236: 3394 4542 M (f)S 

1237: F1S43 Ji 

1238: 3415 4555 M (j)S 

1239: F1S4B Ji 

1240: 3434 4542 M (\(\223Tuesday\224, w)[24 42 42 38 33 29 38 38 33 42 19 48  0]xS 

1241: /F3S32 F3 [50 0 0 -50 0 0 ] mFS

1242: F3S32 Ji 

1243: 3910 4555 M (1)S 

1244: F1S4B Ji 

1245: 3935 4542 M <8577>[67  0]xS 

1246: F1S43 Ji 

1247: 4052 4555 M (i)S 

1248: 4071 4555 M (-)S 

1249: F3S43 Ji 

1250: 4093 4555 M (1)S 

1251: F1S4B Ji 

1252: 4127 4542 M (\)=)[25  0]xS 

1253: F3S4B Ji 

1254: 4203 4542 M (1)S 

1255: F1S4B Ji 

1256: 4241 4542 M ( )S 

1257: F3S4B Ji 

1258: 4289 4542 M (for all)[24 38 25 48 33 21  0]xS 

1259: F1S4B Ji 

1260: 4499 4542 M ( )S 

1261: 2525 4657 M (w)S 

1262: F3S43 Ji 

1263: 2575 4670 M (1)S 

1264: F1S4B Ji 

1265: 2609 4657 M <8577>[67  0]xS 

1266: F1S43 Ji 

1267: 2726 4670 M (i)S 

1268: 2745 4670 M (-)S 

1269: F3S43 Ji 

1270: 2767 4670 M (1)S 

1271: F1S4B Ji 

1272: 2801 4657 M (.  )[19 47  0]xS 

1273: F3S4B Ji 

1274: 2914 4657 M (In principle, any set of indicator functions that )[25 38 46 38 25 21 38 33 21 38 21 33 19 46 33 38 36 46 29 33 21 46 38 24 46 21 38 38 21 33 33 21

1275: 38 25 46 24 38 38 33 21 21 38 38 29 46 21 38 33 21  0]xS 

1276: 2525 4772 M (depends on )[38 33 38 33 38 38 29 20 38 38  0]xS 

1277: F1S4B Ji 

1278: 2888 4772 M (w, w)[50 19 20  0]xS 

1279: F3S43 Ji 

1280: 3027 4785 M (1)S 

1281: F1S4B Ji 

1282: 3061 4772 M <8577>[67  0]xS 

1283: F1S43 Ji 

1284: 3178 4785 M (i)S 

1285: 3197 4785 M (-)S 

1286: F3S43 Ji 

1287: 3219 4785 M (1)S 

1288: F3S4B Ji 

1289: 3253 4772 M ( can be used, including n)[19 33 33 38 19 38 33 19 38 29 33 38 19 19 21 38 33 21 38 38 21 38 37 19  0]xS 

1290: 4003 4772 M (-)S 

1291: 4028 4772 M (grams, caching, )[37 25 33 57 29 19 19 33 33 33 38 21 38 37 19  0]xS 

1292: 2525 4865 M (skipping n)[29 37 21 38 38 21 38 37 19  0]xS 

1293: 2841 4865 M (-)S 

1294: 2866 4865 M (grams, word triggers, etc.)[37 25 33 57 29 19 19 53 38 25 38 19 21 25 21 37 37 33 25 29 19 19 33 21 33  0]xS 

1295: 3630 4865 M ( )S 

1296: 2638 4956 M (The optimal )[45 38 33 40 38 38 21 21 57 33 21  0]xS 

1297: /F5S4B F5 [75 0 0 -75 0 0 ] mFS

1298: F5S4B Ji 

1299: 3063 4956 M (l)S 

1300: F3S4B Ji 

1301: 3105 4956 M (-)S 

1302: 3130 4956 M (values must be learned.  The algorithm )[37 33 21 38 33 29 40 57 38 29 21 40 38 33 40 21 33 33 25 38 33 38 19 40 39 45 38 33 39 33 21 37

1303: 38 25 21 21 38 57  0]xS 

1304: 4461 4956 M <96>S 

1305: 4499 4956 M ( )S 

1306: 2525 5043 M (Generalized Iterative Scaling )[53 33 38 33 25 33 21 21 33 33 38 34 25 21 33 25 33 21 21 37 33 34 42 33 33 21 21 38 37  0]xS 

1307: 3462 5043 M ([1])[26 38  0]xS 

1308: 3552 5043 M ( )S 

1309: 3586 5043 M <96>S 

1310: 3624 5043 M ( for optimizing their values )[34 24 38 25 34 38 38 21 21 57 21 33 21 38 37 33 21 38 33 21 25 33 37 33 21 38 33 29  0]xS 

1311: 2525 5129 M (\(based on some set of training data\) can be very slow.  It requires )

1312: [25 38 33 29 33 38 21 38 38 21 29 38 57 33 20 29 33 21 20 38 24 20 21 25 33 21 38 21 38 37 20 38

1313: 33 21 33 25 20 33 33 38 20 38 33 20 37 33 25 36 20 29 21 38 53 19 20 20 25 21 20 25 33 38 38 21

1314: 25 33 29  0]xS 

1315: 2525 5216 M (many iterations, and at each iteration, it involves a loop over all )

1316: [57 33 38 36 25 21 21 33 25 33 21 21 38 38 29 19 25 33 38 38 25 33 21 25 33 33 33 38 25 21 21 33

1317: 25 33 21 21 38 38 19 24 21 21 24 21 38 37 38 21 37 33 29 24 33 24 21 38 38 38 24 38 37 33 25 24

1318: 33 21 21  0]xS 

1319: 2525 5302 M (words in the training data.  We give here a very rough sk)[53 38 25 38 29 24 21 38 24 21 38 33 24 21 25 33 21 38 21 38 37 24 38 33 21 33 19 24 24 70 33 24

1320: 37 21 37 33 24 38 33 25 33 24 33 24 37 33 25 36 23 25 38 38 37 38 23 29  0]xS 

1321: 4289 5302 M (etch of )[33 21 33 38 23 38 24  0]xS 

1322: 2525 5388 M (the algorithm, with only the inner loop presented in any detail.)

1323: [21 38 33 19 33 21 37 38 25 21 21 38 57 19 19 53 21 21 38 19 38 38 21 36 19 21 38 33 19 21 38 38

1324: 33 25 19 21 38 38 38 19 38 25 33 29 33 38 21 33 38 19 21 38 19 33 38 36 19 38 33 21 33 21 21  0]xS 

1325: 4405 5388 M ( )S 

1326: 2638 5474 M (The inner loop of this code, and the most time)[45 38 33 33 21 38 38 33 25 33 21 38 38 38 33 38 24 32 21 38 21 29 32 33 38 38 33 19 32 33 38 38

1327: 32 21 38 33 32 57 38 29 21 32 21 21 57  0]xS 

1328: 4145 5474 M (-)S 

1329: 4170 5474 M (consuming )[33 38 38 29 38 57 21 38 37  0]xS 

1330: 2525 5561 M (part is lines )[38 33 25 21 25 21 29 25 21 21 38 33 29  0]xS 

1331: 2909 5561 M (4)S 

1332: 2947 5561 M ( to )[25 21 38  0]xS 

1333: 3055 5561 M (12)[38  0]xS 

1334: 3131 5561 M (.  Notice that the inner loop contains several )[19 24 24 54 38 21 21 33 33 24 21 38 33 21 24 21 38 33 24 21 38 38 33 25 24 21 38 38 38 24 33 38

1335: 38 21 33 21 38 29 24 29 33 37 33 25 33 21  0]xS 

1336: 2525 5647 M (loops over all words in the vocabulary \(lines )[21 38 38 38 29 19 38 37 33 25 19 33 21 21 19 53 38 25 38 29 19 21 38 19 21 38 33 19 37 38 33 33

1337: 38 38 21 33 25 36 19 25 21 21 38 33 29  0]xS 

1338: 3882 5647 M (4)S 

1339: 3920 5647 M (, )[19  0]xS 

1340: 3958 5647 M (7)S 

1341: 3996 5647 M ( and )[19 33 38 38  0]xS 

1342: 4143 5647 M (8)S 

1343: 4181 5647 M (\).  Notice )[25 19 19 19 54 38 21 21 33 33  0]xS 

1344: 4482 5647 M ( )S 

1345: LH

1346: (%%[Page: 1]%%) = 

1347: %%PageTrailer

1348: 

1349: %%Page: 2 2

1350: %%PageBoundingBox: 12 12 600 780

1351: %%EndPageComments

1352: %%BeginPageSetup

1353: /DeviceRGB dup setcolorspace /colspABC exch def

1354: mysetup concat colspRefresh

1355: %%EndPageSetup

1356: 

1357: /DeviceGray dup setcolorspace /colspABC exch def

1358: 0 0 scol F3S53 Ji 

1359: 350 6179 M ( )S 

1360: 2150 6179 M ( )S 

1361: F3S4B Ji 

1362: 350 2398 M (that the sum in line )[21 38 33 21 34 21 38 33 34 29 38 57 34 21 38 34 21 21 38 33  0]xS 

1363: 1021 2398 M (5)S 

1364: 1059 2398 M ( is typically bounded by the number of )[34 21 29 34 21 36 38 21 33 33 21 21 36 34 38 38 38 38 38 33 38 34 38 36 34 21 38 33 34 38 38 57

1365: 38 33 25 33 38 24  0]xS 

1366: 350 2485 M (different t)[38 21 24 24 33 25 33 38 21 41  0]xS 

1367: 669 2485 M (ypes of indicator functions.  In particular, a given )[36 38 33 29 41 38 24 40 21 38 38 21 33 33 21 38 25 40 24 38 38 33 21 21 38 38 29 19 40 40 25 38

1368: 40 38 33 25 21 21 33 38 21 33 25 19 40 33 40 37 21 37 33 38  0]xS 

1369: 350 2571 M (system will typically have only a few types of indicators )[29 36 29 21 33 57 29 53 21 21 21 29 21 36 38 21 33 33 21 21 36 29 38 33 37 33 29 38 38 21 36 29

1370: 33 29 24 33 53 28 21 36 38 33 29 28 38 24 28 21 38 38 21 33 33 21 38 25 29  0]xS 

1371: 2150 2571 M <96>S 

1372: 2188 2571 M ( e.g. )[28 33 19 37 19  0]xS 

1373: 350 2657 M (unigram, bigram, and trigram )[38 38 21 37 25 33 57 19 29 38 21 37 25 33 57 19 29 33 38 38 28 21 25 21 37 25 33 57  0]xS 

1374: 1290 2657 M <96>S 

1375: 1328 2657 M ( and typically, for each of these )[28 33 38 38 28 21 36 38 21 33 33 21 21 36 19 28 24 38 25 28 33 33 33 38 28 38 24 28 21 38 33 29

1376: 33  0]xS 

1377: 350 2743 M (types, and for a given word w, and history, there can be only one )

1378: [21 36 38 33 29 19 22 33 38 38 21 24 38 25 21 33 21 37 21 37 33 38 21 53 38 25 38 21 53 19 21 33

1379: 38 38 21 38 21 29 21 38 25 36 19 21 21 38 33 25 33 21 33 33 38 21 38 33 21 38 38 21 36 21 38 38

1380: 33  0]xS 

1381: 350 2830 M (non)[38 38  0]xS 

1382: 464 2830 M (-)S 

1383: 489 2830 M (zero)[33 33 25  0]xS 

1384: 618 2830 M ( indicator function.  This means that the sum in line )[28 21 38 38 21 33 33 21 38 25 28 24 38 38 33 21 21 38 38 19 28 28 45 38 21 29 28 57 33 33 38 29

1385: 28 21 38 33 21 28 21 38 33 28 29 38 57 28 21 38 27 21 21 38 33  0]xS 

1386: 2286 2830 M (5)S 

1387: 2324 2830 M ( )S 

1388: 350 2916 M (and the loop in line )[33 38 38 31 21 38 33 31 21 38 38 38 31 21 38 31 21 21 38 33  0]xS 

1389: 1013 2916 M (9)S 

1390: 1051 2916 M ( are bounded by the number of types of )[31 33 25 33 31 38 38 38 38 38 33 38 31 38 36 31 21 38 33 31 38 38 57 38 33 25 30 38 24 30 21 36

1391: 38 33 29 30 38 24  0]xS 

1392: 350 3002 M (indicator function.  Overall, then, the inner loop of lines )[21 38 38 21 33 33 21 38 25 24 24 38 38 33 21 21 38 38 19 24 24 54 37 33 25 33 21 21 19 24 21 38

1393: 33 38 19 24 21 38 33 24 21 38 38 33 25 24 21 38 38 38 24 38 24 24 21 21 38 33 29  0]xS 

1394: 2105 3002 M (4)S 

1395: 2143 3002 M ( to )[23 21 38  0]xS 

1396: 2248 3002 M (12)[38  0]xS 

1397: 2324 3002 M ( )S 

1398: 350 3088 M (is typically bounded by the number of types of indicator function, )

1399: [21 29 19 21 36 38 21 33 33 21 21 36 19 38 38 38 38 38 33 38 19 38 36 19 21 38 33 19 38 38 57 38

1400: 33 25 19 38 24 19 21 36 38 33 29 19 38 24 19 21 38 38 21 33 33 21 38 25 18 24 38 38 33 21 21 38

1401: 38 19  0]xS 

1402: 350 3175 M (times the vocabulary size.  This means that decreasing the )[21 21 57 33 29 46 21 38 33 45 37 38 33 33 38 38 21 33 25 36 45 29 21 33 33 19 45 45 45 38 21 29

1403: 45 57 33 33 38 29 45 21 38 33 21 45 38 33 33 25 33 33 29 21 38 37 45 21 38 33  0]xS 

1404: 350 3261 M (vocabulary size leads to a decrease in the runtime of the inner )

1405: [37 38 33 33 38 38 21 33 25 36 30 29 21 33 33 30 21 33 33 38 29 30 21 38 30 33 30 38 33 33 25 33

1406: 33 29 33 30 21 38 30 21 38 33 30 25 38 38 21 21 57 33 30 38 24 30 21 38 33 30 21 38 38 33 25  0]xS 

1407: 350 3347 M (loop.  Certain types )[21 38 38 38 19 38 38 50 33 25 21 33 21 38 38 21 36 38 33 29  0]xS 

1408: 1034 3347 M (of indicator functions \(e.g. triggers\) and )[38 24 37 21 38 38 21 33 33 21 38 25 37 24 38 38 33 21 21 38 38 29 37 25 33 19 37 19 37 21 25 21

1409: 37 37 33 25 29 25 37 33 38 38  0]xS 

1410: 350 3455 M (optimizations to line )[38 38 21 21 57 21 33 33 21 21 38 38 29 27 21 38 27 21 21 38 33  0]xS 

1411: 1012 3455 M (5)S 

1412: 1050 3455 M ( \(summing only over non)[27 25 29 38 57 57 21 38 37 27 38 38 21 36 26 38 37 33 25 26 38 38  0]xS 

1413: 1838 3455 M (-)S 

1414: 1863 3455 M (zero f)[33 33 25 38 26  0]xS 

1415: F3S43 Ji 

1416: 2042 3468 M (j)S 

1417: F3S4B Ji 

1418: 2061 3455 M (\) change )[25 26 33 38 33 38 37 33  0]xS 

1419: 350 3548 M (the exact analysis of run)[21 38 33 31 33 37 33 33 21 31 33 38 33 21 36 29 21 29 31 38 24 31 25 38  0]xS 

1420: 1126 3548 M (-)S 

1421: 1151 3548 M (time, but not the intuition that inner)[21 21 57 33 19 31 38 38 21 31 38 38 21 31 21 38 33 31 21 38 21 38 21 21 21 38 38 31 21 38 33 21

1422: 31 21 38 38 33  0]xS 

1423: 2299 3548 M (-)S 

1424: 350 3634 M (loop run)[21 38 38 38 19 25 38  0]xS 

1425: 605 3634 M (-)S 

1426: 630 3634 M (time is roughly proportional to vocabulary size)[21 21 57 33 19 21 29 19 25 38 38 37 38 21 36 19 38 25 38 38 38 25 21 21 38 38 33 21 19 21 38 19

1427: 37 38 33 33 38 38 21 33 25 36 19 29 21 33  0]xS 

1428: 2040 3634 M (.)S 

1429: 2059 3634 M ( )S 

1430: F0S64 Ji 

1431: 654 3797 M (2.)[50  0]xS 

1432: 729 3797 M ( )S 

1433: 789 3797 M (CLASS)[72 61 78 61  0]xS 

1434: 1122 3797 M (-)S 

1435: 1155 3797 M (BASED SPEEDUP)[67 78 61 61 83 25 61 61 61 61 83 78  0]xS 

1436: 1996 3797 M ( )S 

1437: F3S4B Ji 

1438: 350 3941 M (We now describe our spee)[70 33 36 38 38 53 36 38 33 29 33 25 21 38 33 36 38 38 25 36 29 38 33  0]xS 

1439: 1210 3941 M (dup.  We assign each word in the )[38 38 38 19 36 36 70 33 36 33 29 29 21 37 38 36 33 33 33 38 35 53 38 25 38 35 21 38 35 21 38 33

1440:  0]xS 

1441: 350 4028 M (vocabulary to a unique class.  For instance, cat and dog might be )

1442: [37 38 33 33 38 38 21 33 25 36 22 21 38 22 33 22 38 38 21 38 38 33 22 33 21 33 29 29 19 21 21 42

1443: 38 25 21 21 38 29 21 33 38 33 33 19 21 33 33 21 21 33 38 38 21 38 38 37 21 57 21 37 38 21 21 38

1444: 33  0]xS 

1445: 350 4114 M (in the class of ANIMAL, while Tuesday and Wednesday might )[21 38 28 21 38 33 28 33 21 33 29 29 28 38 24 28 53 54 25 67 53 45 19 28 53 38 21 21 33 28 45 38

1446: 33 29 38 33 36 28 33 38 38 28 70 33 38 38 33 29 38 33 36 27 57 21 37 38 21  0]xS 

1447: 350 4200 M (be in the class of WEEKDAY.   Next, we observe that)[38 33 19 21 38 19 21 38 33 19 33 21 33 29 29 19 38 24 19 70 46 46 54 54 53 54 19 19 19 19 54 33

1448: 37 21 19 19 53 33 19 38 38 29 33 25 37 33 19 21 38 33  0]xS 

1449: 1980 4200 M ( )S 

1450: /F3S45 F3 [69.332 0 0 -69.332 0 0 ] mFS

1451: F3S45 Ji 

1452: 2258 4287 M (\)\))[23  0]xS 

1453: 2184 4287 M (\()S 

1454: 2011 4287 M (,)S 

1455: 1835 4287 M (...)[17 17  0]xS 

1456: 1733 4287 M (|)S 

1457: 1642 4287 M (\()S 

1458: 1506 4287 M (\))S 

1459: 1328 4287 M (...)[17 17  0]xS 

1460: 1226 4287 M (|)S 

1461: 1185 4287 M (\))S 

1462: 1111 4287 M (\()S 

1463: 938 4287 M (\()S 

1464: 782 4287 M (\))S 

1465: 604 4287 M (...)[17 17  0]xS 

1466: 503 4287 M (|)S 

1467: 412 4287 M (\()S 

1468: /F3S37 F3 [55.465 0 0 -55.465 0 0 ] mFS

1469: F3S37 Ji 

1470: 1981 4305 M (1)S 

1471: 1807 4305 M (1)S 

1472: 1474 4305 M (1)S 

1473: 1300 4305 M (1)S 

1474: 751 4305 M (1)S 

1475: 576 4305 M (1)S 

1476: /F1S45 F1 [69.332 0 0 -69.332 0 0 ] mFS

1477: F1S45 Ji 

1478: 2211 4287 M (w)S 

1479: 2037 4287 M (class)[31 19 35 27  0]xS 

1480: 1888 4287 M (w)S 

1481: 1768 4287 M (w)S 

1482: 1670 4287 M (w)S 

1483: 1596 4287 M (P)S 

1484: 1381 4287 M (w)S 

1485: 1261 4287 M (w)S 

1486: 1138 4287 M (w)S 

1487: 963 4287 M (class)[31 19 35 27  0]xS 

1488: 892 4287 M (P)S 

1489: 658 4287 M (w)S 

1490: 537 4287 M (w)S 

1491: 439 4287 M (w)S 

1492: 366 4287 M (P)S 

1493: /F1S37 F1 [55.465 0 0 -55.465 0 0 ] mFS

1494: F1S37 Ji 

1495: 1931 4305 M (i)S 

1496: 1425 4305 M (i)S 

1497: 701 4305 M (i)S 

1498: /F5S37 F5 [55.465 0 0 -55.465 0 0 ] mFS

1499: F5S37 Ji 

1500: 1952 4305 M (-)S 

1501: 1445 4305 M (-)S 

1502: 722 4305 M (-)S 

1503: : 1541 4218 38 84 rc /F5S45 F5 [69.332 0 0 -69.332 0 0 ] mFS

1504: F5S45 Ji 

1505: 1541 4287 M <B4>S 

1506: ; : 828 4218 38 84 rc F5S45 Ji 

1507: 828 4287 M (=)S 

1508: ; F3S4B Ji 

1509: 2322 4280 M ( )S 

1510: 350 4390 M (This equality holds becaus)[45 38 21 29 22 33 38 38 33 21 21 21 36 22 38 38 21 38 29 22 38 33 33 33 38  0]xS 

1511: 1158 4390 M (e each word is in a single class, and is )[33 22 33 33 33 38 22 53 38 25 38 22 21 29 22 21 38 22 33 22 29 21 38 37 21 33 22 33 21 33 29 29

1512: 19 22 33 38 38 22 21 29  0]xS 

1513: 350 4477 M (easily proven.  Conceptually, it says that we can decompose the )

1514: [33 33 29 21 21 36 26 38 25 38 37 33 38 19 26 26 50 38 38 33 33 38 21 38 33 21 21 36 19 25 21 21

1515: 25 29 33 36 29 25 21 38 33 21 25 53 33 25 33 33 38 25 38 33 33 38 57 38 38 29 33 25 21 38 33  0]xS 

1516: 350 4563 M (prediction of a word given its history into prediction of its class )

1517: [38 25 33 38 21 33 21 21 38 38 25 38 24 25 33 25 53 38 25 38 25 37 21 37 33 38 25 21 21 29 25 38

1518: 21 29 21 38 25 36 25 21 38 21 38 25 38 25 33 38 21 33 21 21 38 38 24 38 24 24 21 21 29 24 33 21

1519: 33 29 29  0]xS 

1520: 350 4649 M (given the history, and the probability of the word given the )

1521: [37 21 37 33 38 40 21 38 33 40 38 21 29 21 38 25 36 19 39 33 38 38 39 21 38 33 39 38 25 38 38 33

1522: 38 21 21 21 21 36 39 38 24 39 21 38 33 39 53 38 25 38 39 37 21 37 33 38 39 21 38 33  0]xS 

1523: 350 4735 M (history and the class.)[38 21 29 21 38 25 36 32 33 38 38 32 21 38 33 32 33 21 33 29 29  0]xS 

1524: 1019 4735 M (  For \223true\224 probabilities, this equality is )[32 32 42 38 25 32 33 21 25 38 33 33 32 38 25 38 38 33 38 21 21 21 21 21 33 29 19 31 21 38 21 29

1525: 31 33 38 38 33 21 21 21 36 31 21 29  0]xS 

1526: 350 4822 M (exact.  If the probabilities are not true, but instead are, for )

1527: [33 37 33 33 21 19 41 41 25 24 41 21 38 33 41 38 25 38 38 33 38 21 21 21 21 21 33 29 41 33 25 33

1528: 41 38 38 21 41 21 25 38 33 19 41 38 38 21 41 21 38 29 21 33 33 38 40 33 25 33 19 40 24 38 25  0]xS 

1529: 350 4908 M (instance, the results of estimating a model, or are smoothed, or )

1530: [21 38 29 21 33 38 33 33 19 29 21 38 33 28 25 33 29 38 21 21 29 28 38 24 28 33 29 21 21 57 33 21

1531: 21 38 37 28 33 28 57 38 38 33 21 19 28 38 25 28 33 25 33 28 29 57 38 38 21 38 33 38 19 28 38 25

1532:  0]xS 

1533: 350 4994 M (are the results of computing a max)[33 25 33 40 21 38 33 39 25 33 29 38 21 21 29 39 38 24 39 33 38 57 38 38 21 21 38 37 39 33 39 57

1534: 33  0]xS 

1535: 1507 4994 M (imum entropy model, the )[21 57 38 57 39 33 38 21 25 38 38 36 39 57 38 38 33 21 19 39 21 38 33  0]xS 

1536: 350 5080 M (equality will not be ex)[33 38 38 33 21 21 21 36 20 53 21 21 21 20 38 38 21 20 38 33 20 33  0]xS 

1537: 1025 5080 M (act, but will be a very good approximation.  )[33 33 21 19 20 38 38 21 20 53 21 21 21 20 38 33 20 33 19 37 33 25 36 19 37 38 38 38 19 33 38 38

1538: 25 38 37 21 57 33 21 21 38 38 19 19  0]xS 

1539: 350 5167 M (Indeed, the approximation is so good that typically classing is )

1540: [25 38 38 33 33 38 19 32 21 38 33 32 33 38 38 25 38 37 21 57 33 21 21 38 38 32 21 29 32 29 38 32

1541: 37 38 38 38 32 21 38 33 21 32 21 36 38 21 33 33 21 21 36 32 33 21 33 29 29 21 38 37 32 21 29  0]xS 

1542: 350 5253 M (used to lower the perplexity of models.  )[38 29 33 38 19 21 38 19 21 38 53 33 25 19 21 38 33 19 38 33 25 38 21 33 37 21 21 36 19 38 24 19

1543: 57 38 38 33 21 29 19 19  0]xS 

1544: 1561 5253 M ( )S 

1545: 463 5339 M (This decomposition is the basis for our technique.  Rather )[45 38 21 29 33 38 33 33 38 57 38 38 29 21 21 21 38 38 33 21 29 33 21 38 33 33 38 33 29 21 29 33

1546: 24 38 25 32 38 38 25 32 21 33 33 38 38 21 38 38 33 19 32 32 50 33 21 38 33 25  0]xS 

1547: 350 5425 M (than create a single maximum entropy model, we)[21 38 33 38 39 33 25 33 33 21 33 38 33 38 29 21 38 37 21 33 38 57 33 37 21 57 38 57 38 33 38 21

1548: 25 38 38 36 38 57 38 38 33 21 19 38 53  0]xS 

1549: 1958 5425 M ( create two )[38 33 25 33 33 21 33 38 21 53 38  0]xS 

1550: 350 5512 M (different models, the first of which predicts the class of a word )

1551: [38 21 24 24 33 25 33 38 21 28 57 38 38 33 21 29 19 28 21 38 33 28 24 21 25 29 21 28 38 24 27 53

1552: 38 21 33 38 27 38 25 33 38 21 33 21 29 27 21 38 33 27 33 21 33 29 29 27 38 24 27 33 27 53 38 25

1553: 38  0]xS 

1554: 350 5620 M (given its context )[37 21 37 33 38 25 21 21 29 25 33 38 38 21 33 37 21  0]xS 

1555: F1S4B Ji 

1556: 883 5620 M (P)S 

1557: F3S4B Ji 

1558: 929 5620 M (\()S 

1559: F1S4B Ji 

1560: 954 5620 M (class)[33 21 38 29  0]xS 

1561: F3S4B Ji 

1562: 1104 5620 M (\()S 

1563: F1S4B Ji 

1564: 1129 5620 M (w)S 

1565: F3S4B Ji 

1566: 1179 5620 M (\))S 

1567: F1S4B Ji 

1568: 1204 5620 M (|w)[21  0]xS 

1569: F3S43 Ji 

1570: 1275 5633 M (1)S 

1571: F1S4B Ji 

1572: 1309 5620 M <8577>[67  0]xS 

1573: F1S43 Ji 

1574: 1426 5633 M (i)S 

1575: 1445 5633 M (-)S 

1576: F3S43 Ji 

1577: 1467 5633 M (1)S 

1578: F3S4B Ji 

1579: 1501 5620 M (\), and the second of which )[25 19 25 33 38 38 25 21 38 33 25 29 33 33 38 38 38 25 38 24 24 53 38 21 33 38  0]xS 

1580: 2525 590 M (predicts a word given its class and its context, )[38 25 33 38 21 33 21 29 29 33 29 53 38 25 38 29 37 21 37 33 38 29 21 21 29 28 33 21 33 29 29 28

1581: 33 38 38 28 21 21 29 28 33 38 38 21 33 37 21 19  0]xS 

1582: F1S4B Ji 

1583: 4008 590 M (P)S 

1584: F3S4B Ji 

1585: 4054 590 M (\()S 

1586: F1S4B Ji 

1587: 4079 590 M (w | w)[50 28 21 28  0]xS 

1588: F3S43 Ji 

1589: 4256 603 M (1)S 

1590: F1S4B Ji 

1591: 4290 590 M <8577>[67  0]xS 

1592: F1S43 Ji 

1593: 4407 603 M (i)S 

1594: 4426 603 M (-)S 

1595: F3S43 Ji 

1596: 4448 603 M (1)S 

1597: F1S43 Ji 

1598: 4482 603 M (, )[17  0]xS 

1599: F1S4B Ji 

1600: 2525 705 M (class)[33 21 38 29  0]xS 

1601: F3S4B Ji 

1602: 2675 705 M (\()S 

1603: F1S4B Ji 

1604: 2700 705 M (w)S 

1605: F1S43 Ji 

1606: 2750 718 M (i)S 

1607: F3S4B Ji 

1608: 2769 705 M (\)\).  The process of training each of these )[25 25 19 27 27 45 38 33 27 38 25 38 33 33 29 29 27 38 24 27 21 25 33 21 38 21 38 37 27 33 33 33

1609: 38 27 38 24 26 21 38 33 29 33  0]xS 

1610: 4069 705 M (two models is )[21 53 38 26 57 38 38 33 21 29 26 21 29  0]xS 

1611: 2525 798 M (completely separate. If we have 100 classes in our system, then )

1612: [33 38 57 38 21 33 21 33 21 36 27 29 33 38 33 25 33 21 33 19 27 25 24 27 53 33 27 38 33 37 33 27

1613: 38 38 38 27 33 21 33 29 29 33 29 27 21 38 27 38 38 25 26 29 36 29 21 33 57 19 26 21 38 33 38  0]xS 

1614: 2525 884 M (the inner loop of the training code for predicting the class is )

1615: [21 38 33 35 21 38 38 33 25 35 21 38 38 38 35 38 24 35 21 38 33 35 21 25 33 21 38 21 38 37 35 33

1616: 38 38 33 35 24 38 25 35 38 25 33 38 21 33 21 21 38 37 35 21 38 33 34 33 21 33 29 29 34 21 29  0]xS 

1617: 2525 970 M (bounded by a factor of 100, rather than a factor of the voc)[38 38 38 38 38 33 38 20 38 36 20 33 20 24 33 33 21 38 25 20 38 24 20 38 38 38 19 20 25 33 21 38

1618: 33 25 20 21 38 33 38 20 33 20 24 33 33 21 38 25 20 38 24 20 21 38 33 19 37 38  0]xS 

1619: 4275 970 M (abulary )[33 38 38 21 33 25 36  0]xS 

1620: 2525 1057 M (size.  Thus, this model can be computed relati)[29 21 33 33 19 19 19 45 38 38 29 19 19 21 38 21 29 19 57 38 38 33 21 19 33 33 38 19 38 33 19 33

1621: 38 57 38 38 21 33 38 19 25 33 21 33 21  0]xS 

1622: 3902 1057 M (vely quickly. )[37 33 21 36 19 38 38 21 33 37 21 36 19  0]xS 

1623: 4310 1057 M ( )S 

1624: 2638 1143 M (Next, consider the unigram, bigram and trigram indicators for )

1625: [54 33 37 21 19 20 33 38 38 29 21 38 33 25 20 21 38 33 20 38 38 21 37 25 33 57 19 20 38 21 37 25

1626: 33 57 20 33 38 38 20 21 25 21 37 25 33 57 20 21 38 38 21 33 33 21 38 25 29 19 24 38 25  0]xS 

1627: 2525 1229 M (a class)[33 31 33 21 33 29  0]xS 

1628: 2734 1229 M (-)S 

1629: 2759 1229 M (based model.  An example unigram indicator would be )[38 33 29 33 38 31 57 38 38 33 21 19 31 31 53 38 31 33 37 33 57 38 21 33 30 38 38 21 37 25 33 57

1630: 30 21 38 38 21 33 33 21 38 25 30 53 38 38 21 38 30 38 33  0]xS 

1631: F1S4B Ji 

1632: 2525 1337 M (f)S 

1633: F1S43 Ji 

1634: 2546 1350 M (j)S 

1635: F3S4B Ji 

1636: 2565 1337 M (\()S 

1637: F1S4B Ji 

1638: 2590 1337 M (\223Tuesday\224, w)[42 42 38 33 29 38 38 33 42 19 49  0]xS 

1639: F3S43 Ji 

1640: 3043 1350 M (1)S 

1641: F1S4B Ji 

1642: 3077 1337 M <8577>[67  0]xS 

1643: F1S43 Ji 

1644: 3194 1350 M (i)S 

1645: 3213 1350 M (-)S 

1646: F3S43 Ji 

1647: 3235 1350 M (1)S 

1648: F1S4B Ji 

1649: 3269 1337 M (,class)[19 33 21 38 29  0]xS 

1650: F3S4B Ji 

1651: 3438 1337 M (\()S 

1652: F1S4B Ji 

1653: 3463 1337 M (w)S 

1654: F1S43 Ji 

1655: 3513 1350 M (i)S 

1656: F3S4B Ji 

1657: 3532 1337 M (\)\))[25  0]xS 

1658: F1S4B Ji 

1659: 3582 1337 M (=)S 

1660: F3S4B Ji 

1661: 3633 1337 M (1)S 

1662: F1S4B Ji 

1663: 3671 1337 M ( )S 

1664: F3S4B Ji 

1665: 3720 1337 M (if )[21 24  0]xS 

1666: F1S4B Ji 

1667: 3814 1337 M (class)[33 21 38 29  0]xS 

1668: F3S4B Ji 

1669: 3964 1337 M (\()S 

1670: F1S4B Ji 

1671: 3989 1337 M (w)S 

1672: F1S43 Ji 

1673: 4039 1350 M (i)S 

1674: F3S4B Ji 

1675: 4058 1337 M (\))S 

1676: F1S4B Ji 

1677: 4083 1337 M (=WEEKDAY. )[51 62 46 46 50 54 46 42 19  0]xS 

1678: F3S4B Ji 

1679: 4518 1337 M ( )S 

1680: 2525 1452 M (For the bigram indicator, we would have )[42 38 25 19 21 38 33 19 38 21 37 25 33 57 19 21 38 38 21 33 33 21 38 25 19 19 53 33 19 53 38 38

1681: 21 38 19 38 33 37 33  0]xS 

1682: F1S4B Ji 

1683: 3767 1452 M (f)S 

1684: F1S43 Ji 

1685: 3788 1465 M (j)S 

1686: F3S4B Ji 

1687: 3807 1452 M (\()S 

1688: F1S4B Ji 

1689: 3832 1452 M (\223Tuesday\224, w)[42 42 38 33 29 38 38 33 42 19 18  0]xS 

1690: F3S43 Ji 

1691: 4254 1465 M (1)S 

1692: F1S4B Ji 

1693: 4288 1452 M <8577>[67  0]xS 

1694: F1S43 Ji 

1695: 4405 1465 M (i)S 

1696: 4424 1465 M (-)S 

1697: F3S43 Ji 

1698: 4446 1465 M (1)S 

1699: F1S4B Ji 

1700: 4480 1452 M (, )[19  0]xS 

1701: 2525 1567 M (class)[33 21 38 29  0]xS 

1702: F3S4B Ji 

1703: 2675 1567 M (\()S 

1704: F1S4B Ji 

1705: 2700 1567 M (w)S 

1706: F1S43 Ji 

1707: 2750 1580 M (i)S 

1708: F3S4B Ji 

1709: 2769 1567 M (\)\))[25  0]xS 

1710: F1S4B Ji 

1711: 2819 1567 M ( = )[28 51  0]xS 

1712: F3S4B Ji 

1713: 2926 1567 M (1 if )[38 28 21 24  0]xS 

1714: F1S4B Ji 

1715: 3065 1567 M (class)[33 21 38 29  0]xS 

1716: F3S4B Ji 

1717: 3215 1567 M (\()S 

1718: F1S4B Ji 

1719: 3240 1567 M (w)S 

1720: F1S43 Ji 

1721: 3290 1580 M (i)S 

1722: F3S4B Ji 

1723: 3309 1567 M (\))S 

1724: F1S4B Ji 

1725: 3334 1567 M (=WEEKDAY)[51 62 46 46 50 54 46  0]xS 

1726: F3S32 Ji 

1727: 3731 1580 M ( )S 

1728: F3S4B Ji 

1729: 3753 1567 M (and )[33 38 38  0]xS 

1730: F1S4B Ji 

1731: 3890 1567 M (w)S 

1732: F1S43 Ji 

1733: 3940 1580 M (i)S 

1734: 3959 1580 M (-)S 

1735: F3S43 Ji 

1736: 3981 1580 M (1)S 

1737: F1S4B Ji 

1738: 4015 1567 M ( = \223on\224)[28 51 28 42 38 38  0]xS 

1739: F3S4B Ji 

1740: 4282 1567 M (; and a )[21 27 33 38 38 27 33  0]xS 

1741: 2525 1682 M (trigram indicator would be )[21 25 21 37 25 33 57 25 21 38 38 21 33 33 21 38 25 24 53 38 38 21 38 24 38 33  0]xS 

1742: F1S4B Ji 

1743: 3368 1682 M (f)S 

1744: F1S43 Ji 

1745: 3389 1695 M (j)S 

1746: F3S4B Ji 

1747: 3408 1682 M (\()S 

1748: F1S4B Ji 

1749: 3433 1682 M (\223Tuesday\224, w)[42 42 38 33 29 38 38 33 42 19 24  0]xS 

1750: F3S43 Ji 

1751: 3861 1695 M (1)S 

1752: F1S4B Ji 

1753: 3895 1682 M <8577>[67  0]xS 

1754: F1S43 Ji 

1755: 4012 1695 M (i)S 

1756: 4031 1695 M (-)S 

1757: F3S43 Ji 

1758: 4053 1695 M (1)S 

1759: F1S4B Ji 

1760: 4087 1682 M (, class)[19 24 33 21 38 29  0]xS 

1761: F3S4B Ji 

1762: 4280 1682 M (\()S 

1763: F1S4B Ji 

1764: 4305 1682 M (w)S 

1765: F1S43 Ji 

1766: 4355 1695 M (i)S 

1767: F3S4B Ji 

1768: 4374 1682 M (\)\))[25  0]xS 

1769: F1S4B Ji 

1770: 4424 1682 M ( = )[24 51  0]xS 

1771: F3S4B Ji 

1772: 2525 1797 M (1 if )[38 19 21 24  0]xS 

1773: F1S4B Ji 

1774: 2646 1797 M (class)[33 21 38 29  0]xS 

1775: F3S4B Ji 

1776: 2796 1797 M (\()S 

1777: F1S4B Ji 

1778: 2821 1797 M (w)S 

1779: F1S43 Ji 

1780: 2871 1810 M (i)S 

1781: F3S4B Ji 

1782: 2890 1797 M (\))S 

1783: F1S4B Ji 

1784: 2915 1797 M (=WEEKDAY)[51 62 46 46 50 54 46  0]xS 

1785: F3S4B Ji 

1786: 3312 1797 M ( and )[19 33 38 38  0]xS 

1787: F1S4B Ji 

1788: 3459 1797 M (w)S 

1789: F1S43 Ji 

1790: 3509 1810 M (i)S 

1791: 3528 1810 M (-)S 

1792: F3S43 Ji 

1793: 3550 1810 M (1)S 

1794: F1S4B Ji 

1795: 3584 1797 M ( = \223on\224 )[19 51 19 42 38 38 42  0]xS 

1796: F3S4B Ji 

1797: 3852 1797 M (and )[33 38 38  0]xS 

1798: F1S4B Ji 

1799: 3980 1797 M (w)S 

1800: F1S43 Ji 

1801: 4030 1810 M (i)S 

1802: 4049 1810 M (-)S 

1803: F3S43 Ji 

1804: 4071 1810 M (2)S 

1805: F1S4B Ji 

1806: 4105 1797 M ( = \223meet\224.)[19 51 19 42 54 33 33 21 42  0]xS 

1807: F3S4B Ji 

1808: 4438 1797 M (  )[19  0]xS 

1809: 4476 1797 M ( )S 

1810: 2638 1890 M (Notice an important fact: for a word )[54 38 21 21 33 33 25 33 38 25 21 57 38 38 25 21 33 38 21 25 24 33 33 21 21 25 24 38 25 25 33 25

1811: 53 38 25 38  0]xS 

1812: F1S4B Ji 

1813: 3782 1890 M (w )[50  0]xS 

1814: F3S4B Ji 

1815: 3857 1890 M (not in the same class )[38 38 21 25 21 38 24 21 38 33 24 29 33 57 33 24 33 21 33 29 29  0]xS 

1816: 2525 1999 M (as )[33 29  0]xS 

1817: F1S4B Ji 

1818: 2627 1999 M (w)S 

1819: F1S43 Ji 

1820: 2677 2012 M (i)S 

1821: F3S4B Ji 

1822: 2696 1999 M (, )[19  0]xS 

1823: F1S4B Ji 

1824: 2755 1999 M (P)S 

1825: F3S4B Ji 

1826: 2801 1999 M (\()S 

1827: F1S4B Ji 

1828: 2826 1999 M (w|)[50  0]xS 

1829: 2897 1999 M (w)S 

1830: F3S43 Ji 

1831: 2947 2012 M (1)S 

1832: F1S4B Ji 

1833: 2981 1999 M <8577>[67  0]xS 

1834: F1S43 Ji 

1835: 3098 2012 M (i)S 

1836: 3117 2012 M (-)S 

1837: F3S43 Ji 

1838: 3139 2012 M (1)S 

1839: F1S4B Ji 

1840: 3173 1999 M (, class)[19 40 33 21 38 29  0]xS 

1841: F3S4B Ji 

1842: 3382 1999 M (\()S 

1843: F1S4B Ji 

1844: 3407 1999 M (w)S 

1845: F1S43 Ji 

1846: 3457 2012 M (i)S 

1847: F3S4B Ji 

1848: 3476 1999 M (\)\))[25  0]xS 

1849: F1S4B Ji 

1850: 3526 1999 M (=0.  )[51 38 19 40  0]xS 

1851: F3S4B Ji 

1852: 3714 1999 M (This means that )[45 38 21 29 40 57 33 33 38 29 40 21 38 33 21  0]xS 

1853: 4270 1999 M (we can )[53 33 39 33 33 38  0]xS 

1854: 2525 2092 M (modify the loops of lines )[57 38 38 21 24 36 34 21 38 33 34 21 38 38 38 29 34 38 24 34 21 21 38 33 29  0]xS 

1855: 3369 2092 M (4)S 

1856: 3407 2092 M (, )[19  0]xS 

1857: 3460 2092 M (7)S 

1858: 3498 2092 M ( and )[34 33 38 38  0]xS 

1859: 3675 2092 M (8)S 

1860: 3713 2092 M ( to loop only over those )[34 21 38 34 21 38 38 38 33 38 38 21 36 33 38 37 33 25 33 21 38 38 29 33  0]xS 

1861: 2525 2200 M (words )[53 38 25 38 29  0]xS 

1862: F1S4B Ji 

1863: 2738 2200 M (w)S 

1864: F3S4B Ji 

1865: 2788 2200 M ( such that )[30 29 38 33 38 30 21 38 33 21  0]xS 

1866: F1S4B Ji 

1867: 3129 2200 M (w)S 

1868: F3S4B Ji 

1869: 3179 2200 M ( is in the same class as )[30 21 29 30 21 38 30 21 38 33 30 29 33 57 33 30 33 21 33 29 29 30 33 29  0]xS 

1870: F1S4B Ji 

1871: 3949 2200 M (w)S 

1872: F1S43 Ji 

1873: 3999 2213 M (i)S 

1874: F3S4B Ji 

1875: 4018 2200 M (.  )[19 29  0]xS 

1876: 4095 2200 M (Now, if each)[54 38 53 19 29 21 24 29 33 33 33  0]xS 

1877: 4499 2200 M ( )S 

1878: 2525 2293 M (class has 100 words, then the run time of the inner loop is )[33 21 33 29 29 39 38 33 29 39 38 38 38 39 53 38 25 38 29 19 39 21 38 33 38 39 21 38 33 39 25 38

1879: 38 39 21 21 57 33 39 38 24 38 21 38 33 38 21 38 38 33 25 38 21 38 38 38 38 21 29  0]xS 

1880: 2525 2379 M (bounded by a factor of 100.)[38 38 38 38 38 33 38 22 38 36 22 33 22 24 33 33 21 38 25 22 38 24 22 38 38 38  0]xS 

1881: 3372 2379 M ( \(If we were to explicitly perform the )[22 25 25 24 22 53 33 22 53 33 25 33 21 21 38 21 33 37 38 21 21 33 21 21 21 36 21 38 33 25 24 38

1882: 25 57 21 21 38 33  0]xS 

1883: 2525 2487 M (computations, the unigram )[33 38 57 38 38 21 33 21 21 38 38 29 19 41 21 38 33 41 38 38 21 37 25 33 57  0]xS 

1884: F5S4B Ji 

1885: 3413 2487 M (l)S 

1886: F3S4B Ji 

1887: 3455 2487 M (\222s for words )[25 29 41 24 38 25 41 53 38 25 38 29  0]xS 

1888: F1S4B Ji 

1889: 3902 2487 M (w)S 

1890: F3S4B Ji 

1891: 3952 2487 M ( not in )[41 38 38 21 41 21 38  0]xS 

1892: F1S4B Ji 

1893: 4230 2487 M (class)[33 21 38 29  0]xS 

1894: F3S4B Ji 

1895: 4380 2487 M (\()S 

1896: F1S4B Ji 

1897: 4405 2487 M (w)S 

1898: F1S43 Ji 

1899: 4455 2500 M (i)S 

1900: F3S4B Ji 

1901: 4474 2487 M (\))S 

1902: F1S4B Ji 

1903: 4499 2487 M ( )S 

1904: F3S4B Ji 

1905: 2525 2585 M (would be set to )[53 38 38 21 38 23 38 33 23 29 33 21 23 21 38  0]xS 

1906: 3018 2585 M (-)S 

1907: F5S4B Ji 

1908: 3043 2585 M <A5>S 

1909: F3S4B Ji 

1910: 3096 2585 M (, and the unnormalized probabilities would be )[19 23 33 38 38 23 21 38 33 22 38 38 38 38 25 57 33 21 21 33 33 38 22 38 25 38 38 33 38 21 21 21

1911: 21 21 33 29 22 53 38 38 21 38 22 38 33  0]xS 

1912: 2525 2672 M (0)S 

1913: 2563 2672 M (, lea)[19 19 21 33  0]xS 

1914: 2688 2672 M (ding to no contribution in lines )[38 21 38 37 19 21 38 19 38 38 19 33 38 38 21 25 21 38 38 21 21 38 38 19 21 38 19 21 21 38 33 29

1915:  0]xS 

1916: 3642 2672 M (7)S 

1917: 3680 2672 M ( and )[19 33 38 38  0]xS 

1918: 3827 2672 M (10)[38  0]xS 

1919: 3903 2672 M (\).)[25  0]xS 

1920: 3947 2672 M ( )S 

1921: 2638 2758 M (Consider a hypothetical example, with a 10,000 word )[50 38 38 29 21 38 33 25 55 33 55 38 36 38 38 21 38 33 21 21 33 33 21 55 33 37 33 57 38 21 33 19

1922: 55 53 21 21 38 55 33 55 38 38 19 38 38 38 55 53 38 25 38  0]xS 

1923: 2525 2845 M (vocabulary, 100 classes and 100 words per class.  The inner loop )

1924: [37 38 33 33 38 38 21 33 25 36 19 21 38 38 38 21 33 21 33 29 29 33 29 21 33 38 38 21 38 38 38 21

1925: 53 38 25 38 29 21 38 33 25 21 33 21 33 29 29 19 21 21 45 38 33 21 21 38 38 33 25 20 21 38 38 38

1926:  0]xS 

1927: 2525 2931 M (of the standard training algorithm wo)[38 24 73 21 38 33 73 29 21 33 38 38 33 25 38 73 21 25 33 21 38 21 38 37 72 33 21 37 38 25 21 21

1928: 38 57 72 53  0]xS 

1929: 3913 2931 M (uld require time )[38 21 38 72 25 33 38 38 21 25 33 72 21 21 57 33  0]xS 

1930: 2525 3017 M (proportional to 10,000.  Alternatively, we can use the class)[38 25 38 38 38 25 21 21 38 38 33 21 20 21 38 20 38 38 19 38 38 38 19 20 20 53 21 21 33 25 38 33

1931: 21 21 37 33 21 36 19 20 53 33 20 33 33 38 20 38 29 33 19 21 38 33 19 33 21 33 29  0]xS 

1932: 4303 3017 M (-)S 

1933: 4328 3017 M (based )[38 33 29 33 38  0]xS 

1934: 2525 3103 M (speedup.  Both the inner loop for learning the class model, and )

1935: [29 38 33 33 38 38 38 19 27 27 50 38 21 38 27 21 38 33 27 21 38 38 33 25 27 21 38 38 38 27 24 38

1936: 25 27 21 33 33 25 38 21 38 37 27 21 38 33 26 33 21 33 29 29 26 57 38 38 33 21 19 26 33 38 38  0]xS 

1937: 2525 3190 M (the inner loop for running the word)[21 38 33 49 21 38 38 33 25 48 21 38 38 38 48 24 38 25 48 25 38 38 38 21 38 37 48 21 38 33 48 53

1938: 38 25  0]xS 

1939: 3764 3190 M (-)S 

1940: 3789 3190 M (given)[37 21 37 33  0]xS 

1941: 3955 3190 M (-)S 

1942: 3980 3190 M (class model are )[33 21 33 29 29 48 57 38 38 33 21 48 33 25 33  0]xS 

1943: 2525 3276 M (bounded by a factor of 100, leading to an overall )[38 38 38 38 38 33 38 31 38 36 31 33 31 24 33 33 21 38 25 31 38 24 31 38 38 38 19 31 21 33 33 38

1944: 21 38 37 31 21 38 30 33 38 30 38 37 33 25 33 21 21  0]xS 

1945: 4128 3276 M (hypothetical )[38 36 38 38 21 38 33 21 21 33 33 21  0]xS 

1946: 2525 3362 M (improvement of 10,000 / \(100+100\) = 50.)[21 57 38 25 38 37 33 57 33 38 21 19 38 24 19 38 38 19 38 38 38 19 21 19 25 38 38 38 42 38 38 38

1947: 25 19 42 19 38 38  0]xS 

1948: 3786 3362 M ( )S 

1949: 2638 3448 M (We can extend this result to 3 or more levels, by predicting )

1950: [70 33 27 33 33 38 27 33 37 21 33 38 38 27 21 38 21 29 27 25 33 29 38 21 21 27 21 38 27 38 27 38

1951: 25 26 57 38 25 33 26 21 33 37 33 21 29 19 26 38 36 26 38 25 33 38 21 33 21 21 38 37  0]xS 

1952: 2525 3535 M (first a super)[24 21 25 29 21 65 33 65 29 38 38 33  0]xS 

1953: 2971 3535 M (-)S 

1954: 2996 3535 M (class, e.g. NOUN, and then a class, e.g. )[33 21 33 29 29 19 65 33 19 37 19 65 54 54 54 54 19 65 33 38 38 65 21 38 33 38 64 33 64 33 21 33

1955: 29 29 19 64 33 19 37 19  0]xS 

1956: 2525 3621 M (WEEKDAY, and finally the word, Tuesday.  Such a decom)[70 46 46 54 54 53 54 19 37 33 38 38 37 24 21 38 33 21 21 36 37 21 38 33 37 53 38 25 38 19 37 45

1957: 38 33 29 38 33 36 19 37 37 42 38 33 38 36 33 36 38 33 33 38  0]xS 

1958: 4474 3621 M (-)S 

1959: 2525 3707 M (position will further reduce the )[38 38 29 21 21 21 38 38 33 53 21 21 21 33 24 38 25 21 38 33 25 33 25 33 38 38 33 33 32 21 38 33

1960:  0]xS 

1961: 3544 3707 M (maximum number of indicator )[57 33 37 21 57 38 57 32 38 38 57 38 33 25 32 38 24 32 21 38 38 21 33 33 21 38 25  0]xS 

1962: 2525 3793 M (functions, but, since there is some overhead to each level, we )

1963: [24 38 38 33 21 21 38 38 29 19 33 38 38 21 19 33 29 21 38 33 33 33 21 38 33 25 33 33 21 29 33 29

1964: 38 57 33 33 38 37 33 25 38 33 33 38 33 21 38 32 33 33 33 38 32 21 33 37 33 21 19 32 53 33  0]xS 

1965: 2525 3880 M (have not found improvements by extending beyond 3 levels.)[38 33 37 33 19 38 38 21 19 24 38 38 38 38 19 21 57 38 25 38 37 33 57 33 38 21 29 19 38 36 19 33

1966: 37 21 33 38 38 21 38 37 19 38 33 36 38 38 38 19 38 19 21 33 37 33 21 29  0]xS 

1967: 4340 3880 M ( )S 

1968: F0S64 Ji 

1969: 2873 4042 M (3.)[50  0]xS 

1970: 2948 4042 M ( )S 

1971: 3008 4042 M (PREVIOUS RESEARCH)[61 72 61 78 39 83 78 61 25 72 61 61 61 78 72 72  0]xS 

1972: 4126 4042 M ( )S 

1973: F3S4B Ji 

1974: 2525 4186 M (Maximum entropy has been well studied.  )[67 33 37 21 57 38 57 27 33 38 21 25 38 38 36 27 38 33 29 27 38 33 33 38 27 53 33 21 21 26 29 21

1975: 38 38 21 33 38 19 26  0]xS 

1976: 3857 4186 M ([1])[26 38  0]xS 

1977: 3947 4186 M (  gives the class)[26 26 37 21 37 33 29 26 21 38 33 26 33 21 33 29  0]xS 

1978: 4445 4186 M (ic )[21 33  0]xS 

1979: 2525 4273 M (Generalized Iterative Scaling algorithm, although in a form )[53 33 38 33 25 33 21 21 33 33 38 47 25 21 33 25 33 21 21 37 33 47 42 33 33 21 21 38 37 47 33 21

1980: 37 38 25 21 21 38 57 19 47 33 21 21 38 38 38 37 38 47 21 38 47 33 47 24 38 25 57  0]xS 

1981: 2525 4359 M (suitable for joint probabilities, as opposed to the conditional )

1982: [29 38 21 21 33 38 21 33 40 24 38 25 39 21 38 21 38 21 39 38 25 38 38 33 38 21 21 21 21 21 33 29

1983: 19 39 33 29 39 38 38 38 38 29 33 38 39 21 38 39 21 38 33 39 33 38 38 38 21 21 21 38 38 33 21  0]xS 

1984: 2525 4445 M (probabilities given here, and is somewhat dense; )[38 25 38 38 33 38 21 21 21 21 21 33 29 26 37 21 37 33 38 26 38 33 25 33 19 26 33 38 38 26 21 29

1985: 26 29 38 57 33 53 38 33 21 26 38 33 38 29 33 21  0]xS 

1986: 4051 4445 M ([2])[26 38  0]xS 

1987: 4141 4445 M ( is a classic )[26 21 29 25 33 25 33 21 33 29 29 21 33  0]xS 

1988: 2525 4531 M (introduction to the use of maximum)[21 38 21 25 38 38 38 33 21 21 38 38 21 21 38 21 21 38 33 21 38 29 33 21 38 24 21 57 33 37 21 57

1989: 38  0]xS 

1990: 3613 4531 M ( entropy models for language )[21 33 38 21 25 38 38 36 21 57 38 38 33 21 29 21 24 38 25 21 21 33 38 37 38 33 37 33  0]xS 

1991: 2525 4618 M (modeling, but despite the fact that )[57 38 38 33 21 21 38 37 19 61 38 38 21 61 38 33 29 38 21 21 33 61 21 38 33 61 24 33 33 21 61 21

1992: 38 33 21  0]xS 

1993: 3819 4618 M ([2])[26 38  0]xS 

1994: 3909 4618 M ( uses conditional )[61 38 29 33 29 60 33 38 38 38 21 21 21 38 38 33 21  0]xS 

1995: 2525 4704 M (probabilities, most of the discussion is of joint probabilities.)

1996: [38 25 38 38 33 38 21 21 21 21 21 33 29 19 19 57 38 29 21 19 38 24 19 21 38 33 19 38 21 29 33 38

1997: 29 29 21 38 38 19 21 29 19 38 24 19 21 38 21 38 21 19 38 25 38 38 33 38 21 21 21 21 21 33 29  0]xS 

1998: 4333 4704 M ( )S 

1999: 2638 4790 M ([2])[26 38  0]xS 

2000: 2728 4790 M ( has previously used a simple form of classe)[49 38 33 29 49 38 25 33 37 21 38 38 29 21 36 49 38 29 33 38 49 33 49 29 21 57 38 21 33 49 24 38

2001: 25 57 49 38 24 48 33 21 33 29 29  0]xS 

2002: 4289 4790 M (s with )[29 48 53 21 21 38  0]xS 

2003: 2525 4876 M (maximum entropy)[57 33 37 21 57 38 57 26 33 38 21 25 38 38  0]xS 

2004: 3080 4876 M (-)S 

2005: 3105 4876 M (based language models.  However, they were )[38 33 29 33 38 26 21 33 38 37 38 33 37 33 26 57 38 38 33 21 29 19 26 26 54 38 53 33 37 33 25 19

2006: 25 21 38 33 36 25 53 33 25 33  0]xS 

2007: 2525 4963 M (used only as )[38 29 33 38 25 38 38 21 36 25 33 29  0]xS 

2008: F1S4B Ji 

2009: 2933 4963 M (conditioning)[33 38 38 38 21 21 21 38 38 21 38  0]xS 

2010: F3S4B Ji 

2011: 3316 4963 M ( variables; i.e., indicator functions like )[25 37 33 25 21 33 38 21 33 29 21 24 21 19 33 19 19 24 21 38 38 21 33 33 21 38 25 24 24 38 38 33

2012: 21 21 38 38 29 24 21 21 37 33  0]xS 

2013: F1S4B Ji 

2014: 2525 5071 M (f)S 

2015: F1S43 Ji 

2016: 2546 5084 M (j)S 

2017: F3S4B Ji 

2018: 2565 5071 M (\()S 

2019: F1S4B Ji 

2020: 2590 5071 M (w|w)[50 21  0]xS 

2021: F1S43 Ji 

2022: 2711 5084 M (i)S 

2023: 2730 5084 M (-)S 

2024: F3S43 Ji 

2025: 2752 5084 M (1)S 

2026: F3S4B Ji 

2027: 2786 5071 M (\))S 

2028: F1S4B Ji 

2029: 2811 5071 M (= )[51  0]xS 

2030: F3S4B Ji 

2031: 2881 5071 M (1)S 

2032: F1S4B Ji 

2033: 2919 5071 M ( if w=y and class)[19 21 21 19 50 51 33 19 38 38 38 19 33 21 38 29  0]xS 

2034: F3S4B Ji 

2035: 3435 5071 M (\()S 

2036: F1S4B Ji 

2037: 3460 5071 M (w)S 

2038: F1S43 Ji 

2039: 3510 5084 M (i)S 

2040: 3529 5084 M (-)S 

2041: F3S43 Ji 

2042: 3551 5084 M (1)S 

2043: F3S4B Ji 

2044: 3585 5071 M (\))S 

2045: F1S4B Ji 

2046: 3610 5071 M (=x)[51  0]xS 

2047: F3S4B Ji 

2048: 3694 5071 M ( were used.  They were not )[19 53 33 25 33 19 38 29 33 38 19 19 18 45 38 33 36 18 53 33 25 33 18 38 38 21  0]xS 

2049: 2525 5164 M (used for )[38 29 33 38 19 24 38 25  0]xS 

2050: F1S4B Ji 

2051: 2788 5164 M (predicting)[38 29 33 38 21 33 21 21 38  0]xS 

2052: F3S4B Ji 

2053: 3098 5164 M ( outputs, and thus did not lead to spe)[19 38 38 21 38 38 21 29 19 19 33 38 38 19 21 38 38 29 19 38 21 38 19 38 38 21 19 21 33 33 38 19

2054: 21 38 19 29 38  0]xS 

2055: 4205 5164 M (edups.  )[33 38 38 38 29 19 19  0]xS 

2056: 4438 5164 M ( )S 

2057: 2638 5250 M (Word classes have, of course, been used extensively in )[70 38 25 38 46 33 21 33 29 29 33 29 46 38 33 37 33 19 46 38 24 46 33 38 38 25 29 33 19 46 38 33

2058: 33 38 46 38 29 33 38 46 33 37 21 33 38 29 21 37 33 21 36 46 21 38  0]xS 

2059: 2525 5336 M (language modeling, including )[21 33 38 37 38 33 37 33 32 57 38 38 33 21 21 38 37 19 31 21 38 33 21 38 38 21 38 37  0]xS 

2060: 3476 5336 M ([2])[26 38  0]xS 

2061: 3566 5336 M ([3])[26 38  0]xS 

2062: 3656 5336 M ([4])[26 38  0]xS 

2063: 3746 5336 M ([5])[26 38  0]xS 

2064: 3836 5336 M (.  However, previous )[19 31 31 54 38 53 33 37 33 25 19 31 38 25 33 37 21 38 38 29  0]xS 

2065: 2525 5423 M (research has f)[25 33 29 33 33 25 33 38 22 38 33 29 22  0]xS 

2066: 2942 5423 M (ocused mostly on improving perplexity or reducing )[38 33 38 29 33 38 22 57 38 29 21 21 36 22 38 38 22 21 57 38 25 38 37 21 38 37 21 38 33 25 38 21

2067: 33 37 21 21 36 21 38 25 21 25 33 38 38 33 21 38 37  0]xS 

2068: 2525 5509 M (language model size, and never to our knowledge for increasing )

2069: [21 33 38 37 38 33 37 33 25 57 38 38 33 21 25 29 21 33 33 19 25 33 38 38 25 38 33 37 33 25 25 21

2070: 38 25 38 38 25 25 37 38 38 53 21 33 38 37 33 25 24 38 25 24 21 38 33 25 33 33 29 21 38 37  0]xS 

2071: 2525 5595 M (speed.  Note that we have previously used a model form very )[29 38 33 33 38 19 32 32 54 38 21 33 32 21 38 33 21 32 53 33 32 38 33 37 33 32 38 25 33 37 21 38

2072: 38 29 21 36 32 38 29 33 38 32 33 31 57 38 38 33 21 31 24 38 25 57 31 37 33 25 36  0]xS 

2073: 

2074: %%IncludeResource: font Courier

2075: F /F7 0 /256 T /Courier mF 

2076: /F7S4B F7 [75 0 0 -75 0 0 ] mFS

2077: F7S4B Ji 

2078: 350 567 M (1)S 

2079: 395 567 M ( )S 

2080: 500 567 M (For each iteration)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2081: 1310 567 M ( )S 

2082: 350 665 M (2)S 

2083: 395 665 M ( )S 

2084: 500 665 M (  observed[1..# of indicators] )[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2085: F5S4B Ji 

2086: 1895 665 M <AC>S 

2087: F7S4B Ji 

2088: 1969 665 M ( {0})[45 45 45  0]xS 

2089: 2149 665 M ( )S 

2090: 350 750 M (3)S 

2091: 395 750 M ( )S 

2092: 500 750 M (  For i = 1 to |training data|)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2093: 1850 750 M ( )S 

2094: 350 835 M (4)S 

2095: 395 835 M ( )S 

2096: 500 835 M (    For each word w in Vocabulary)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45

2097:  0]xS 

2098: 1985 835 M ( )S 

2099: 350 919 M (5)S 

2100: 395 919 M ( )S 

2101: 500 919 M (      )[45 45 45 45 45  0]xS 

2102: 770 944 M (unnormalized[w])[45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2103: F5S4B Ji 

2104: 1445 944 M <AC>S 

2105: /F3S40 F3 [64.07 0 0 -64.07 0 0 ] mFS

2106: F3S40 Ji 

2107: 2192 950 M (\))S 

2108: 2170 950 M (\))S 

2109: 2008 950 M (...)[16 16  0]xS 

2110: 1920 950 M (,)S 

2111: 1855 950 M (\()S 

2112: 1530 950 M (exp\()[28 32 32  0]xS 

2113: /F3S33 F3 [51.258 0 0 -51.258 0 0 ] mFS

2114: F3S33 Ji 

2115: 2142 966 M (1)S 

2116: 1983 966 M (1)S 

2117: : 1645 879 68 106 rc /F6S60 F6 [96.109 0 0 -96.109 0 0 ] mFS

2118: F6S60 Ji 

2119: 1645 964 M <07>S 

2120: ; /F5S33 F5 [51.258 0 0 -51.258 0 0 ] mFS

2121: F5S33 Ji 

2122: 2115 966 M (-)S 

2123: /F1S33 F1 [51.258 0 0 -51.258 0 0 ] mFS

2124: F1S33 Ji 

2125: 1676 1024 M (j)S 

2126: 2095 966 M (i)S 

2127: 1831 966 M (j)S 

2128: 1767 966 M (j)S 

2129: /F1S40 F1 [64.07 0 0 -64.07 0 0 ] mFS

2130: F1S40 Ji 

2131: 2055 950 M (w)S 

2132: 1947 950 M (w)S 

2133: 1879 950 M (w)S 

2134: 1799 950 M (f)S 

2135: /F5S40I15Y00 F5 [64.07 0 21.914 -64.07 0 0 ] mFS

2136: F5S40I15Y00 Ji 

2137: 1722 950 M (l)S 

2138: F7S4B Ji 

2139: 2230 919 M ( )S 

2140: 350 1116 M (6)S 

2141: 395 1116 M ( )S 

2142: 500 1116 M (    Next w)[45 45 45 45 45 45 45 45 45  0]xS 

2143: 950 1116 M ( )S 

2144: 350 1229 M (7)S 

2145: 395 1229 M ( )S 

2146: 500 1229 M (    z )[45 45 45 45 45  0]xS 

2147: F5S4B Ji 

2148: 770 1229 M <AC>S 

2149: : 859 1138 83 135 rc /F6S74 F6 [116.477 0 0 -116.477 0 0 ] mFS

2150: F6S74 Ji 

2151: 859 1248 M <07>S 

2152: ; /F1S42 F1 [66.559 0 0 -66.559 0 0 ] mFS

2153: F1S42 Ji 

2154: 879 1326 M (w)S 

2155: F7S4B Ji 

2156: 1002 1229 M (unnormalized[w])[45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2157: 1677 1229 M ( )S 

2158: 350 1408 M (8)S 

2159: 395 1408 M ( )S 

2160: 500 1408 M (    For each word w in Vocabul)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2161: 1850 1408 M (ary)[45 45  0]xS 

2162: 1985 1408 M ( )S 

2163: 350 1526 M (9)S 

2164: 395 1526 M ( )S 

2165: 500 1526 M (      For)[45 45 45 45 45 45 45 45  0]xS 

2166: /F7S21 F7 [33 0 0 -33 0 0 ] mFS

2167: F7S21 Ji 

2168: 905 1526 M ( )S 

2169: F7S4B Ji 

2170: 925 1526 M (each j such that f)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2171: /F7S43 F7 [67 0 0 -67 0 0 ] mFS

2172: F7S43 Ji 

2173: 1735 1539 M (j)S 

2174: F7S4B Ji 

2175: 1775 1526 M (\(w,)[45 45  0]xS 

2176: F1S4B Ji 

2177: 1910 1526 M ( w)[19  0]xS 

2178: F1S43 Ji 

2179: 1979 1539 M (1)S 

2180: F1S4B Ji 

2181: 2013 1526 M <8577>[67  0]xS 

2182: F1S43 Ji 

2183: 2130 1539 M (i)S 

2184: 2149 1539 M (-)S 

2185: 2171 1539 M (1)S 

2186: F7S4B Ji 

2187: 2205 1526 M (\))S 

2188: F5S4B Ji 

2189: 2250 1526 M ( )S 

2190: 2269 1526 M <B9>S 

2191: F7S4B Ji 

2192: 2310 1526 M (0)S 

2193: : 2355 1467 30 80 rc 2355 1526 M ( )S 

2194: ; 350 1640 M (10)[45  0]xS 

2195: 440 1640 M ( )S 

2196: 500 1640 M (        observed[j]+=f)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2197: F7S43 Ji 

2198: 1490 1653 M (j)S 

2199: F5S4B Ji 

2200: 1530 1640 M ( )S 

2201: 1549 1640 M <B4>S 

2202: F7S4B Ji 

2203: 1590 1640 M (unnormalized[w]/z)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2204: : 2355 1581 30 80 rc 2355 1640 M ( )S 

2205: ; 350 1730 M (11)[45  0]xS 

2206: 440 1730 M ( )S 

2207: 500 1730 M (      Next j)[45 45 45 45 45 45 45 45 45 45 45  0]xS 

2208: 1040 1730 M ( )S 

2209: 350 1815 M (12)[45  0]xS 

2210: 440 1815 M ( )S 

2211: 500 1815 M (    Next w)[45 45 45 45 45 45 45 45 45  0]xS 

2212: 950 1815 M ( )S 

2213: 350 1900 M (13)[45  0]xS 

2214: 440 1900 M ( )S 

2215: 500 1900 M (  Next i)[45 45 45 45 45 45 45  0]xS 

2216: 860 1900 M ( )S 

2217: 350 2008 M (14)[45  0]xS 

2218: 440 2008 M ( )S 

2219: 500 2008 M (  For each indicator f)[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2220: F7S43 Ji 

2221: 1490 2021 M (j)S 

2222: F7S4B Ji 

2223: 1530 2008 M ( )S 

2224: 350 2121 M (15)[45  0]xS 

2225: 440 2121 M ( )S 

2226: 500 2121 M (    re)[45 45 45 45 45  0]xS 

2227: 770 2121 M (-)S 

2228: 815 2121 M (estimate )[45 45 45 45 45 45 45 45  0]xS 

2229: F5S4B Ji 

2230: 1220 2121 M (l)S 

2231: F7S43 Ji 

2232: 1262 2134 M (j)S 

2233: F7S4B Ji 

2234: 1302 2121 M ( using observed[j] )[45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2235: 2157 2121 M ( )S 

2236: 350 2211 M (16)[45  0]xS 

2237: 440 2211 M ( )S 

2238: 500 2211 M (  Next j)[45 45 45 45 45 45 45  0]xS 

2239: 860 2211 M ( )S 

2240: F3S4B Ji 

2241: 350 2304 M (17)[38  0]xS 

2242: 425 2304 M ( )S 

2243: F7S4B Ji 

2244: 500 2304 M (Next iteration)[45 45 45 45 45 45 45 45 45 45 45 45 45  0]xS 

2245: F3S4B Ji 

2246: 1130 2304 M ( )S 

2247: : N 303 500 4 4 rp C 

2248:  L ; : N 303 500 4 4 rp C 

2249:  L ; : N 307 500 2078 4 rp C 

2250:  L ; : N 2385 500 4 4 rp C 

2251:  L ; : N 2385 500 4 4 rp C 

2252:  L ; : N 303 504 4 1822 rp C 

2253:  L ; : N 303 2326 4 4 rp C 

2254:  L ; : N 303 2326 4 4 rp C 

2255:  L ; : N 307 2326 2078 4 rp C 

2256:  L ; : N 2385 504 4 1822 rp C 

2257:  L ; : N 2385 2326 4 4 rp C 

2258:  L ; : N 2385 2326 4 4 rp C 

2259:  L ; LH

2260: (%%[Page: 2]%%) = 

2261: %%PageTrailer

2262: 

2263: %%Page: 3 3

2264: %%PageBoundingBox: 12 12 600 780

2265: %%EndPageComments

2266: %%BeginPageSetup

2267: /DeviceRGB dup setcolorspace /colspABC exch def

2268: mysetup concat colspRefresh

2269: %%EndPageSetup

2270: 

2271: /DeviceGray dup setcolorspace /colspABC exch def

2272: 0 0 scol F3S53 Ji 

2273: 350 6179 M ( )S 

2274: 2150 6179 M ( )S 

2275: F3S4B Ji 

2276: 350 568 M (similar to the one used here for reducing language model size, by )

2277: [29 21 57 21 21 33 25 21 21 38 21 21 38 33 21 38 38 33 21 38 29 33 38 20 38 33 25 33 20 24 38 25

2278: 20 25 33 38 38 33 21 38 37 20 21 33 38 37 38 33 37 33 20 57 38 38 33 21 20 29 21 33 33 19 20 38

2279: 36  0]xS 

2280: 350 654 M (up to a factor of)[38 38 19 21 38 19 33 19 24 33 33 21 38 25 19 38  0]xS 

2281: 830 654 M ( 3, at the same perplexity )[19 38 19 19 33 21 19 21 38 33 19 29 33 57 33 19 38 33 25 38 21 33 37 21 21 36  0]xS 

2282: 1602 654 M ([5])[26 38  0]xS 

2283: 1692 654 M (.)S 

2284: 1711 654 M ( )S 

2285: 463 740 M (There have been three noteworthy previous attempts to speed )[45 38 33 25 33 22 38 33 37 33 22 38 33 33 38 22 21 38 25 33 33 21 38 38 21 33 53 38 25 21 38 36

2286: 21 38 25 33 37 21 38 38 29 21 33 21 21 33 57 38 21 29 21 21 38 21 29 38 33 33 38  0]xS 

2287: 350 827 M (up maximum entropy models: unigram caching, Improved )[38 38 57 57 33 37 21 57 38 57 57 33 38 21 25 38 38 36 57 57 38 38 33 21 29 21 57 38 38 21 37 25

2288: 33 57 56 33 33 33 38 21 38 37 19 56 25 57 38 25 38 37 33 38  0]xS 

2289: 350 913 M (Iterative Scaling \(IIS\) )[25 21 33 25 33 21 21 37 33 19 42 33 33 21 21 38 37 19 25 25 25 42 25  0]xS 

2290: 1023 913 M ([6])[26 38  0]xS 

2291: 1113 913 M (, and cluster expansion )[19 19 33 38 38 19 33 21 38 29 21 33 25 19 33 37 38 33 38 29 21 38 38  0]xS 

2292: 1822 913 M ([7])[26 38  0]xS 

2293: 1912 913 M ([8])[26 38  0]xS 

2294: 2002 913 M (.  )[19 19  0]xS 

2295: 2059 913 M ( )S 

2296: 463 999 M (Unigram caching makes use of the following observation: )[54 38 21 37 25 33 57 37 33 33 33 38 21 38 37 37 57 33 37 33 29 37 38 29 33 37 38 24 37 21 38 33

2297: 36 24 38 21 21 38 53 21 38 37 36 38 38 29 33 25 37 33 21 21 38 38 21  0]xS 

2298: 350 1085 M (most bigram and trigram indicators are not used in practice \(e.g. )

2299: [57 38 29 21 24 38 21 37 25 33 57 24 33 38 38 24 21 25 21 37 25 33 57 24 21 38 38 21 33 33 21 38

2300: 25 29 24 33 25 33 24 38 38 21 24 38 29 33 38 24 21 38 23 38 25 33 33 21 21 33 33 23 25 33 19 37

2301: 19  0]xS 

2302: 350 1172 M (if the string \223York Francisco\224 never occurred in the training data, )

2303: [21 24 20 21 38 33 20 29 21 25 21 38 37 20 33 54 38 25 37 20 42 25 33 38 33 21 29 33 38 33 20 38

2304: 33 37 33 25 20 38 33 33 38 25 25 33 38 19 21 38 19 21 38 33 19 21 25 33 21 38 21 38 37 19 38 33

2305: 21 33 19  0]xS 

2306: 350 1258 M (then there will not be any bigram indicators for that case\).  On )

2307: [21 38 33 38 28 21 38 33 25 33 28 53 21 21 21 28 38 38 21 28 38 33 28 33 38 36 28 38 21 37 25 33

2308: 57 28 21 38 38 21 33 33 21 38 25 29 27 24 38 25 27 21 38 33 21 27 33 33 29 33 25 19 27 27 54 38

2309:  0]xS 

2310: 350 1344 M (the other hand, all possible unigram indicators typically )[21 38 33 20 38 21 38 33 25 20 38 33 38 38 19 20 33 21 21 20 38 38 29 29 21 38 21 33 20 38 38 21

2311: 37 25 33 57 19 21 38 38 21 33 33 21 38 25 29 19 21 36 38 21 33 33 21 21 36  0]xS 

2312: F1S4B Ji 

2313: 2048 1344 M (are)[38 29  0]xS 

2314: F3S4B Ji 

2315: 2148 1344 M ( used.  )[19 38 29 33 38 19 19  0]xS 

2316: 350 1430 M (This means that typically, the vast majority of indicator functions )

2317: [45 38 21 29 21 57 33 33 38 29 21 21 38 33 21 21 21 36 38 21 33 33 21 21 36 19 21 21 38 33 21 37

2318: 33 29 21 20 57 33 21 38 25 21 21 36 20 38 24 20 21 38 38 21 33 33 21 38 25 20 24 38 38 33 21 21

2319: 38 38 29  0]xS 

2320: 350 1517 M (that are non)[21 38 33 21 22 33 25 33 22 38 38  0]xS 

2321: 712 1517 M (-)S 

2322: 737 1517 M (zero for a given context are unigram indi)[33 33 25 38 22 24 38 25 22 33 22 37 21 37 33 38 22 33 38 38 21 33 37 21 21 33 25 33 21 38 38 21

2323: 37 25 33 57 21 21 38 38  0]xS 

2324: 1982 1517 M (cators; also )[33 33 21 38 25 29 21 21 33 21 29 38  0]xS 

2325: 350 1603 M (notice that these unigram indicators are independent of context, )

2326: [38 38 21 21 33 33 27 21 38 33 21 27 21 38 33 29 33 27 38 38 21 37 25 33 57 27 21 38 38 21 33 33

2327: 21 38 25 29 27 33 25 33 27 21 38 38 33 38 33 38 38 33 38 21 27 38 24 26 33 38 38 21 33 37 21 19

2328:  0]xS 

2329: 350 1689 M (meaning computation can be easily shared.  In unigram caching, )

2330: [57 33 33 38 21 38 37 24 33 38 57 38 38 21 33 21 21 38 38 24 33 33 38 24 38 33 24 33 33 29 21 21

2331: 36 24 29 38 33 25 33 38 19 24 24 25 38 23 38 38 21 37 25 33 57 23 33 33 33 38 21 38 37 19  0]xS 

2332: 350 1775 M (the effect of the unigram indicators is pre)[21 38 33 39 33 24 24 33 33 21 39 38 24 39 21 38 33 39 38 38 21 37 25 33 57 38 21 38 38 21 33 33

2333: 21 38 25 29 38 21 29 38 38 25  0]xS 

2334: 1726 1775 M (-)S 

2335: 1751 1775 M (computed and the )[33 38 57 38 38 21 33 38 38 33 38 38 38 21 38 33  0]xS 

2336: 350 1862 M (computations of the inner loop are rearranged so that )[33 38 57 38 38 21 33 21 21 38 38 29 19 38 24 19 21 38 33 19 21 38 38 33 25 19 21 38 38 38 19 33

2337: 25 33 19 25 33 33 25 25 33 38 37 33 38 19 29 38 19 21 38 33 21  0]xS 

2338: 1960 1862 M (they depend )[21 38 33 36 18 38 33 38 33 38 38  0]xS 

2339: 350 1948 M (only on those non)[38 38 21 36 23 38 38 23 21 38 38 29 33 23 38 38  0]xS 

2340: 901 1948 M (-)S 

2341: 926 1948 M (unigram indicators that take a non)[38 38 21 37 25 33 57 23 21 38 38 21 33 33 21 38 25 29 22 21 38 33 21 22 21 33 37 33 22 33 22 38

2342: 38  0]xS 

2343: 1967 1948 M (-)S 

2344: 1992 1948 M (zero value.   )[33 33 25 38 22 37 33 21 38 33 19 19 19  0]xS 

2345: 350 2034 M (In practice, the number of non)[25 38 31 38 25 33 33 21 21 33 33 19 31 21 38 33 31 38 38 57 38 33 25 31 38 24 31 38 38  0]xS 

2346: 1321 2034 M (-)S 

2347: 1346 2034 M (zero indicators still tends to be )[33 33 25 38 30 21 38 38 21 33 33 21 38 25 29 30 29 21 21 21 21 30 21 33 38 38 29 30 21 38 30 38

2348: 33  0]xS 

2349: 350 2120 M (proportional to the vocabulary size \(since the number of non)[38 25 38 38 38 25 21 21 38 38 33 21 19 21 38 19 21 38 33 19 37 38 33 33 38 38 21 33 25 36 19 29

2350: 21 33 33 19 25 29 21 38 33 33 19 21 38 33 19 38 38 57 38 33 25 19 38 24 19 38 38  0]xS 

2351: 2170 2120 M (-)S 

2352: 2195 2120 M (zero )[33 33 25 38  0]xS 

2353: 350 2207 M (bigrams, trigrams, and similar indicator)[38 21 37 25 33 57 29 19 39 21 25 21 37 25 33 57 29 19 39 33 38 38 39 29 21 57 21 21 33 25 39 21

2354: 38 38 21 33 33 21 38  0]xS 

2355: 1616 2207 M (-)S 

2356: 1641 2207 M (funct)[24 38 38 33  0]xS 

2357: 1795 2207 M (ions for a given )[21 38 38 29 39 24 38 25 39 33 39 37 21 37 33 38  0]xS 

2358: 350 2293 M (history is bounded by the vocabulary size\).)[38 21 29 21 38 25 36 19 21 29 19 38 38 38 38 38 33 38 19 38 36 19 21 38 33 19 37 38 33 33 38 38

2359: 21 33 25 36 19 29 21 33 33 25  0]xS 

2360: 1641 2293 M ( )S 

2361: 463 2379 M ( We have implemented unigram caching and it leads to )[19 70 33 46 38 33 37 33 46 21 57 38 21 33 57 33 38 21 33 38 45 38 38 21 37 25 33 57 45 33 33 33

2362: 38 21 38 37 45 33 38 38 45 21 21 45 21 33 33 38 29 45 21 38  0]xS 

2363: 350 2465 M (considerable speedups over the na\357ve implementation.  Our 35 )

2364: [33 38 38 29 21 38 33 25 33 38 21 33 32 29 38 33 33 38 38 38 29 31 38 37 33 25 31 21 38 33 31 38

2365: 33 21 37 33 31 21 57 38 21 33 57 33 38 21 33 21 21 38 38 19 31 31 54 38 25 31 38 38  0]xS 

2366: 350 2552 M (times speedup is a speedup over unigram caching.  Our technique )

2367: [21 21 57 33 29 20 29 38 33 33 38 38 38 20 21 29 19 33 19 29 38 33 33 38 38 38 19 38 37 33 25 19

2368: 38 38 21 37 25 33 57 19 33 33 33 38 21 38 37 19 19 19 54 38 25 19 21 33 33 38 38 21 38 38 33  0]xS 

2369: 350 2638 M (can be used with )[33 33 38 34 38 33 34 38 29 33 38 34 53 21 21 38  0]xS 

2370: 932 2638 M (or without unigram caching, but because of )[38 25 34 53 21 21 38 38 38 21 34 38 38 21 37 25 33 57 34 33 33 33 38 21 38 37 19 34 38 38 21 33

2371: 38 33 33 33 38 29 33 33 38 24  0]xS 

2372: 350 2724 M (some extra overhead involved in unigram caching, and because )

2373: [29 38 57 33 28 33 37 21 25 33 28 38 37 33 25 38 33 33 38 28 21 38 37 38 21 37 33 38 28 21 38 28

2374: 38 38 21 37 25 33 57 28 33 33 33 38 21 38 37 19 28 33 38 38 28 38 33 33 33 38 29 33  0]xS 

2375: 350 2810 M (our technique drastically reduces the number of unigrams, it is )

2376: [38 38 25 30 21 33 33 38 38 21 38 38 33 30 38 25 33 29 21 21 33 33 21 21 36 30 25 33 38 38 33 33

2377: 29 30 21 38 33 30 38 38 57 38 33 25 30 38 24 30 38 38 21 37 25 33 57 29 19 29 21 21 29 21 29  0]xS 

2378: 350 2897 M (usually best to use our technique without unigram caching.)[38 29 38 33 21 21 36 19 38 33 29 21 19 21 38 19 38 29 33 19 38 38 25 19 21 33 33 38 38 21 38 38

2379: 33 19 53 21 21 38 38 38 21 19 38 38 21 37 25 33 57 19 33 33 33 38 21 38 37  0]xS 

2380: 2123 2897 M ( )S 

2381: 463 2983 M (In Improved Iterative Scaling)[25 38 62 25 57 38 25 38 37 33 38 62 25 21 33 25 33 21 21 37 33 62 42 33 33 21 21 38  0]xS 

2382: 1477 2983 M ( )S 

2383: 1539 2983 M ([6])[26 38  0]xS 

2384: 1629 2983 M (, a different update )[19 62 33 62 38 21 24 24 33 25 33 38 21 61 38 38 38 33 21 33  0]xS 

2385: 350 3069 M (technique is used. It introduces additional overhead that slows )

2386: [21 33 33 38 38 21 38 38 33 32 21 29 32 38 29 33 38 19 32 25 21 32 21 38 21 25 38 38 38 33 33 29

2387: 32 33 38 38 21 21 21 38 38 33 21 32 38 37 33 25 38 33 33 38 31 21 38 33 21 31 29 21 38 53 29  0]xS 

2388: 350 3155 M (down the time for each iteration of the iterative scaling )[38 38 53 38 56 21 38 33 56 21 21 57 33 55 24 38 25 55 33 33 33 38 55 21 21 33 25 33 21 21 38 38

2389: 55 38 24 55 21 38 33 55 21 21 33 25 33 21 21 37 33 55 29 33 33 21 21 38 37  0]xS 

2390: 350 3242 M (algorithm, but allows larger steps to be taken at each time, )

2391: [33 21 37 38 25 21 21 38 57 19 42 38 38 21 42 33 21 21 38 53 29 42 21 33 25 37 33 25 42 29 21 33

2392: 38 29 41 21 38 41 38 33 41 21 33 37 33 38 41 33 21 41 33 33 33 38 41 21 21 57 33 19  0]xS 

2393: 350 3328 M (leading to fewer it)[21 33 33 38 21 38 37 34 21 38 34 24 33 53 33 25 34 21  0]xS 

2394: 942 3328 M (erations, and overall faster performance.  It )[33 25 33 21 21 38 38 29 19 34 33 38 38 34 38 37 33 25 33 21 21 34 24 33 29 21 33 25 33 38 33 25

2395: 24 38 25 57 33 38 33 33 19 33 33 25 21  0]xS 

2396: 350 3414 M (also introduces additional memory overhead and coding )[33 21 29 38 67 21 38 21 25 38 38 38 33 33 29 67 33 38 38 21 21 21 38 38 33 21 67 57 33 57 38 25

2397: 36 67 38 37 33 25 38 33 33 38 67 33 38 38 67 33 38 38 21 38 37  0]xS 

2398: 350 3500 M (complexity.  The main benefits from improved iterative scaling )

2399: [33 38 57 38 21 33 37 21 21 36 19 29 29 45 38 33 29 57 33 21 38 29 38 33 38 33 24 21 21 29 29 24

2400: 25 38 57 29 21 57 38 25 38 37 33 38 28 21 21 33 25 33 21 21 37 33 28 29 33 33 21 21 38 37  0]xS 

2401: 350 3586 M (come from certain models in which the total number of indicator )

2402: [33 38 57 33 23 24 25 38 57 23 33 33 25 21 33 21 38 22 57 38 38 33 21 29 22 21 38 22 53 38 21 33

2403: 38 22 21 38 33 22 21 38 21 33 21 22 38 38 57 38 33 25 22 38 24 22 21 38 38 21 33 33 21 38 25  0]xS 

2404: 350 3673 M (functions that can be true fo)[24 38 38 33 21 21 38 38 29 31 21 38 33 21 31 33 33 38 30 38 33 30 21 25 38 33 30 24  0]xS 

2405: 1249 3673 M (r a certain time is highly variable.  )[25 30 33 30 33 33 25 21 33 21 38 30 21 21 57 33 30 21 29 30 38 21 37 38 21 36 30 37 33 25 21 33

2406: 38 21 33 19 19  0]xS 

2407: 350 3759 M (The learning speed of Generalized Iterative Scaling )[45 38 33 28 21 33 33 25 38 21 38 37 28 29 38 33 33 38 28 38 24 28 53 33 38 33 25 33 21 21 33 33

2408: 38 28 25 21 33 25 33 21 21 37 33 27 42 33 33 21 21 38 37  0]xS 

2409: 1974 3759 M (is inversely )[21 29 27 21 38 37 33 25 29 33 21 36  0]xS 

2410: 350 3869 M (proportional to the )[38 25 38 38 38 25 21 21 38 38 33 21 42 21 38 41 21 38 33  0]xS 

2411: 999 3869 M (value of )[37 33 21 38 33 41 38 24  0]xS 

2412: : 1639 3777 89 138 rc /F6S7D F6 [125 0 0 -125 0 0 ] mFS

2413: F6S7D Ji 

2414: 1639 3888 M <07>S 

2415: ; /F5S42 F5 [66.664 0 0 -66.664 0 0 ] mFS

2416: F5S42 Ji 

2417: 2186 3890 M (-)S 

2418: /F1S42 F1 [66.664 0 0 -66.664 0 0 ] mFS

2419: F1S42 Ji 

2420: 1679 3966 M (j)S 

2421: 2163 3890 M (i)S 

2422: 1803 3890 M (j)S 

2423: 1541 3937 M (i)S 

2424: 1372 3937 M (w)S 

2425: /F1S53 F1 [83.332 0 0 -83.332 0 0 ] mFS

2426: F1S53 Ji 

2427: 2107 3869 M (w)S 

2428: 1958 3869 M (w)S 

2429: 1869 3869 M (w)S 

2430: 1758 3869 M (f)S 

2431: /F3S53 F3 [83.332 0 0 -83.332 0 0 ] mFS

2432: F3S53 Ji 

2433: 2259 3869 M (\))S 

2434: 2046 3869 M (...)[21 21  0]xS 

2435: 1923 3869 M (,)S 

2436: 1838 3869 M (\()S 

2437: 1480 3869 M (max)[63 37  0]xS 

2438: 1320 3869 M (max)[63 37  0]xS 

2439: /F3S42 F3 [66.664 0 0 -66.664 0 0 ] mFS

2440: F3S42 Ji 

2441: 2218 3890 M (1)S 

2442: 2009 3890 M (1)S 

2443: F3S4B Ji 

2444: 2305 3869 M (.  )[19 19  0]xS 

2445: 350 4070 M (GIS uses the maximum of this value to slow learning, while IIS )

2446: [53 25 42 26 38 29 33 29 26 21 38 33 25 57 33 37 21 57 38 57 25 38 24 25 21 38 21 29 25 37 33 21

2447: 38 33 25 21 38 25 29 21 38 53 25 21 33 33 25 38 21 38 37 19 25 53 38 21 21 33 25 25 25 42  0]xS 

2448: 350 4156 M (slows learning on a c)[29 21 38 53 29 20 21 33 33 25 38 21 38 37 20 38 38 20 33 19  0]xS 

2449: 987 4156 M (ase)[33 29  0]xS 

2450: 1082 4156 M (-)S 

2451: 1107 4156 M (by)[38  0]xS 

2452: 1181 4156 M (-)S 

2453: 1206 4156 M (case basis.  In some models, this sum )[33 33 29 33 19 38 33 29 21 29 19 19 19 25 38 19 29 38 57 33 19 57 38 38 33 21 29 19 19 21 38 21

2454: 29 19 29 38 57  0]xS 

2455: 350 4243 M (can be very different for different )[33 33 38 29 38 33 29 37 33 25 36 29 38 21 24 24 33 25 33 38 21 29 24 38 25 29 38 21 24 24 33 25

2456: 33 38 21  0]xS 

2457: F1S4B Ji 

2458: 1431 4243 M (w, )[50 19  0]xS 

2459: 1529 4243 M (i)S 

2460: F3S4B Ji 

2461: 1550 4243 M (.    In particular, models )[19 29 28 28 28 25 38 28 38 33 25 21 21 33 38 21 33 25 19 28 57 38 38 33 21 29  0]xS 

2462: 350 4329 M (using caching and triggering techniques can lead to these )[38 29 21 38 37 52 33 33 33 38 21 38 37 52 33 38 38 51 21 25 21 37 37 33 25 21 38 37 51 21 33 33

2463: 38 38 21 38 38 33 29 51 33 33 38 51 21 33 33 38 51 21 38 51 21 38 33 29 33  0]xS 

2464: 350 4415 M (different numbers of active indicators.  In other models, such as )

2465: [38 21 24 24 33 25 33 38 21 25 38 38 57 38 33 25 29 25 38 24 25 33 33 21 21 37 33 25 21 38 38 21

2466: 33 33 21 38 25 29 19 25 25 25 38 25 38 21 38 33 25 25 57 38 38 33 21 29 19 25 29 38 33 38 25 33

2467: 29  0]xS 

2468: 350 4501 M (n)S 

2469: 388 4501 M (-)S 

2470: 413 4501 M (gram)[37 25 33  0]xS 

2471: 565 4501 M (-)S 

2472: 590 4501 M (style m)[29 21 36 21 33 21  0]xS 

2473: 808 4501 M (odels, there is a fixed maximum for the number of )[38 38 33 21 29 19 21 21 38 33 25 33 21 21 29 20 33 20 24 21 37 33 38 20 57 33 37 21 57 38 57 20

2474: 24 38 25 20 21 38 33 20 38 38 57 38 33 25 20 38 24  0]xS 

2475: 350 4588 M (non)[38 38  0]xS 

2476: 464 4588 M (-)S 

2477: 489 4588 M (zero indicators.  In these models, IIS would lead to little or )

2478: [33 33 25 38 25 21 38 38 21 33 33 21 38 25 29 19 25 25 25 38 25 21 38 33 29 33 25 57 38 38 33 21

2479: 29 19 25 25 25 42 25 53 38 38 21 38 25 21 33 33 38 25 21 38 24 21 21 21 21 21 33 24 38 25  0]xS 

2480: 350 4674 M (no reduction in the number of iterations of iterative scaling, and, )

2481: [38 38 23 25 33 38 38 33 21 21 38 38 23 21 38 23 21 38 33 23 38 38 57 38 33 25 23 38 24 22 21 21

2482: 33 25 33 21 21 38 38 29 22 38 24 22 21 21 33 25 33 21 21 37 33 22 29 33 33 21 21 38 37 19 22 33

2483: 38 38 19  0]xS 

2484: 350 4760 M (because of the additional overhead for each iteration, might )

2485: [38 33 33 33 38 29 33 43 38 24 43 21 38 33 42 33 38 38 21 21 21 38 38 33 21 42 38 37 33 25 38 33

2486: 33 38 42 24 38 25 42 33 33 33 38 42 21 21 33 25 33 21 21 38 38 19 42 57 21 37 38 21  0]xS 

2487: 350 4846 M (actually l)[33 33 21 38 33 21 21 36 19  0]xS 

2488: 626 4846 M (ead to a slowdown.)[33 33 38 19 21 38 19 33 19 29 21 38 53 38 38 53 38  0]xS 

2489: 1206 4846 M ( )S 

2490: 463 4933 M (The last technique we consider is the most powerful one, )[45 38 33 36 21 33 29 21 36 21 33 33 38 38 21 38 38 33 36 53 33 36 33 38 38 29 21 38 33 25 36 21

2491: 29 36 21 38 33 36 57 38 29 21 35 38 38 53 33 25 24 38 21 35 38 38 33 19  0]xS 

2492: 350 5019 M (cluster expansion, introduced in )[33 21 38 29 21 33 25 20 33 37 38 33 38 29 21 38 38 19 20 21 38 21 25 38 38 38 33 33 38 20 21 38

2493:  0]xS 

2494: 1336 5019 M ([7])[26 38  0]xS 

2495: 1426 5019 M ( and expanded in )[20 33 38 38 20 33 37 38 33 38 38 33 38 19 21 38  0]xS 

2496: 1960 5019 M ([8])[26 38  0]xS 

2497: 2050 5019 M (.  Cluster )[19 19 19 50 21 38 29 21 33 25  0]xS 

2498: 350 5105 M (expansion can be regarded as a natural extension of uni)[33 37 38 33 38 29 21 38 38 37 33 33 38 37 38 33 37 25 33 37 33 25 38 33 38 37 33 29 37 33 36 38

2499: 33 21 38 25 33 21 36 33 37 21 33 38 29 21 38 38 36 38 24 36 38 38  0]xS 

2500: 2172 5105 M (gram )[37 25 33 57  0]xS 

2501: 350 5191 M (caching to n)[33 33 33 38 21 38 37 34 21 38 34  0]xS 

2502: 748 5191 M (-)S 

2503: 773 5191 M (grams.  Consider a simple trigram model.  With )[37 25 33 57 29 19 34 34 50 38 38 29 21 38 33 25 34 33 34 29 21 57 38 21 33 34 21 25 21 37 25 33

2504: 57 34 57 38 38 33 21 19 34 34 70 21 21 38  0]xS 

2505: 350 5278 M (some straightforward rearrangement of the equations, for two )

2506: [29 38 57 33 38 29 21 25 33 21 37 38 21 24 38 25 53 33 25 38 38 25 33 33 25 25 33 38 37 33 57 33

2507: 38 21 38 38 24 38 21 38 33 38 33 38 38 33 21 21 38 38 29 19 37 24 38 25 37 21 53 38  0]xS 

2508: 350 5364 M (trigrams with a common bigram, most of the computation can be )

2509: [21 25 21 37 25 33 57 29 22 53 21 21 38 22 33 22 33 38 57 57 38 38 22 38 21 37 25 33 57 19 22 57

2510: 38 29 21 22 38 24 22 21 38 33 22 33 38 57 38 38 21 33 21 21 38 38 22 33 33 38 21 38 33  0]xS 

2511: 350 5450 M (shared.  In )[29 38 33 25 33 38 19 36 36 25 38  0]xS 

2512: 736 5450 M ([8])[26 38  0]xS 

2513: 826 5450 M (, this technique is exte)[19 35 21 38 21 29 35 21 33 33 38 38 21 38 38 33 35 21 29 35 33 37 21  0]xS 

2514: 1561 5450 M (nded to handle cases in )[38 38 33 38 35 21 38 35 38 33 38 38 21 33 35 33 33 29 33 29 35 21 38  0]xS 

2515: 350 5536 M (which there is limited interaction between hierarchical )[53 38 21 33 38 76 21 38 33 25 33 76 21 29 76 21 21 57 21 21 33 38 76 21 38 21 33 25 33 33 21 21

2516: 38 38 75 38 33 21 53 33 33 38 75 38 21 33 25 33 25 33 38 21 33 33 21  0]xS 

2517: 350 5623 M (constraints, and still achieves good speedups \(a factor of 15.\)  )

2518: [33 38 38 29 21 25 33 21 38 21 29 19 35 33 38 38 35 29 21 21 21 21 35 33 33 38 21 33 37 33 29 34

2519: 37 38 38 38 34 29 38 33 33 38 38 38 29 34 25 33 34 24 33 33 21 38 25 34 38 24 34 38 38 19 25 19

2520:  0]xS 

2521: 350 5709 M (However, as )[54 38 53 33 37 33 25 19 32 33 29  0]xS 

2522: 767 5709 M ([7])[26 38  0]xS 

2523: 857 5709 M ( concludes, cluster expansion \223is limited in its )[31 33 38 38 33 21 38 38 33 29 19 31 33 21 38 29 21 33 25 31 33 37 38 33 38 29 21 38 38 31 33 21

2524: 29 31 21 21 57 21 21 33 38 31 21 38 31 21 21 29  0]xS 

2525: 2525 568 M (usefulness\205 )[38 29 33 24 38 21 38 33 29 29 75  0]xS 

2526: 2953 568 M (W)S 

2527: 3023 568 M (hen)[38 33  0]xS 

2528: 3132 568 M ( the number of  interacting constraints is )[41 21 38 33 41 38 38 57 38 33 25 41 38 24 41 41 21 38 21 33 25 33 33 21 21 38 37 41 33 38 38 29

2529: 21 25 33 21 38 21 29 41 21 29  0]xS 

2530: 2525 654 M (large\205the cluster expansion is of little use in computing the )

2531: [21 33 25 37 33 75 21 38 33 37 33 21 38 29 21 33 25 37 33 37 38 33 38 29 21 38 38 37 21 29 37 38

2532: 24 37 21 21 21 21 21 33 37 38 29 33 37 21 38 36 33 38 57 38 38 21 21 38 37 36 21 38 33  0]xS 

2533: 2525 740 M (exact maximum entropy solution.\224  We believe that the same )[33 37 33 33 21 36 57 33 37 21 57 38 57 35 33 38 21 25 38 38 36 35 29 38 21 38 21 21 38 38 19 33

2534: 35 35 70 33 35 38 33 21 21 33 37 33 35 21 38 33 21 35 21 38 33 35 29 33 57 33  0]xS 

2535: 2525 827 M (conclusion applies to )[33 38 38 33 21 38 29 21 38 38 23 33 38 38 21 21 33 29 23 21 38  0]xS 

2536: 3193 827 M ([8])[26 38  0]xS 

2537: 3283 827 M (.  In particular, a simple model combi)[19 23 23 25 38 23 38 33 25 21 21 33 38 21 33 25 19 23 33 22 29 21 57 38 21 33 22 57 38 38 33 21

2538: 22 33 38 57 38  0]xS 

2539: 4436 827 M (n)S 

2540: 4474 827 M (-)S 

2541: 2525 913 M (ing bigram 1)[21 38 37 31 38 21 37 25 33 57 31  0]xS 

2542: 2932 913 M (-)S 

2543: 2957 913 M (back, 2)[38 33 33 37 19 31  0]xS 

2544: 3186 913 M (-)S 

2545: 3211 913 M (back, \205, 5)[38 33 33 37 19 31 75 19 31  0]xS 

2546: 3565 913 M (-)S 

2547: 3590 913 M (back constraints would prob)[38 33 33 37 30 33 38 38 29 21 25 33 21 38 21 29 30 53 38 38 21 38 30 38 25 38  0]xS 

2548: 4474 913 M (-)S 

2549: 2525 999 M (ably show no, or only small, gains from the techniques of )[33 38 21 36 24 29 38 38 53 24 38 38 19 24 38 25 24 38 38 21 36 24 29 57 33 21 21 19 24 37 33 21

2550: 38 29 24 24 25 38 57 24 21 38 33 24 21 33 33 38 38 21 38 38 33 29 24 38 24  0]xS 

2551: 4323 999 M ([7])[26 38  0]xS 

2552: 4413 999 M ( or )[23 38 25  0]xS 

2553: 2525 1085 M ([8])[26 38  0]xS 

2554: 2615 1085 M (, while for our technique, the gains would be about the same.)

2555: [19 19 53 38 21 21 33 19 24 38 25 19 38 38 25 19 21 33 33 38 38 21 38 38 33 19 19 21 38 33 19 37

2556: 33 21 38 29 19 53 38 38 21 38 19 38 33 19 33 38 38 38 21 19 21 38 33 19 29 33 57 33  0]xS 

2557: 4449 1085 M ( )S 

2558: 2638 1172 M (In the)[25 38 27 21 38  0]xS 

2559: 2820 1172 M (ory, our speedup can be used in conjunction with IIS, )[38 25 36 19 27 38 38 25 27 29 38 33 33 38 38 38 27 33 33 38 27 38 33 27 38 29 33 38 27 21 38 26

2560: 33 38 38 21 38 38 33 21 21 38 38 26 53 21 21 38 26 25 25 42 19  0]xS 

2561: 2525 1258 M (unigram caching, or cluster expansion.  However, in conjunction )

2562: [38 38 21 37 25 33 57 23 33 33 33 38 21 38 37 19 23 38 25 22 33 21 38 29 21 33 25 22 33 37 38 33

2563: 38 29 21 38 38 19 22 22 54 38 53 33 37 33 25 19 22 21 38 22 33 38 38 21 38 38 33 21 21 38 38  0]xS 

2564: 2525 1344 M (with unigram caching, in experiments, it typically leads to only )

2565: [53 21 21 38 28 38 38 21 37 25 33 57 28 33 33 33 38 21 38 37 19 28 21 38 28 33 37 38 33 25 21 57

2566: 33 38 21 29 19 28 21 21 28 21 36 38 21 33 33 21 21 36 27 21 33 33 38 29 27 21 38 27 38 38 21 36

2567:  0]xS 

2568: 2525 1430 M (small speed improvements, or sometimes actual slowdowns )[29 57 33 21 21 50 29 38 33 33 38 50 21 57 38 25 38 37 33 57 33 38 21 29 19 50 38 25 50 29 38 57

2569: 33 21 21 57 33 29 50 33 33 21 38 33 21 49 29 21 38 53 38 38 53 38 29  0]xS 

2570: 2525 1517 M (\(because unigram )[25 38 33 33 33 38 29 33 28 38 38 21 37 25 33 57  0]xS 

2571: 3092 1517 M (caching introduces overhead in other parts of )[33 33 33 38 21 38 37 28 21 38 21 25 38 38 38 33 33 29 27 38 37 33 25 38 33 33 38 27 21 38 27 38

2572: 21 38 33 25 27 38 33 25 21 29 27 38 24  0]xS 

2573: 2525 1603 M (the algorithm\).  Similarly, we suspect that with cluster expansion )

2574: [21 38 33 22 33 21 37 38 25 21 21 38 57 25 19 22 22 42 21 57 21 21 33 25 21 36 19 22 53 33 21 29

2575: 38 29 38 33 33 21 21 21 38 33 21 21 53 21 21 38 21 33 21 38 29 21 33 25 21 33 37 38 33 38 29 21

2576: 38 38  0]xS 

2577: 2525 1689 M (speedup might be limited.  We have not tested our algorithm with )

2578: [29 38 33 33 38 38 38 20 57 21 37 38 21 19 38 33 19 21 21 57 21 21 33 38 19 19 19 70 33 19 38 33

2579: 37 33 19 38 38 21 19 21 33 29 21 33 38 19 38 38 25 19 33 21 37 38 25 21 21 38 57 19 53 21 21 38

2580:  0]xS 

2581: 2525 1775 M (IIS, but in principle, there is no reason they could not be )[25 25 42 19 44 38 38 21 44 21 38 44 38 25 21 38 33 21 38 21 33 19 44 21 38 33 25 33 44 21 29 44

2582: 38 38 44 25 33 33 29 38 38 44 21 38 33 36 44 33 38 38 21 38 44 38 38 21 44 38 33  0]xS 

2583: 2525 1862 M (combined, and we g)[33 38 57 38 21 38 33 38 19 19 33 38 38 19 53 33 19  0]xS 

2584: 3129 1862 M (uess the combination would work well.)[38 33 29 29 19 21 38 33 19 33 38 57 38 21 38 33 21 21 38 38 19 53 38 38 21 38 19 53 38 25 37 19

2585: 53 33 21 21  0]xS 

2586: 4309 1862 M ( )S 

2587: F4S64 Ji 

2588: 3204 2022 M (4.)[50  0]xS 

2589: 3279 2022 M ( )S 

2590: 3339 2022 M (RESULTS)[72 67 56 72 67 67  0]xS 

2591: 3796 2022 M ( )S 

2592: F3S4B Ji 

2593: 2525 2163 M (We ran our experiments using four different learning techniques: )

2594: [70 33 22 25 33 38 22 38 38 25 22 33 37 38 33 25 21 57 33 38 21 29 22 38 29 21 38 37 22 24 38 38

2595: 25 22 38 21 24 24 33 25 33 38 21 22 21 33 33 25 38 21 38 37 21 21 33 33 38 38 21 38 38 33 29 21

2596:  0]xS 

2597: 2525 2249 M (simple GIS, GIS with unigram caching, GIS with a two level )[29 21 57 38 21 33 34 53 25 42 19 34 53 25 42 34 53 21 21 38 34 38 38 21 37 25 33 57 34 33 33 33

2598: 38 21 38 37 19 34 53 25 42 34 53 21 21 38 34 33 34 21 53 38 33 21 33 37 33 21  0]xS 

2599: 2525 2335 M (clustering, and GIS with a three level clustering.  We ran on four )

2600: [33 21 38 29 21 33 25 21 38 37 19 22 33 38 38 22 53 25 42 21 53 21 21 38 21 33 21 21 38 25 33 33

2601: 21 21 33 37 33 21 21 33 21 38 29 21 33 25 21 38 37 19 21 21 70 33 21 25 33 38 21 38 38 21 24 38

2602: 38 25  0]xS 

2603: 2525 2422 M (different sizes of )[38 21 24 24 33 25 33 38 21 27 29 21 33 33 29 27 38 24  0]xS 

2604: 3070 2422 M (training data.  The model used is a \223skipping, )[21 25 33 21 38 21 38 37 27 38 33 21 33 19 27 27 45 38 33 27 57 38 38 33 21 27 38 29 33 38 27 21

2605: 29 27 33 27 33 29 37 21 38 38 21 38 37 19  0]xS 

2606: 2525 2508 M (classing\224 model with the following types of indicator functions, )

2607: [33 21 33 29 29 21 38 37 33 27 57 38 38 33 21 27 53 21 21 38 27 21 38 33 26 24 38 21 21 38 53 21

2608: 38 37 26 21 36 38 33 29 26 38 24 26 21 38 38 21 33 33 21 38 25 26 24 38 38 33 21 21 38 38 29 19

2609:  0]xS 

2610: 2525 2594 M (where )[53 38 33 25 33  0]xS 

2611: F1S4B Ji 

2612: 2730 2594 M (W)S 

2613: F3S4B Ji 

2614: 2792 2594 M (, )[19  0]xS 

2615: F1S4B Ji 

2616: 2833 2594 M (Y)S 

2617: F3S4B Ji 

2618: 2875 2594 M (, and )[19 22 33 38 38  0]xS 

2619: F1S4B Ji 

2620: 3047 2594 M (Z)S 

2621: F3S4B Ji 

2622: 3089 2594 M ( are variables filled in for specific instances of )[22 33 25 33 22 37 33 25 21 33 38 21 33 29 22 24 21 21 21 33 38 22 21 38 22 24 38 25 22 29 38 33

2623: 33 21 24 21 33 22 21 38 29 21 33 38 33 33 29 22 38 24  0]xS 

2624: 2525 2680 M (the indicator functions.  )[21 38 33 19 21 38 38 21 33 33 21 38 25 19 24 38 38 33 21 21 38 38 29 19 19  0]xS 

2625: 3260 2680 M ( )S 

2626: F1S4B Ji 

2627: 2639 2788 M (f)S 

2628: F1S43 Ji 

2629: 2660 2801 M (j)S 

2630: F3S43 Ji 

2631: 2679 2742 M (unigram)[34 34 19 33 22 30  0]xS 

2632: F3S4B Ji 

2633: 2903 2788 M (\()S 

2634: F1S4B Ji 

2635: 2928 2788 M (w)S 

2636: F3S4B Ji 

2637: 2978 2788 M (, )[19  0]xS 

2638: F1S4B Ji 

2639: 3016 2788 M (w)S 

2640: F1S43 Ji 

2641: 3066 2801 M (1)S 

2642: F3S4B Ji 

2643: 3100 2788 M <85>S 

2644: F1S4B Ji 

2645: 3175 2788 M (w)S 

2646: F1S43 Ji 

2647: 3225 2801 M (i)S 

2648: 3244 2801 M (-)S 

2649: 3266 2801 M (1)S 

2650: F3S4B Ji 

2651: 3300 2788 M (\)=1 if )[25 42 38 19 21 24  0]xS 

2652: F1S4B Ji 

2653: 3488 2788 M (w)S 

2654: F3S4B Ji 

2655: 3538 2788 M (=)S 

2656: F1S4B Ji 

2657: 3580 2788 M (W)S 

2658: F3S4B Ji 

2659: 3642 2788 M ( )S 

2660: F1S4B Ji 

2661: 2639 2913 M (f)S 

2662: F1S43 Ji 

2663: 2660 2926 M (j)S 

2664: F3S43 Ji 

2665: 2679 2867 M (class)[30 18 30 26  0]xS 

2666: 2809 2867 M (-)S 

2667: 2831 2867 M (bigram)[34 19 33 22 30  0]xS 

2668: F3S4B Ji 

2669: 3021 2913 M ( \()[19  0]xS 

2670: F1S4B Ji 

2671: 3065 2913 M (w)S 

2672: F3S4B Ji 

2673: 3115 2913 M (,)S 

2674: 3134 2913 M ( )S 

2675: F1S4B Ji 

2676: 3153 2913 M (w)S 

2677: F1S43 Ji 

2678: 3203 2926 M (1)S 

2679: F3S4B Ji 

2680: 3237 2913 M <85>S 

2681: F1S4B Ji 

2682: 3312 2913 M (w)S 

2683: F1S43 Ji 

2684: 3362 2926 M (i)S 

2685: 3381 2926 M (-)S 

2686: 3403 2926 M (1)S 

2687: F3S4B Ji 

2688: 3437 2913 M (\)=1 )[25 42 38  0]xS 

2689: 3561 2913 M (if )[21 24  0]xS 

2690: F1S4B Ji 

2691: 3625 2913 M (w)S 

2692: F3S4B Ji 

2693: 3675 2913 M (=)S 

2694: F1S4B Ji 

2695: 3717 2913 M (W)S 

2696: F3S4B Ji 

2697: 3779 2913 M ( and )[19 33 38 38  0]xS 

2698: F1S4B Ji 

2699: 3926 2913 M (class)[33 21 38 29  0]xS 

2700: F3S4B Ji 

2701: 4076 2913 M (\()S 

2702: F1S4B Ji 

2703: 4101 2913 M (w)S 

2704: F1S43 Ji 

2705: 4151 2926 M (i)S 

2706: 4170 2926 M (-)S 

2707: F3S43 Ji 

2708: 4192 2926 M (1)S 

2709: F3S4B Ji 

2710: 4226 2913 M (\)=)[25  0]xS 

2711: F1S4B Ji 

2712: 4293 2913 M (Z)S 

2713: F3S4B Ji 

2714: 4335 2913 M ( )S 

2715: F1S4B Ji 

2716: 2639 3038 M (f)S 

2717: F1S43 Ji 

2718: 2660 3051 M (j)S 

2719: F3S43 Ji 

2720: 2679 2992 M (class)[30 18 30 26  0]xS 

2721: 2809 2992 M (-)S 

2722: 2831 2992 M (skip)[26 34 19  0]xS 

2723: 2944 2992 M (-)S 

2724: 2966 2992 M (bigram)[34 19 33 22 30  0]xS 

2725: F3S4B Ji 

2726: 3156 3038 M ( \()[19  0]xS 

2727: F1S4B Ji 

2728: 3200 3038 M (w)S 

2729: F3S4B Ji 

2730: 3250 3038 M (, )[19  0]xS 

2731: F1S4B Ji 

2732: 3288 3038 M (w)S 

2733: F1S43 Ji 

2734: 3338 3051 M (1)S 

2735: F3S4B Ji 

2736: 3372 3038 M <85>S 

2737: F1S4B Ji 

2738: 3447 3038 M (w)S 

2739: F1S43 Ji 

2740: 3497 3051 M (i)S 

2741: 3516 3051 M (-)S 

2742: 3538 3051 M (1)S 

2743: F3S4B Ji 

2744: 3572 3038 M (\)=1 )[25 42 38  0]xS 

2745: 3696 3038 M (if w=W and class\(w)[21 24 19 53 42 70 19 33 38 38 19 33 21 33 29 29 25  0]xS 

2746: F1S43 Ji 

2747: 4295 3051 M (i)S 

2748: 4314 3051 M (-)S 

2749: 4336 3051 M (2)S 

2750: F3S4B Ji 

2751: 4370 3038 M (\)=Z)[25 42  0]xS 

2752: 4483 3038 M ( )S 

2753: F1S4B Ji 

2754: 2639 3164 M (f)S 

2755: F1S43 Ji 

2756: 2660 3177 M (j)S 

2757: F3S43 Ji 

2758: 2679 3118 M (bigram)[34 19 33 22 30  0]xS 

2759: F3S4B Ji 

2760: 2869 3164 M ( \()[19  0]xS 

2761: F1S4B Ji 

2762: 2913 3164 M (w)S 

2763: F3S4B Ji 

2764: 2963 3164 M (, )[19  0]xS 

2765: F1S4B Ji 

2766: 3001 3164 M (w)S 

2767: F1S43 Ji 

2768: 3051 3177 M (1)S 

2769: F3S4B Ji 

2770: 3085 3164 M <85>S 

2771: F1S4B Ji 

2772: 3160 3164 M (w)S 

2773: F1S43 Ji 

2774: 3210 3177 M (i)S 

2775: 3229 3177 M (-)S 

2776: 3251 3177 M (1)S 

2777: F3S4B Ji 

2778: 3285 3164 M (\)=1 )[25 42 38  0]xS 

2779: 3409 3164 M (if w=W and w)[21 24 19 53 42 70 19 33 38 38 19  0]xS 

2780: F1S43 Ji 

2781: 3838 3177 M (i)S 

2782: 3857 3177 M (-)S 

2783: 3879 3177 M (1)S 

2784: F3S4B Ji 

2785: 3913 3164 M (=Z)[42  0]xS 

2786: 4001 3164 M ( )S 

2787: F1S4B Ji 

2788: 2639 3289 M (f)S 

2789: F1S43 Ji 

2790: 2660 3302 M (j)S 

2791: F3S43 Ji 

2792: 2679 3243 M (skip)[26 34 19  0]xS 

2793: 2792 3243 M (-)S 

2794: 2814 3243 M (bigram)[34 19 33 22 30  0]xS 

2795: F3S4B Ji 

2796: 3004 3289 M ( \()[19  0]xS 

2797: F1S4B Ji 

2798: 3048 3289 M (w)S 

2799: F3S4B Ji 

2800: 3098 3289 M (, )[19  0]xS 

2801: F1S4B Ji 

2802: 3136 3289 M (w)S 

2803: F1S43 Ji 

2804: 3186 3302 M (1)S 

2805: F3S4B Ji 

2806: 3220 3289 M <85>S 

2807: F1S4B Ji 

2808: 3295 3289 M (w)S 

2809: F1S43 Ji 

2810: 3345 3302 M (i)S 

2811: 3364 3302 M (-)S 

2812: 3386 3302 M (1)S 

2813: F3S4B Ji 

2814: 3420 3289 M (\)=1 )[25 42 38  0]xS 

2815: 3544 3289 M (if w=W and w)[21 24 19 53 42 70 19 33 38 38 19  0]xS 

2816: F1S43 Ji 

2817: 3973 3302 M (i)S 

2818: 3992 3302 M (-)S 

2819: 4014 3302 M (2)S 

2820: F3S4B Ji 

2821: 4048 3289 M (=Z)[42  0]xS 

2822: 4136 3289 M ( )S 

2823: F1S4B Ji 

2824: 2639 3414 M (f)S 

2825: F1S43 Ji 

2826: 2660 3427 M (j)S 

2827: F3S43 Ji 

2828: 2679 3368 M (class)[30 18 30 26  0]xS 

2829: 2809 3368 M (-)S 

2830: 2831 3368 M (trigram)[19 22 19 33 22 30  0]xS 

2831: F3S4B Ji 

2832: 3028 3414 M ( \()[22  0]xS 

2833: F1S4B Ji 

2834: 3075 3414 M (w)S 

2835: F3S4B Ji 

2836: 3125 3414 M (, )[19  0]xS 

2837: F1S4B Ji 

2838: 3166 3414 M (w)S 

2839: F1S43 Ji 

2840: 3216 3427 M (1)S 

2841: F3S4B Ji 

2842: 3250 3414 M <85>S 

2843: F1S4B Ji 

2844: 3325 3414 M (w)S 

2845: F1S43 Ji 

2846: 3375 3427 M (i)S 

2847: 3394 3427 M (-)S 

2848: 3416 3427 M (1)S 

2849: F3S4B Ji 

2850: 3450 3414 M (\)=1 )[25 42 38  0]xS 

2851: 3576 3414 M (if w=W and class\(w)[21 24 21 53 42 70 21 33 38 38 21 33 21 33 29 29 25  0]xS 

2852: F1S43 Ji 

2853: 4181 3427 M (i)S 

2854: 4200 3427 M (-)S 

2855: 4222 3427 M (1)S 

2856: F3S4B Ji 

2857: 4256 3414 M (\)=Z and )[25 42 46 21 33 38 38  0]xS 

2858: 2525 3506 M (class\(w)[33 21 33 29 29 25  0]xS 

2859: F1S43 Ji 

2860: 2748 3519 M (i)S 

2861: 2767 3519 M (-)S 

2862: 2789 3519 M (2)S 

2863: F1S4B Ji 

2864: 2823 3506 M (\))S 

2865: F3S4B Ji 

2866: 2848 3506 M (=Y)[42  0]xS 

2867: 2944 3506 M ( )S 

2868: F1S4B Ji 

2869: 2639 3632 M (f)S 

2870: F1S43 Ji 

2871: 2660 3645 M (j)S 

2872: F3S43 Ji 

2873: 2679 3586 M (class)[30 18 30 26  0]xS 

2874: 2809 3586 M (-)S 

2875: 2831 3586 M (bigram)[34 19 33 22 30  0]xS 

2876: 3021 3586 M (-)S 

2877: 3043 3586 M (skip)[26 34 19  0]xS 

2878: 3156 3586 M (-)S 

2879: 3178 3586 M (bigram)[34 19 33 22 30  0]xS 

2880: F3S4B Ji 

2881: 3368 3632 M ( \()[67  0]xS 

2882: F1S4B Ji 

2883: 3460 3632 M (w)S 

2884: F3S4B Ji 

2885: 3510 3632 M (, )[19  0]xS 

2886: F1S4B Ji 

2887: 3595 3632 M (w)S 

2888: F1S43 Ji 

2889: 3645 3645 M (1)S 

2890: F3S4B Ji 

2891: 3679 3632 M <85>S 

2892: F1S4B Ji 

2893: 3754 3632 M (w)S 

2894: F1S43 Ji 

2895: 3804 3645 M (i)S 

2896: 3823 3645 M (-)S 

2897: 3845 3645 M (1)S 

2898: F3S4B Ji 

2899: 3879 3632 M (\)=1 )[25 42 38  0]xS 

2900: 4050 3632 M (if )[21 24  0]xS 

2901: F1S4B Ji 

2902: 4161 3632 M (w=W)[50 51  0]xS 

2903: F3S4B Ji 

2904: 4324 3632 M ( and )[66 33 38 38  0]xS 

2905: F1S4B Ji 

2906: 2525 3723 M (class)[33 21 38 29  0]xS 

2907: F3S4B Ji 

2908: 2675 3723 M (\()S 

2909: F1S4B Ji 

2910: 2700 3723 M (w)S 

2911: F1S43 Ji 

2912: 2750 3736 M (i)S 

2913: 2769 3736 M (-)S 

2914: F3S43 Ji 

2915: 2791 3736 M (1)S 

2916: F3S4B Ji 

2917: 2825 3723 M (\)=)[25  0]xS 

2918: F1S4B Ji 

2919: 2892 3723 M (Z)S 

2920: F3S4B Ji 

2921: 2934 3723 M ( and )[19 33 38 38  0]xS 

2922: F1S4B Ji 

2923: 3081 3723 M (w)S 

2924: F1S43 Ji 

2925: 3131 3736 M (i)S 

2926: F3S43 Ji 

2927: 3150 3736 M (-)S 

2928: 3172 3736 M (2)S 

2929: F3S4B Ji 

2930: 3206 3723 M (=)S 

2931: F1S4B Ji 

2932: 3248 3723 M (Y)S 

2933: F3S4B Ji 

2934: 3290 3723 M ( )S 

2935: F1S4B Ji 

2936: 2639 3849 M (f)S 

2937: F1S43 Ji 

2938: 2660 3862 M (j)S 

2939: F3S43 Ji 

2940: 2679 3803 M (bigram)[34 19 33 22 30  0]xS 

2941: 2869 3803 M (-)S 

2942: 2891 3803 M (class)[30 18 30 26  0]xS 

2943: 3021 3803 M (-)S 

2944: 3043 3803 M (skip)[26 34 19  0]xS 

2945: 3156 3803 M (-)S 

2946: 3178 3803 M (bigram)[34 19 33 22 30  0]xS 

2947: F3S4B Ji 

2948: 3368 3849 M ( \()[22  0]xS 

2949: F1S4B Ji 

2950: 3415 3849 M (w)S 

2951: F3S4B Ji 

2952: 3465 3849 M (, )[19  0]xS 

2953: F1S4B Ji 

2954: 3506 3849 M (w)S 

2955: F1S43 Ji 

2956: 3556 3862 M (1)S 

2957: F3S4B Ji 

2958: 3590 3849 M <85>S 

2959: F1S4B Ji 

2960: 3665 3849 M (w)S 

2961: F1S43 Ji 

2962: 3715 3862 M (i)S 

2963: 3734 3862 M (-)S 

2964: 3756 3862 M (1)S 

2965: F3S4B Ji 

2966: 3790 3849 M (\)=1 )[25 42 38  0]xS 

2967: 3917 3849 M (if )[21 24  0]xS 

2968: F1S4B Ji 

2969: 3984 3849 M (w)S 

2970: F3S4B Ji 

2971: 4034 3849 M (=)S 

2972: F1S4B Ji 

2973: 4076 3849 M (W)S 

2974: F3S4B Ji 

2975: 4138 3849 M ( and )[22 33 38 38  0]xS 

2976: F1S4B Ji 

2977: 4290 3849 M (w)S 

2978: F1S43 Ji 

2979: 4340 3862 M (i)S 

2980: 4359 3862 M (-)S 

2981: F3S43 Ji 

2982: 4381 3862 M (1)S 

2983: F3S4B Ji 

2984: 4415 3849 M (=)S 

2985: F1S4B Ji 

2986: 4457 3849 M (Z)S 

2987: F3S4B Ji 

2988: 4499 3849 M ( )S 

2989: 2525 3940 M (and )[33 38 38  0]xS 

2990: F1S4B Ji 

2991: 2653 3940 M (class)[33 21 38 29  0]xS 

2992: F3S4B Ji 

2993: 2803 3940 M (\()S 

2994: F1S4B Ji 

2995: 2828 3940 M (w)S 

2996: F1S43 Ji 

2997: 2878 3953 M (i)S 

2998: 2897 3953 M (-)S 

2999: F3S43 Ji 

3000: 2919 3953 M (2)S 

3001: F3S4B Ji 

3002: 2953 3940 M (\)=)[25  0]xS 

3003: F1S4B Ji 

3004: 3020 3940 M (Y)S 

3005: F3S4B Ji 

3006: 3062 3940 M ( )S 

3007: 2638 4045 M (We used all and only indicator functions where there were at )

3008: [70 33 23 38 29 33 38 23 33 21 21 23 33 38 38 23 38 38 21 36 23 21 38 38 21 33 33 21 38 25 22 24

3009: 38 38 33 21 21 38 38 29 22 53 38 33 25 33 22 21 38 33 25 33 22 53 33 25 33 22 33 21  0]xS 

3010: 2525 4131 M (least three matching cases in the t)[21 33 33 29 21 32 21 38 25 33 33 32 57 33 21 33 38 21 38 37 31 33 33 29 33 29 31 21 38 31 21 38

3011: 33 31  0]xS 

3012: 3607 4131 M (raining data.  We found our )[25 33 21 38 21 38 37 31 38 33 21 33 19 31 31 70 33 31 24 38 38 38 38 31 38 38 25  0]xS 

3013: 2525 4218 M (word classes by using a top)[53 38 25 38 51 33 21 33 29 29 33 29 51 38 36 51 38 29 21 38 37 50 33 50 21 38  0]xS 

3014: 3506 4218 M (-)S 

3015: 3531 4218 M (down splitting algorithm that )[38 38 53 38 50 29 38 21 21 21 21 21 38 37 50 33 21 37 38 25 21 21 38 57 50 21 38 33 21  0]xS 

3016: 2525 4304 M (at)[33  0]xS 

3017: 2579 4304 M (tempted to minimize entropy loss, as described in )[21 33 57 38 21 33 38 20 21 38 20 57 21 38 21 57 21 33 33 20 33 38 21 25 38 38 36 20 21 38 29 29

3018: 19 20 33 29 20 38 33 29 33 25 21 38 33 38 19 21 38  0]xS 

3019: 4092 4304 M ([9])[26 38  0]xS 

3020: 4182 4304 M (.  We used )[19 19 19 70 33 19 38 29 33 38  0]xS 

3021: 2525 4390 M (dif)[38 21  0]xS 

3022: 2608 4390 M (ferent numbers of classes for different purposes.  For the two)

3023: [24 33 25 33 38 21 23 38 38 57 38 33 25 29 23 38 24 23 33 21 33 29 29 33 29 23 24 38 25 23 38 21

3024: 24 24 33 25 33 38 21 22 38 38 25 38 38 29 33 29 19 22 22 42 38 25 22 21 38 33 22 21 53  0]xS 

3025: 4474 4390 M (-)S 

3026: 2525 4476 M (lev)[21 33  0]xS 

3027: 2616 4476 M (el spl)[33 21 30 29 38  0]xS 

3028: 2788 4476 M (itting, we used approximately 60)[21 21 21 21 38 37 19 30 53 33 30 38 29 33 38 30 33 38 38 25 38 37 21 57 33 21 33 21 36 30 38  0]xS 

3029: 3817 4476 M (-)S 

3030: 3842 4476 M (250 classes.  For the )[38 38 38 30 33 21 33 29 29 33 29 19 30 30 42 38 25 30 21 38 33  0]xS 

3031: 2525 4563 M (three)[21 38 25 33  0]xS 

3032: 2675 4563 M (-)S 

3033: 2700 4563 M (level splitting, we used approximately 8)[21 33 37 33 21 26 29 38 21 21 21 21 21 38 37 19 26 53 33 26 38 29 33 38 26 33 38 38 25 38 37 21

3034: 57 33 21 33 21 36 26  0]xS 

3035: 3934 4563 M (-)S 

3036: 3959 4563 M (30 classes for the )[38 38 26 33 21 33 29 29 33 29 26 24 38 25 26 21 38 33  0]xS 

3037: 2525 4649 M (first level, and 100)[24 21 25 29 21 36 21 33 37 33 21 19 36 33 38 38 36 38 38  0]xS 

3038: 3140 4649 M (-)S 

3039: 3165 4649 M (2000 classes for the second level.  In all )[38 38 38 38 36 33 21 33 29 29 33 29 36 24 38 25 36 21 38 33 36 29 33 33 38 38 38 36 21 33 37 33

3040: 21 19 35 35 25 38 35 33 21 21  0]xS 

3041: 2525 4735 M (cases, we optimized the number of classes by running one iter)

3042: [33 33 29 33 29 19 28 53 33 28 38 38 21 21 57 21 33 33 38 28 21 38 33 28 38 38 57 38 33 25 28 38

3043: 24 28 33 21 33 29 29 33 29 28 38 36 28 25 38 38 38 21 38 37 28 38 38 33 27 21 21 33  0]xS 

3044: 4474 4735 M (-)S 

3045: 2525 4821 M (ation of t)[33 21 21 38 38 20 38 24 20  0]xS 

3046: 2799 4821 M (raining with varying numbers of classes, and picking the )[25 33 21 38 21 38 37 20 53 21 21 38 20 37 33 25 36 21 38 37 20 38 38 57 38 33 25 29 19 38 24 19

3047: 33 21 33 29 29 33 29 19 19 33 38 38 19 38 21 33 37 21 38 37 19 21 38 33  0]xS 

3048: 2525 4908 M (fast)[24 33 29  0]xS 

3049: 2632 4908 M (est.  The classes used in the indicator functions are not typ)

3050: [33 29 21 19 27 27 45 38 33 27 33 21 33 29 29 33 29 27 38 29 33 38 27 21 38 27 21 38 33 27 21 38

3051: 38 21 33 33 21 38 25 27 24 38 38 33 21 21 38 38 29 27 33 25 33 27 38 38 21 27 21 36  0]xS 

3052: 4474 4908 M (-)S 

3053: 2525 4994 M (ically the same as the classes used in our factoring; for the in)

3054: [21 33 33 21 21 36 22 21 38 33 22 29 33 57 33 22 33 29 22 21 38 33 22 33 21 33 29 29 33 29 22 38

3055: 29 33 38 22 21 38 22 38 38 25 21 24 33 33 21 38 25 21 38 37 21 21 24 38 25 21 21 38 33 21 21  0]xS 

3056: 4382 4994 M (dic)[38 21  0]xS 

3057: 4474 4994 M (-)S 

3058: 2525 5080 M (at)[33  0]xS 

3059: 2579 5080 M (or classes, we used 64 classes.  We )[38 25 28 33 21 33 29 29 33 29 19 28 53 33 28 38 29 33 38 28 38 38 28 33 21 33 29 29 33 29 19 28

3060: 28 70 33  0]xS 

3061: 3720 5080 M (linearly )[21 21 38 33 33 25 21 36  0]xS 

3062: 3975 5080 M (interpolated each )[21 38 21 33 25 38 38 21 33 21 33 38 27 33 33 33 38  0]xS 

3063: 2525 5166 M (max)[57 33  0]xS 

3064: 2652 5166 M (imum ent)[21 57 38 57 31 33 38  0]xS 

3065: 2948 5166 M (ropy model with a trigram model, to smooth and )[25 38 38 36 31 57 38 38 33 21 31 53 21 21 38 31 33 31 21 25 21 37 25 33 57 31 57 38 38 33 21 19

3066: 31 21 38 31 29 57 38 38 21 38 30 33 38 38  0]xS 

3067: 2525 5253 M (a)S 

3068: 2558 5253 M (void zero probabilities.  )[37 38 21 38 29 33 33 25 38 29 38 25 38 38 33 38 21 21 21 21 21 33 29 19 28  0]xS 

3069: 3331 5253 M (O)S 

3070: 3385 5253 M (ur technique )[38 25 28 21 33 33 38 38 21 38 38 33  0]xS 

3071: 3797 5253 M (interpolated )[21 38 21 33 25 38 38 21 33 21 33 38  0]xS 

3072: 4185 5253 M (with a tri)[53 21 21 38 28 33 28 21 25  0]xS 

3073: 4474 5253 M (-)S 

3074: 2525 5339 M (gram model reduced overall perplexity from 1% to 5% versus a )

3075: [37 25 33 57 26 57 38 38 33 21 26 25 33 38 38 33 33 38 26 38 37 33 25 33 21 21 26 38 33 25 38 21

3076: 33 37 21 21 36 26 24 25 38 57 26 38 62 26 21 38 26 38 62 26 37 33 25 29 38 29 25 33  0]xS 

3077: 2525 5425 M (max)[57 33  0]xS 

3078: 2652 5425 M (imum)[21 57 38  0]xS 

3079: 2825 5425 M ( entropy model without our technique )[21 33 38 21 25 38 38 36 20 57 38 38 33 21 20 53 21 21 38 38 38 21 20 38 38 25 20 21 33 33 38 38

3080: 21 38 38 33  0]xS 

3081: 3986 5425 M (interpolated)[21 38 21 33 25 38 38 21 33 21 33  0]xS 

3082: 4346 5425 M ( with )[20 53 21 21 38  0]xS 

3083: 2525 5511 M (the same trigram mod)[21 38 33 33 29 33 57 33 32 21 25 21 37 25 33 57 32 57 38  0]xS 

3084: 3218 5511 M (el; we were not able to run the base)[33 21 21 32 53 33 32 53 33 25 33 32 38 38 21 32 33 38 21 33 32 21 38 32 25 38 38 32 21 38 33 32

3085: 38 33 29  0]xS 

3086: 4386 5511 M (line )[21 21 38 33  0]xS 

3087: 2525 5598 M (per)[38 33  0]xS 

3088: 2621 5598 M (plexity at 10,000,000 words, because the version without our )

3089: [38 21 33 37 21 21 36 24 33 21 24 38 38 19 38 38 38 19 38 38 38 24 53 38 25 38 29 19 24 38 33 33

3090: 33 38 29 33 24 21 38 33 24 37 33 25 29 21 38 38 24 53 21 21 38 38 38 21 24 38 38 25  0]xS 

3091: 2525 5684 M (speed)[29 38 33 33  0]xS 

3092: 2696 5684 M (ups was too slow.  We used subsets of Wall Street Journal )[38 38 29 24 53 33 29 24 21 38 38 24 29 21 38 53 19 24 24 70 33 24 38 29 33 38 24 29 38 38 29 33

3093: 21 29 24 38 24 24 70 33 21 21 24 42 21 25 33 33 21 24 29 38 38 25 38 33 21  0]xS 

3094: LH

3095: (%%[Page: 3]%%) = 

3096: %%PageTrailer

3097: 

3098: %%Page: 4 4

3099: %%PageBoundingBox: 12 12 600 780

3100: %%EndPageComments

3101: %%BeginPageSetup

3102: /DeviceRGB dup setcolorspace /colspABC exch def

3103: mysetup concat colspRefresh

3104: %%EndPageSetup

3105: 

3106: /DeviceGray dup setcolorspace /colspABC exch def

3107: 0 0 scol F3S53 Ji 

3108: 350 6179 M ( )S 

3109: 2150 6179 M ( )S 

3110: F3S4B Ji 

3111: 350 568 M (data, building the classes from scratch at each size, and using the )

3112: [38 33 21 33 19 22 38 38 21 21 38 21 38 37 22 21 38 33 21 33 21 33 29 29 33 29 21 24 25 38 57 21

3113: 29 33 25 33 21 33 38 21 33 21 21 33 33 33 38 21 29 21 33 33 19 21 33 38 38 21 38 29 21 38 37 21

3114: 21 38 33  0]xS 

3115: 350 654 M (60,000 most common w)[38 38 19 38 38 38 24 57 38 29 21 24 33 38 57 57 38 38 24  0]xS 

3116: 1090 654 M (ords in the training data, or all words, if )[38 25 38 29 24 21 38 24 21 38 33 24 21 25 33 21 38 21 38 37 24 38 33 21 33 19 24 38 25 24 33 21

3117: 21 23 53 38 25 38 29 19 23 21 24  0]xS 

3118: 350 740 M (there were fewer than 60,000 unique words in the training data.)

3119: [21 38 33 25 33 19 53 33 25 33 19 24 33 53 33 25 19 21 38 33 38 19 38 38 19 38 38 38 19 38 38 21

3120: 38 38 33 19 53 38 25 38 29 19 21 38 19 21 38 33 19 21 25 33 21 38 21 38 37 19 38 33 21 33  0]xS 

3121: 2259 740 M ( )S 

3122: : N 384 793 1944 1427 rp C 

3123: 1 0 scol  L ; solid N 2328 792 M 383 792 I 383 2220 I 2328 2220 I C 

3124: K 

3125: N 696 849 M 2152 849 I 2152 1985 I 696 1985 I 696 849 I C 

3126: : N 696 845 M 696 1989 I 2152 1989 I 2152 845 I C 

3127: 701 1642 M 701 1984 I 1668 1984 I 1668 1642 I C 

3128: eoclip N 696 1701 M 2152 1701 I K 

3129: ; : 696 1417 1456 1 rc N 696 1417 M 2152 1417 I K 

3130: ; : 696 1133 1456 1 rc N 696 1133 M 2152 1133 I K 

3131: ; : 696 849 1456 1 rc N 696 849 M 2152 849 I K 

3132: ; 0 Lj 2 Lc 8 Lw solid N 696 849 M 2152 849 I : [ 1 0 0 1 350 759 ] concat K 

3133: ; N 2152 849 M 2152 1985 I : [ 1 0 0 1 350 759 ] concat K 

3134: ; : N 350 759 M 350 2253 I 2362 2253 I 2362 759 I C 

3135: 701 1642 M 701 1984 I 1668 1984 I 1668 1642 I C 

3136: eoclip N 2152 1985 M 696 1985 I : [ 1 0 0 1 350 759 ] concat K 

3137: ; ; : N 350 759 M 350 2253 I 2362 2253 I 2362 759 I C 

3138: 701 1642 M 701 1984 I 1668 1984 I 1668 1642 I C 

3139: eoclip N 696 1985 M 696 849 I : [ 1 0 0 1 350 759 ] concat K 

3140: ; ; 1 Lw solid N 696 849 M 696 1985 I K 

3141: N 676 1985 M 696 1985 I K 

3142: N 676 1701 M 696 1701 I K 

3143: N 676 1417 M 696 1417 I K 

3144: N 676 1133 M 696 1133 I K 

3145: N 676 849 M 696 849 I K 

3146: N 696 1985 M 2152 1985 I K 

3147: N 696 2005 M 696 1985 I K 

3148: N 1181 2005 M 1181 1985 I K 

3149: N 1667 2005 M 1667 1985 I K 

3150: N 2152 2005 M 2152 1985 I K 

3151: 0.055 0 scol 8 Lw solid : 675 1058 582 151 rc N 696 1133 M 1181 1133 I : [ 1 0 0 1 350 759 ] concat K 

3152: ; ; N 1181 1133 M 1667 1133 I : [ 1 0 0 1 350 759 ] concat K 

3153: ; N 691 1327 M 747 1333 I 748 1325 I 692 1319 I C 

3154: 0.406 0 scol  O N 771 1338 M 820 1344 I 821 1336 I 772 1330 I C 

3155:  O N 812 1344 M 819 1345 I 820 1337 I 813 1336 I C 

3156:  O N 851 1349 M 881 1353 I 882 1345 I 852 1341 I C 

3157:  O N 873 1353 M 899 1356 I 900 1348 I 874 1345 I C 

3158:  O N 931 1361 M 941 1362 I 942 1354 I 932 1353 I C 

3159:  O N 933 1362 M 979 1368 I 980 1360 I 934 1354 I C 

3160:  O N 1003 1372 M 1059 1378 I 1060 1370 I 1004 1364 I C 

3161:  O N 1083 1380 M 1123 1383 I 1124 1375 I 1084 1372 I C 

3162:  O N 1116 1375 M 1132 1375 I 1132 1383 I 1116 1383 I C 

3163:  O N 1155 1384 M 1184 1385 I 1185 1377 I 1156 1376 I C 

3164:  O N 1177 1377 M 1215 1377 I 1215 1385 I 1177 1385 I C 

3165:  O N 1207 1377 M 1225 1377 I 1225 1385 I 1207 1385 I C 

3166:  O N 1248 1375 M 1275 1374 I 1276 1382 I 1249 1383 I C 

3167:  O N 1267 1374 M 1296 1372 I 1297 1380 I 1268 1382 I C 

3168:  O N 1320 1370 M 1366 1366 I 1367 1374 I 1321 1378 I C 

3169:  O N 1358 1366 M 1368 1365 I 1369 1373 I 1359 1374 I C 

3170:  O N 1400 1361 M 1427 1358 I 1428 1366 I 1401 1369 I C 

3171:  O N 1419 1358 M 1448 1355 I 1449 1363 I 1420 1366 I C 

3172:  O N 1480 1351 M 1488 1350 I 1489 1358 I 1481 1359 I C 

3173:  O N 1480 1350 M 1528 1345 I 1529 1353 I 1481 1358 I C 

3174:  O N 1552 1342 M 1608 1336 I 1609 1344 I 1553 1350 I C 

3175:  O N 1632 1334 M 1670 1332 I 1671 1340 I 1633 1342 I C 

3176:  O N 1662 1332 M 1718 1330 I 1719 1338 I 1663 1340 I C 

3177:  O N 1742 1329 M 1791 1328 I 1792 1336 I 1743 1337 I C 

3178:  O N 1784 1328 M 1791 1328 I 1791 1336 I 1784 1336 I C 

3179:  O N 1823 1328 M 1853 1328 I 1853 1336 I 1823 1336 I C 

3180:  O N 1845 1328 M 1871 1328 I 1871 1336 I 1845 1336 I C 

3181:  O N 1903 1328 M 1914 1328 I 1914 1336 I 1903 1336 I C 

3182:  O N 1905 1336 M 1950 1337 I 1951 1329 I 1906 1328 I C 

3183:  O N 1982 1337 M 2034 1338 I 2035 1330 I 1983 1329 I C 

3184:  O N 2027 1330 M 2031 1330 I 2031 1338 I 2027 1338 I C 

3185:  O N 2063 1330 M 2119 1330 I 2119 1338 I 2063 1338 I C 

3186:  O N 2151 1330 M 2156 1330 I 2156 1338 I 2151 1338 I C 

3187:  O N 691 1292 M 745 1308 I 746 1300 I 692 1284 I C 

3188: 0.453 0 scol  O N 767 1318 M 783 1322 I 784 1314 I 768 1310 I C 

3189:  O N 813 1332 M 820 1334 I 821 1326 I 814 1324 I C 

3190:  O N 812 1334 M 859 1350 I 860 1342 I 813 1326 I C 

3191:  O N 889 1360 M 905 1365 I 906 1357 I 890 1352 I C 

3192:  O N 935 1375 M 941 1377 I 942 1369 I 936 1367 I C 

3193:  O N 933 1377 M 982 1391 I 983 1383 I 934 1369 I C 

3194:  O N 1005 1400 M 1020 1405 I 1021 1397 I 1006 1392 I C 

3195:  O N 1051 1414 M 1063 1418 I 1064 1410 I 1052 1406 I C 

3196:  O N 1055 1418 M 1097 1430 I 1098 1422 I 1056 1410 I C 

3197:  O N 1120 1438 M 1135 1442 I 1136 1434 I 1121 1430 I C 

3198:  O N 1167 1450 M 1184 1455 I 1185 1447 I 1168 1442 I C 

3199:  O N 1176 1455 M 1230 1468 I 1231 1460 I 1177 1447 I C 

3200:  O N 1254 1476 M 1270 1479 I 1271 1471 I 1255 1468 I C 

3201:  O N 1301 1486 M 1305 1487 I 1306 1479 I 1302 1478 I C 

3202:  O N 1297 1487 M 1348 1498 I 1349 1490 I 1298 1479 I C 

3203:  O N 1372 1505 M 1388 1508 I 1389 1500 I 1373 1497 I C 

3204:  O N 1419 1514 M 1427 1515 I 1428 1507 I 1420 1506 I C 

3205:  O N 1419 1515 M 1466 1525 I 1467 1517 I 1420 1507 I C 

3206:  O N 1498 1531 M 1514 1535 I 1515 1527 I 1499 1523 I C 

3207:  O N 1545 1541 M 1549 1542 I 1550 1534 I 1546 1533 I C 

3208:  O N 1541 1542 M 1592 1553 I 1593 1545 I 1542 1534 I C 

3209:  O N 1624 1559 M 1639 1563 I 1640 1555 I 1625 1551 I C 

3210:  O N 1662 1569 M 1717 1581 I 1718 1573 I 1663 1561 I C 

3211:  O N 1748 1588 M 1764 1591 I 1765 1583 I 1749 1580 I C 

3212:  O N 1795 1598 M 1850 1610 I 1851 1602 I 1796 1590 I C 

3213:  O N 1882 1617 M 1897 1621 I 1898 1613 I 1883 1609 I C 

3214:  O N 1921 1627 M 1976 1640 I 1977 1632 I 1922 1619 I C 

3215:  O N 2007 1646 M 2023 1650 I 2024 1642 I 2008 1638 I C 

3216:  O N 2054 1657 M 2109 1669 I 2110 1661 I 2055 1649 I C 

3217:  O N 2140 1676 M 2155 1679 I 2156 1671 I 2141 1668 I C 

3218:  O 0.203 0 scol : 675 1190 582 343 rc N 696 1265 M 817 1314 I 938 1363 I 1060 1411 I 1120 1434 I 1181 1457 I : [ 1 0 0 1 350 759 ] concat K 

3219: ; ; N 1181 1457 M 1302 1501 I 1424 1543 I 1546 1584 I 1667 1624 I : [ 1 0 0 1 350 759 ] concat K 

3220: ; : 1592 1549 582 301 rc N 1667 1624 M 1788 1663 I 1910 1700 I 2031 1737 I 2152 1774 I : [ 1 0 0 1 350 759 ] concat K 

3221: ; ; 0.055 0 scol : 675 1037 118 193 rc N 696 1112 M 717 1133 I 696 1154 I 675 1133 I 696 1112 I C 

3222: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3223: ; ; N 1181 1112 M 1202 1133 I 1181 1154 I 1160 1133 I 1181 1112 I C 

3224: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3225: ; N 1667 1112 M 1688 1133 I 1667 1154 I 1646 1133 I 1667 1112 I C 

3226: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3227: ; 0.406 0 scol : 675 1227 117 192 rc N 716 1302 M 675 1302 I 675 1343 I 716 1343 I C 

3228: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3229: ; ; N 1201 1360 M 1160 1360 I 1160 1401 I 1201 1401 I C 

3230: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3231: ; N 1687 1315 M 1646 1315 I 1646 1356 I 1687 1356 I C 

3232: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3233: ; : 2056 1238 118 192 rc N 2172 1313 M 2131 1313 I 2131 1354 I 2172 1354 I C 

3234: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3235: ; ; 0.453 0 scol : 675 1192 118 193 rc N 696 1267 M 717 1309 I 675 1309 I 696 1267 I C 

3236: : 0.891 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3237: ; ; N 1181 1430 M 1202 1472 I 1160 1472 I 1181 1430 I C 

3238: : 0.891 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3239: ; N 1667 1544 M 1688 1586 I 1646 1586 I 1667 1544 I C 

3240: : 0.891 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3241: ; : 2056 1579 118 193 rc N 2152 1654 M 2173 1696 I 2131 1696 I 2152 1654 I C 

3242: : 0.891 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3243: ; ; 0.203 0 scol : 675 1169 97 172 rc N 696 1265 M 675 1244 I : [ 1 0 0 1 350 759 ] concat K 

3244: ; ; : 675 1190 118 172 rc N 696 1265 M 717 1286 I : [ 1 0 0 1 350 759 ] concat K 

3245: ; ; : 675 1190 97 172 rc N 696 1265 M 675 1286 I : [ 1 0 0 1 350 759 ] concat K 

3246: ; ; : 675 1169 118 172 rc N 696 1265 M 717 1244 I : [ 1 0 0 1 350 759 ] concat K 

3247: ; ; : 675 1169 97 172 rc N 696 1265 M 696 1244 I : [ 1 0 0 1 350 759 ] concat K 

3248: ; ; : 675 1190 97 172 rc N 696 1265 M 696 1286 I : [ 1 0 0 1 350 759 ] concat K 

3249: ; ; N 1181 1457 M 1160 1436 I : [ 1 0 0 1 350 759 ] concat K 

3250: ; N 1181 1457 M 1202 1478 I : [ 1 0 0 1 350 759 ] concat K 

3251: ; N 1181 1457 M 1160 1478 I : [ 1 0 0 1 350 759 ] concat K 

3252: ; N 1181 1457 M 1202 1436 I : [ 1 0 0 1 350 759 ] concat K 

3253: ; N 1181 1457 M 1181 1436 I : [ 1 0 0 1 350 759 ] concat K 

3254: ; N 1181 1457 M 1181 1478 I : [ 1 0 0 1 350 759 ] concat K 

3255: ; N 1667 1624 M 1646 1603 I : [ 1 0 0 1 350 759 ] concat K 

3256: ; N 1667 1624 M 1688 1645 I : [ 1 0 0 1 350 759 ] concat K 

3257: ; N 1667 1624 M 1646 1645 I : [ 1 0 0 1 350 759 ] concat K 

3258: ; N 1667 1624 M 1688 1603 I : [ 1 0 0 1 350 759 ] concat K 

3259: ; N 1667 1624 M 1667 1603 I : [ 1 0 0 1 350 759 ] concat K 

3260: ; N 1667 1624 M 1667 1645 I : [ 1 0 0 1 350 759 ] concat K 

3261: ; : 2056 1678 118 172 rc N 2152 1774 M 2131 1753 I : [ 1 0 0 1 350 759 ] concat K 

3262: ; ; : 2077 1699 97 172 rc N 2152 1774 M 2173 1795 I : [ 1 0 0 1 350 759 ] concat K 

3263: ; ; : 2056 1699 118 172 rc N 2152 1774 M 2131 1795 I : [ 1 0 0 1 350 759 ] concat K 

3264: ; ; : 2077 1678 97 172 rc N 2152 1774 M 2173 1753 I : [ 1 0 0 1 350 759 ] concat K 

3265: ; N 2152 1774 M 2152 1753 I : [ 1 0 0 1 350 759 ] concat K 

3266: ; ; : 2077 1699 97 172 rc N 2152 1774 M 2152 1795 I : [ 1 0 0 1 350 759 ] concat K 

3267: ; ; 0 0 scol 

3268: %%IncludeResource: font Helvetica

3269: F /F8 0 /256 T /Helvetica mF 

3270: /F8S43 F8 [67 0 0 -67 0 0 ] mFS

3271: F8S43 Ji 

3272: 479 2005 M (0.001)[37 19 37 37  0]xS 

3273: 516 1721 M (0.01)[37 19 37  0]xS 

3274: 553 1437 M (0.1)[37 19  0]xS 

3275: 609 1153 M (1)S 

3276: 572 869 M (10)[37  0]xS 

3277: 594 2102 M (10,000)[37 37 19 37 37  0]xS 

3278: 1061 2102 M (100,000)[37 37 37 19 37 37  0]xS 

3279: 1519 2102 M (1,000,000)[37 19 37 37 37 19 37 37  0]xS 

3280: 1985 2102 M (10,000,000)[37 37 19 37 37 37 19 37 37  0]xS 

3281: 

3282: %%IncludeResource: font Helvetica-Bold

3283: F /F9 0 /256 T /Helvetica-Bold mF 

3284: /F9S43 F9 [67 0 0 -67 0 0 ] mFS

3285: F9S43 Ji 

3286: 1219 2180 M (Training size)[40 26 37 19 41 19 41 41 19 37 19 34  0]xS 

3287: /F9S00IFFFFFFBD F9 [0 -67 -67 0 0 0 ] mFS

3288: F9S00IFFFFFFBD Ji 

3289: 461 1819 M (Time\(relative to baseline\))[-41 -19 -60 -37 -22 -26 -37 -19 -37 -22 -19 -37 -37 -19 -22 -41 -19 -41 -37 -37 -37 -19 -19 -41 -37  0]yS 

3290: 16 Lw N 1667 1642 M 701 1642 I 701 1984 I 1667 1984 I C 

3291: : 0.754 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3292: ; 0.055 0 scol 8 Lw : 710 1651 357 118 rc N 741 1693 M 991 1693 I : [ 1 0 0 1 350 759 ] concat K 

3293: ; ; : 770 1651 193 139 rc N 866 1672 M 887 1693 I 866 1714 I 845 1693 I 866 1672 I C 

3294: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3295: ; ; 0 0 scol : 1009 1651 493 86 rc /F8S4D F8 [77 0 0 -77 0 0 ] mFS

3296: F8S4D Ji 

3297: 1009 1720 M (baseline \(GIS\))[43 43 39 43 17 17 43 43 21 26 60 21 51  0]xS 

3298: ; N 737 1759 M 793 1759 I 793 1767 I 737 1767 I C 

3299: 0.406 0 scol  O N 825 1759 M 881 1759 I 881 1767 I 825 1767 I C 

3300:  O N 913 1759 M 969 1759 I 969 1767 I 913 1767 I C 

3301:  O N 886 1742 M 845 1742 I 845 1783 I 886 1783 I C 

3302: :  O ; : [ 1 0 0 1 350 759 ] concat K 

3303: ; 0 0 scol F8S4D Ji 

3304: 1009 1790 M (unigram caching)[43 43 17 43 26 43 65 21 39 43 39 43 17 43  0]xS 

3305: N 737 1829 M 793 1829 I 793 1837 I 737 1837 I C 

3306: 0.453 0 scol  O N 825 1829 M 841 1829 I 841 1837 I 825 1837 I C 

3307:  O N 873 1829 M 929 1829 I 929 1837 I 873 1837 I C 

3308:  O N 961 1829 M 977 1829 I 977 1837 I 961 1837 I C 

3309:  O N 866 1812 M 887 1854 I 845 1854 I 866 1812 I C 

3310: : 0.891 0 scol  O ; : [ 1 0 0 1 350 759 ] concat K 

3311: ; 0 0 scol 1009 1860 M (clustered)[39 17 43 39 21 43 26 43  0]xS 

3312: 0.203 0 scol : 710 1828 357 147 rc N 741 1903 M 991 1903 I : [ 1 0 0 1 350 759 ] concat K 

3313: ; ; : 770 1807 172 168 rc N 866 1903 M 845 1882 I : [ 1 0 0 1 350 759 ] concat K 

3314: ; ; : 791 1828 172 147 rc N 866 1903 M 887 1924 I : [ 1 0 0 1 350 759 ] concat K 

3315: ; ; : 770 1828 172 147 rc N 866 1903 M 845 1924 I : [ 1 0 0 1 350 759 ] concat K 

3316: ; ; : 791 1807 172 168 rc N 866 1903 M 887 1882 I : [ 1 0 0 1 350 759 ] concat K 

3317: ; ; : 791 1807 151 168 rc N 866 1903 M 866 1882 I : [ 1 0 0 1 350 759 ] concat K 

3318: ; ; : 791 1828 151 147 rc N 866 1903 M 866 1924 I : [ 1 0 0 1 350 759 ] concat K 

3319: ; ; 0 0 scol 1009 1930 M (clustered \( 3 level\))[39 17 43 39 21 43 26 43 43 21 26 21 43 21 17 43 39 43 17  0]xS 

3320: 1 Lw solid N 2328 792 M 383 792 I 383 2220 I 2328 2220 I C 

3321: K 

3322: F3S4B Ji 

3323: 350 2321 M ( )S 

3324: /F4S53 F4 [83 0 0 -83 0 0 ] mFS

3325: F4S53 Ji 

3326: 350 2466 M (Figure )[51 23 42 46 37 37  0]xS 

3327: 607 2466 M (1)S 

3328: 649 2466 M (: Speedup Results)[28 21 46 46 37 37 46 46 46 21 60 37 32 46 23 28  0]xS 

3329: 1281 2466 M ( )S 

3330: F3S4B Ji 

3331: 463 2603 M (Figure )[42 21 37 38 25 33  0]xS 

3332: 692 2603 M (1)S 

3333: 730 2603 M ( shows our results, giving relativ)[33 29 38 38 53 29 32 38 38 25 32 25 33 29 38 21 21 29 19 32 37 21 37 21 38 37 32 25 33 21 33 21

3334: 21  0]xS 

3335: 1776 2603 M (e speeds.  Notice )[33 32 29 38 33 33 38 29 19 32 32 54 38 21 21 33 33  0]xS 

3336: 350 2689 M (that we achieve a speedup of up to a factor of 35 over the )[21 38 33 21 39 53 33 39 33 33 38 21 33 37 33 39 33 39 29 38 33 33 38 38 38 39 38 24 39 38 38 38

3337: 21 38 38 33 38 24 33 33 21 38 25 38 38 24 38 38 38 38 38 37 33 25 38 21 38 33  0]xS 

3338: 350 2775 M (unigram caching result.  We believe this is the largest speedup )

3339: [38 38 21 37 25 33 57 29 33 33 33 38 21 38 37 29 25 33 29 38 21 21 19 29 29 70 33 29 38 33 21 21

3340: 33 37 33 29 21 38 21 29 29 21 29 29 21 38 33 29 21 33 25 37 33 29 21 29 29 38 33 33 38 38 38  0]xS 

3341: 350 2862 M (reported.  Notice also that at the smallest data size, the classing )

3342: [25 33 38 38 25 21 33 38 19 27 26 54 38 21 21 33 33 26 33 21 29 38 26 21 38 33 21 26 33 21 26 21

3343: 38 33 26 29 57 33 21 21 33 29 21 26 38 33 21 33 26 29 21 33 33 19 26 21 38 33 26 33 21 33 29 29

3344: 21 38 37  0]xS 

3345: 350 2948 M (methods actually result in minor slowdowns comp)[57 33 21 38 38 38 29 56 33 33 21 38 33 21 21 36 56 25 33 29 38 21 21 56 21 38 55 57 21 38 38 25

3346: 55 29 21 38 53 38 38 53 38 29 55 33 38 57  0]xS 

3347: 2081 2948 M (ared to )[33 25 33 38 55 21 38  0]xS 

3348: 350 3034 M (unigram caching, but that as the training data size increases, the )

3349: [38 38 21 37 25 33 57 25 33 33 33 38 21 38 37 19 25 38 38 21 25 21 38 33 21 25 33 29 25 21 38 33

3350: 25 21 25 33 21 38 21 38 37 25 38 33 21 33 25 29 21 33 33 25 21 38 33 25 33 33 29 33 29 19 24 21

3351: 38 33  0]xS 

3352: 350 3120 M (speedup from our technique also rapidly increases.)[29 38 33 33 38 38 38 19 24 25 38 57 19 38 38 25 19 21 33 33 38 38 21 38 38 33 19 33 21 29 38 19

3353: 25 33 38 21 38 21 36 19 21 38 33 25 33 33 29 33 29  0]xS 

3354: 1875 3120 M ( )S 

3355: F4S64 Ji 

3356: 951 3281 M (5.)[50  0]xS 

3357: 1026 3281 M ( )S 

3358: 1086 3281 M (DISCUSSION)[72 39 56 72 72 56 56 39 78  0]xS 

3359: 1698 3281 M ( )S 

3360: F3S4B Ji 

3361: 350 3422 M (We have discussed our speedup technique in the context of )[70 33 41 38 33 37 33 41 38 21 29 33 38 29 29 33 38 41 38 38 25 41 29 38 33 33 38 38 38 41 21 33

3362: 33 38 38 21 38 38 33 41 21 38 41 21 38 33 40 33 38 38 21 33 37 21 40 38 24  0]xS 

3363: 350 3508 M (training.  However, it can in many cases also be used for tes)

3364: [21 25 33 21 38 21 38 37 19 23 23 54 38 53 33 37 33 25 19 23 21 21 22 33 33 38 22 21 38 22 57 33

3365: 38 36 22 33 33 29 33 29 22 33 21 29 38 22 38 33 22 38 29 33 38 22 24 38 25 22 21 33  0]xS 

3366: 2188 3508 M (ting.  )[21 21 38 37 19 19  0]xS 

3367: 350 3594 M (In particular, if in the test situation one needs the probabilities of )

3368: [25 38 22 38 33 25 21 21 33 38 21 33 25 19 22 21 24 21 21 38 21 21 38 33 21 21 33 29 21 21 29 21

3369: 21 38 33 21 21 38 38 21 38 38 33 21 38 33 33 38 29 21 21 38 33 21 38 25 38 38 33 38 21 21 21 21

3370: 21 33 29 21 38 24  0]xS 

3371: 350 3680 M (most or all words in a particular context, our speedup will not be )

3372: [57 38 29 21 22 38 25 22 33 21 21 22 53 38 25 38 29 22 21 38 21 33 21 38 33 25 21 21 33 38 21 33

3373: 25 21 33 38 38 21 33 37 21 19 21 38 38 25 21 29 38 33 33 38 38 38 21 53 21 21 21 21 38 38 21 21

3374: 38 33  0]xS 

3375: 350 3767 M (helpful.  On the other hand, if one were to use maximum entropy )

3376: [38 33 21 38 24 38 21 19 21 21 54 38 21 21 38 33 21 38 21 38 33 25 21 38 33 38 38 19 21 21 24 21

3377: 38 38 33 21 53 33 25 33 21 21 38 21 38 29 33 21 57 33 37 21 57 38 57 20 33 38 21 25 38 38 36  0]xS 

3378: 350 3853 M (models to rescore n)[57 38 38 33 21 29 31 21 38 31 25 33 29 33 38 25 33 31  0]xS 

3379: 972 3853 M (-)S 

3380: 997 3853 M (best lists, the speedup wou)[38 33 29 21 31 21 21 29 21 29 19 31 21 38 33 31 29 38 33 33 38 38 38 31 53 38  0]xS 

3381: 1850 3853 M (ld work just as )[21 38 31 53 38 25 37 30 21 38 29 21 30 33 29  0]xS 

3382: 350 3939 M (well for testing as for training.  In the case of, say, rescoring )

3383: [53 33 21 21 34 24 38 25 33 21 33 29 21 21 38 37 33 33 29 33 24 38 25 33 21 25 33 21 38 21 38 37

3384: 19 33 33 25 38 33 21 38 33 33 33 33 29 33 33 38 24 19 33 29 33 36 19 33 25 33 29 33 38 25 21 38

3385: 37  0]xS 

3386: 350 4025 M (lattices, the speedup would be helpful, as long as the lattices did )

3387: [21 33 21 21 21 33 33 29 19 24 21 38 33 23 29 38 33 33 38 38 38 23 53 38 38 21 38 23 38 33 23 38

3388: 33 21 38 24 38 21 19 23 33 29 23 21 38 38 37 23 33 29 23 21 38 33 23 21 33 21 21 21 33 33 29 23

3389: 38 21 38  0]xS 

3390: 350 4112 M (not allow for too many words in each context.)[38 38 21 19 33 21 21 38 53 19 24 38 25 19 21 38 38 19 57 33 38 36 19 53 38 25 38 29 19 21 38 19

3391: 33 33 33 38 19 33 38 38 21 33 37 21  0]xS 

3392: 1732 4112 M ( )S 

3393: 463 4198 M (Notice that our speedup technique could be applied to a )[54 38 21 21 33 33 40 21 38 33 21 40 38 38 25 40 29 38 33 33 38 38 38 39 21 33 33 38 38 21 38 38

3394: 33 39 33 38 38 21 38 39 38 33 39 33 38 38 21 21 33 38 39 21 38 39 33  0]xS 

3395: 350 4284 M (vari)[37 33 25  0]xS 

3396: 466 4284 M (ety of other problems and to a variety of other learning )[33 21 36 41 38 24 41 38 21 38 33 25 40 38 25 38 38 21 33 57 29 40 33 38 38 40 21 38 40 33 40 37

3397: 33 25 21 33 21 36 40 38 24 40 38 21 38 33 25 40 21 33 33 25 38 21 38 37  0]xS 

3398: 350 4370 M (methods.  In particular, there is nothing in particular specific to )

3399: [57 33 21 38 38 38 29 19 27 27 25 38 26 38 33 25 21 21 33 38 21 33 25 19 26 21 38 33 25 33 26 21

3400: 29 26 38 38 21 38 21 38 37 26 21 38 26 38 33 25 21 21 33 38 21 33 25 26 29 38 33 33 21 24 21 33

3401: 26 21 38  0]xS 

3402: 350 4457 M (language modeling in our speedup technique, except that we are )

3403: [21 33 38 37 38 33 37 33 25 57 38 38 33 21 21 38 37 24 21 38 24 38 38 25 24 29 38 33 33 38 38 38

3404: 24 21 33 33 38 38 21 38 38 33 19 24 33 37 33 33 38 21 24 21 38 33 21 24 53 33 24 33 25 33  0]xS 

3405: 350 4543 M (predicting the probabilities of a large number of outputs \(possible )

3406: [38 25 33 38 21 33 21 21 38 37 19 21 38 33 19 38 25 38 38 33 38 21 21 21 21 21 33 29 19 38 24 19

3407: 33 19 21 33 25 37 33 19 38 38 57 38 33 25 19 38 24 19 38 38 21 38 38 21 29 18 25 38 38 29 29 21

3408: 38 21 33  0]xS 

3409: 350 4629 M (next words\).  Any other problem predicting the probabilities of a )

3410: [38 33 37 21 22 53 38 25 38 29 25 19 22 22 53 38 36 22 38 21 38 33 25 22 38 25 38 38 21 33 57 22

3411: 38 25 33 38 21 33 21 21 38 37 22 21 38 33 21 38 25 38 38 33 38 21 21 21 21 21 33 29 21 38 24 21

3412: 33  0]xS 

3413: 350 4715 M (large number of outputs could benefit from these methods.  )[21 33 25 37 33 46 38 38 57 38 33 25 46 38 24 46 38 38 21 38 38 21 29 46 33 38 38 21 38 45 38 33

3414: 38 33 24 21 21 45 24 25 38 57 45 21 38 33 29 33 45 57 33 21 38 38 38 29 19 19  0]xS 

3415: 350 4802 M (Similarly, there is little that is specific to maximum entropy )

3416: [42 21 57 21 21 33 25 21 36 19 40 21 38 33 25 33 40 21 29 40 21 21 21 21 21 33 40 21 38 33 21 40

3417: 21 29 40 29 38 33 33 21 24 21 33 39 21 38 39 57 33 37 21 57 38 57 39 33 38 21 25 38 38 36  0]xS 

3418: 350 4888 M (models in our technique.  For instance, consider training a neural )

3419: [57 38 38 33 21 29 21 21 38 21 38 38 25 21 21 33 33 38 38 21 38 38 33 19 21 21 42 38 25 21 21 38

3420: 29 21 33 38 33 33 19 20 33 38 38 29 21 38 33 25 20 21 25 33 21 38 21 38 37 20 33 20 38 33 38 25

3421: 33 21  0]xS 

3422: 350 4974 M (n)S 

3423: 388 4974 M (etwork to learn the probabilities of 10,000 outputs.  Each step of )

3424: [33 21 53 38 25 37 19 21 38 19 21 33 33 25 38 19 21 38 33 19 38 25 38 38 33 38 21 21 21 21 21 33

3425: 29 19 38 24 19 38 38 19 38 38 38 19 38 38 21 38 38 21 29 19 18 18 46 33 33 38 18 29 21 33 38 18

3426: 38 24  0]xS 

3427: 350 5060 M (training would require back)[21 25 33 21 38 21 38 37 24 53 38 38 21 38 24 25 33 38 38 21 25 33 24 38 33 33  0]xS 

3428: 1198 5060 M (-)S 

3429: 1223 5060 M (propagating 9,999 zeros, and one 1.  )[38 25 38 38 33 37 33 21 21 38 37 24 38 19 38 38 38 24 33 33 25 38 29 19 24 33 38 38 24 38 38 33

3430: 23 38 19 19  0]xS 

3431: 350 5147 M (Alternatively, one could place the outputs into 100 classes.  A )

3432: [53 21 21 33 25 38 33 21 21 37 33 21 36 19 31 38 38 33 30 33 38 38 21 38 30 38 21 33 33 33 30 21

3433: 38 33 30 38 38 21 38 38 21 29 30 21 38 21 38 30 38 38 38 30 33 21 33 29 29 33 29 19 30 30 53  0]xS 

3434: 350 5233 M (first network could be trained to learn the class)[24 21 25 29 21 37 38 33 21 53 38 25 37 37 33 38 38 21 38 37 38 33 37 21 25 33 21 38 33 38 37 21

3435: 38 37 21 33 33 25 38 36 21 38 33 36 33 21 33 29  0]xS 

3436: 1903 5233 M (-)S 

3437: 1928 5233 M (probabilit)[38 25 38 38 33 38 21 21 21  0]xS 

3438: 2222 5233 M (ies.  )[21 33 29 19 19  0]xS 

3439: 350 5319 M (Each step of training would require back)[46 33 33 38 31 29 21 33 38 31 38 24 31 21 25 33 21 38 21 38 37 31 53 38 38 21 38 31 25 33 38 38

3440: 21 25 33 31 38 33 33  0]xS 

3441: 1645 5319 M (-)S 

3442: 1670 5319 M (propagating 99 zeros )[38 25 38 38 33 37 33 21 21 38 37 31 38 38 30 33 33 25 38 29  0]xS 

3443: 350 5405 M (and one 1.  Next, we would learn 100 neural networks, for )[33 38 38 39 38 38 33 39 38 19 39 39 54 33 37 21 19 39 53 33 39 53 38 38 21 38 39 21 33 33 25 38

3444: 39 38 38 38 39 38 33 38 25 33 21 39 38 33 21 53 38 25 37 29 19 39 24 38 25  0]xS 

3445: 350 5491 M (predicting the probability of the outputs given the class, one )

3446: [38 25 33 38 21 33 21 21 38 37 38 21 38 33 38 38 25 38 38 33 38 21 21 21 21 36 38 38 24 38 21 38

3447: 33 38 38 38 21 38 38 21 29 38 37 21 37 33 38 37 21 38 33 37 33 21 33 29 29 19 37 38 38 33  0]xS 

3448: 350 5578 M (neural network for each class, predicting a probability for each )

3449: [38 33 38 25 33 21 29 38 33 21 53 38 25 37 29 24 38 25 29 33 33 33 38 29 33 21 33 29 29 19 29 38

3450: 25 33 38 21 33 21 21 38 37 29 33 29 38 25 38 38 33 38 21 21 21 21 36 29 24 38 25 29 33 33 33 38

3451:  0]xS 

3452: 350 5664 M (ou)[38  0]xS 

3453: 426 5664 M (tput in that class.  Network )[21 38 38 21 40 21 38 40 21 38 33 21 39 33 21 33 29 29 19 39 39 54 33 21 53 38 25 37  0]xS 

3454: F1S4B Ji 

3455: 1377 5664 M (i)S 

3456: F3S4B Ji 

3457: 1398 5664 M ( would learn the conditional )[39 53 38 38 21 38 39 21 33 33 25 38 39 21 38 33 39 33 38 38 38 21 21 21 38 38 33 21  0]xS 

3458: 2525 568 M (probabilities of outputs in class )[38 25 38 38 33 38 21 21 21 21 21 33 29 34 38 24 34 38 38 21 38 38 21 29 33 21 38 33 33 21 33 29

3459: 29  0]xS 

3460: F1S4B Ji 

3461: 3558 568 M (i)S 

3462: F3S4B Ji 

3463: 3579 568 M ( given that class )[33 37 21 37 33 38 33 21 38 33 21 33 33 21 33 29 29  0]xS 

3464: F1S4B Ji 

3465: 4135 568 M (i)S 

3466: F3S4B Ji 

3467: 4156 568 M ( is correct.  )[33 21 29 33 33 38 25 25 33 33 21 19 19  0]xS 

3468: 2525 654 M (Each step of training would need only train the network )[46 33 33 38 53 29 21 33 38 52 38 24 52 21 25 33 21 38 21 38 37 52 53 38 38 21 38 52 38 33 33 38

3469: 52 38 38 21 36 52 21 25 33 21 38 52 21 38 33 52 38 33 21 53 38 25 37  0]xS 

3470: 2525 740 M (corresponding to the correct class, meaning that again, only 99 )

3471: [33 38 25 25 33 29 38 38 38 38 21 38 37 29 21 38 29 21 38 33 29 33 38 25 25 33 33 21 29 33 21 33

3472: 29 29 19 29 57 33 33 38 21 38 37 29 21 38 33 21 29 33 37 33 21 38 19 29 38 38 21 36 28 38 38  0]xS 

3473: 2525 827 M (zeros and)[33 33 25 38 29 23 33 38  0]xS 

3474: 2815 827 M ( one 1 would need to be back)[23 38 38 33 23 38 23 53 38 38 21 38 23 38 33 33 38 23 21 38 23 38 33 23 38 33 33  0]xS 

3475: 3724 827 M (-)S 

3476: 3749 827 M (propagated.  Presumably )[38 25 38 38 33 37 33 21 33 38 19 22 22 43 25 33 29 38 57 33 38 21 36  0]xS 

3477: 2525 913 M (the number of hidden units of these smaller networks predicting )

3478: [21 38 33 25 38 38 57 38 33 25 25 38 24 25 38 21 38 38 33 38 25 38 38 21 21 29 25 38 24 25 21 38

3479: 33 29 33 25 29 57 33 21 21 33 25 25 38 33 21 53 38 25 37 29 24 38 25 33 38 21 33 21 21 38 37  0]xS 

3480: 2525 999 M (only 100 values \(for the class, or the outputs given the class\) )

3481: [38 38 21 36 33 38 38 38 33 37 33 21 38 33 29 33 25 24 38 25 33 21 38 33 33 33 21 33 29 29 19 33

3482: 38 25 33 21 38 33 33 38 38 21 38 38 21 29 33 37 21 37 33 38 33 21 38 33 32 33 21 33 29 29 25  0]xS 

3483: 2525 1085 M (would also be much smaller than the number of hidden units of a )

3484: [53 38 38 21 38 21 33 21 29 38 21 38 33 21 57 38 33 38 21 29 57 33 21 21 33 25 21 21 38 33 38 21

3485: 21 38 33 21 38 38 57 38 33 25 21 38 24 20 38 21 38 38 33 38 20 38 38 21 21 29 20 38 24 20 33  0]xS 

3486: 2525 1172 M (network for)[38 33 21 53 38 25 37 19 24 38  0]xS 

3487: 2876 1172 M ( predicting 10,000 outputs directly.)[19 38 25 33 38 21 33 21 21 38 37 19 38 38 19 38 38 38 19 38 38 21 38 38 21 29 19 38 21 25 33 33

3488: 21 21 36  0]xS 

3489: 3936 1172 M ( )S 

3490: 2638 1258 M (Similarly, there are at least two ways to train decision trees to )

3491: [42 21 57 21 21 33 25 21 36 19 21 21 38 33 25 33 21 33 25 33 20 33 21 20 21 33 33 29 21 20 21 53

3492: 38 20 53 33 36 29 20 21 38 20 21 25 33 21 38 20 38 33 33 21 29 21 38 38 20 21 25 33 33 29 20 21

3493: 38  0]xS 

3494: 2525 1344 M (handle large numbers of outputs \(train decision trees with many )

3495: [38 33 38 38 21 33 26 21 33 25 37 33 26 38 38 57 38 33 25 29 26 38 24 26 38 38 21 38 38 21 29 25

3496: 25 21 25 33 21 38 25 38 33 33 21 29 21 38 38 25 21 25 33 33 29 25 53 21 21 38 25 57 33 38 36  0]xS 

3497: 2525 1430 M (outputs at each leaf, or train a binary decision tree for each )

3498: [38 38 21 38 38 21 29 38 33 21 38 33 33 33 38 38 21 33 33 24 19 38 38 25 38 21 25 33 21 38 38 33

3499: 38 38 21 38 33 25 36 38 38 33 33 21 29 21 38 38 38 21 25 33 33 38 24 38 25 38 33 33 33 38  0]xS 

3500: 2525 1517 M (possible output and normal)[38 38 29 29 21 38 21 33 30 38 38 21 38 38 21 30 33 38 38 30 38 38 25 57 33  0]xS 

3501: 3377 1517 M (ize\).  Again, in both of these cases, )[21 33 33 25 19 30 29 53 37 33 21 38 19 29 21 38 29 38 38 21 38 29 38 24 29 21 38 33 29 33 29 33

3502: 33 29 33 29 19  0]xS 

3503: 2525 1603 M (our method can be applied.)[38 38 25 19 57 33 21 38 38 38 19 33 33 38 19 38 33 19 33 38 38 21 21 33 38  0]xS 

3504: 3343 1603 M ( )S 

3505: 2638 1689 M (More generally, almost any learning algorithm that is slowed )

3506: [67 38 25 33 23 37 33 38 33 25 33 21 21 36 19 23 33 21 57 38 29 21 23 33 38 36 23 21 33 33 25 38

3507: 21 38 37 23 33 21 37 38 25 21 21 38 57 23 21 38 33 21 23 21 29 23 29 21 38 53 33 38  0]xS 

3508: 2525 1775 M (at training time when there are a large number of outputs could )

3509: [33 21 27 21 25 33 21 38 21 38 37 26 21 21 57 33 26 53 38 33 38 26 21 38 33 25 33 26 33 25 33 26

3510: 33 26 21 33 25 37 33 26 38 38 57 38 33 25 26 38 24 26 38 38 21 38 38 21 29 26 33 38 38 21 38  0]xS 

3511: 2525 1862 M (benefit from our approach.  Similarly, any algorithm slowed at )

3512: [38 33 38 33 24 21 21 30 24 25 38 57 30 38 38 25 30 33 38 38 25 38 33 33 38 19 30 30 42 21 57 21

3513: 21 33 25 21 36 19 29 33 38 36 29 33 21 37 38 25 21 21 38 57 29 29 21 38 53 33 38 29 33 21  0]xS 

3514: 2525 1948 M (tes)[21 33  0]xS 

3515: 2608 1948 M (t time by a large number of outputs, but used in a situation in )

3516: [21 24 21 21 57 33 24 38 36 24 33 24 21 33 25 37 33 23 38 38 57 38 33 25 23 38 24 23 38 38 21 38

3517: 38 21 29 19 23 38 38 21 23 38 29 33 38 23 21 38 23 33 23 29 21 21 38 33 21 21 38 38 23 21 38  0]xS 

3518: 2525 2034 M (which only a few of those outputs are needed, would benefit.)[53 38 21 33 38 19 38 38 21 36 19 33 19 24 33 53 19 38 24 19 21 38 38 29 33 19 38 38 21 38 38 21

3519: 29 19 33 25 33 19 38 33 33 38 33 38 19 19 53 38 38 21 38 19 38 33 38 33 24 21 21  0]xS 

3520: 4356 2034 M ( )S 

3521: 2638 2120 M (Our technique is an extremely promising one.  Although it is )

3522: [54 38 25 23 21 33 33 38 38 21 38 38 33 23 21 29 23 33 38 23 33 37 21 25 33 57 33 21 36 23 38 25

3523: 38 57 21 29 21 38 37 23 38 38 33 19 22 22 53 21 21 38 38 38 37 38 22 21 21 22 21 29  0]xS 

3524: 2525 2207 M (only an approximation, rather than an exact technique, it is one )

3525: [38 38 21 36 27 33 38 27 33 38 38 25 38 37 21 57 33 21 21 38 38 19 27 25 33 21 38 33 25 27 21 38

3526: 33 38 27 33 38 27 33 37 33 33 21 27 21 33 33 38 38 21 38 38 33 19 26 21 21 26 21 29 26 38 38 33

3527:  0]xS 

3528: 2525 2293 M (that )[21 38 33 21  0]xS 

3529: 2660 2293 M (both theoretically and empirically reduces perplexity; it adds )

3530: [38 38 21 38 22 21 38 33 38 25 33 21 21 33 33 21 21 36 22 33 38 38 21 33 57 38 21 25 21 33 33 21

3531: 21 36 21 25 33 38 38 33 33 29 21 38 33 25 38 21 33 37 21 21 36 21 21 21 21 21 33 38 38 29  0]xS 

3532: 2525 2379 M (very little complexity to coding; it leads to perhaps the largest )

3533: [37 33 25 36 31 21 21 21 21 21 33 31 33 38 57 38 21 33 37 21 21 36 31 21 38 31 33 38 38 21 38 37

3534: 21 31 21 21 30 21 33 33 38 29 30 21 38 30 38 33 25 38 33 38 29 30 21 38 33 30 21 33 25 37 33 29

3535: 21  0]xS 

3536: 2525 2465 M (reported speedups )[25 33 38 38 25 21 33 38 33 29 38 33 33 38 38 38 29  0]xS 

3537: 3118 2465 M <96>S 

3538: 3156 2465 M ( a factor of 35; these speedups are largest )[33 33 32 24 33 33 21 38 25 32 38 24 32 38 38 21 32 21 38 33 29 33 32 29 38 33 33 38 38 38 29 32

3539: 33 25 33 32 21 33 25 37 33 29 21  0]xS 

3540: 2525 2552 M (when needed most, on large, complex problems; it can be applied )

3541: [53 38 33 38 20 38 33 33 38 33 38 20 57 38 29 21 19 20 38 38 20 21 33 25 37 33 19 20 33 38 57 38

3542: 21 33 37 19 38 25 38 38 21 33 57 29 21 19 21 21 19 33 33 38 19 38 33 19 33 38 38 21 21 33 38  0]xS 

3543: 2525 2638 M (independently of the form of the model; and it can be applied )

3544: [21 38 38 33 38 33 38 38 33 38 21 21 36 31 38 24 31 21 38 33 31 24 38 25 57 31 38 24 31 21 38 33

3545: 31 57 38 38 33 21 21 31 33 38 38 31 21 21 31 33 33 38 31 38 33 30 33 38 38 21 21 33 38  0]xS 

3546: 2525 2724 M (both to other learning algorithms and to other problem domains.   )

3547: [38 38 21 38 24 21 38 24 38 21 38 33 25 24 21 33 33 25 38 21 38 37 24 33 21 37 38 25 21 21 38 57

3548: 29 24 33 38 38 24 21 38 23 38 21 38 33 25 23 38 25 38 38 21 33 57 23 38 38 57 33 21 38 29 19 19

3549: 19  0]xS 

3550: 2525 2810 M (We are hopeful that others will use it, both for maximum entropy )

3551: [70 33 21 33 25 33 21 38 38 38 33 24 38 21 21 21 38 33 21 21 38 21 38 33 25 29 21 53 21 21 21 20

3552: 38 29 33 20 21 21 19 20 38 38 21 38 20 24 38 25 20 57 33 37 21 57 38 57 20 33 38 21 25 38 38 36

3553:  0]xS 

3554: 2525 2897 M (modeling applied to language modeling and in many other fields.)

3555: [57 38 38 33 21 21 38 37 19 33 38 38 21 21 33 38 19 21 38 19 21 33 38 37 38 33 37 33 19 57 38 38

3556: 33 21 21 38 37 19 33 38 38 19 21 38 19 57 33 38 36 19 38 21 38 33 25 19 24 21 33 21 38 29  0]xS 

3557: 4485 2897 M ( )S 

3558: F4S64 Ji 

3559: 3096 3057 M (6.)[50  0]xS 

3560: 3171 3057 M ( )S 

3561: 3231 3057 M (REFERENCES)[72 67 60 67 72 67 72 72 67  0]xS 

3562: 3903 3057 M ( )S 

3563: F3S4B Ji 

3564: 2525 3198 M ([1])[25 37  0]xS 

3565: 2612 3198 M ( )S 

3566: 2675 3198 M (Darroch, J. N. and Ratcliff, D \(1972\). Generalized Iterative )

3567: [54 33 25 25 38 33 38 19 22 29 19 22 54 19 22 33 38 38 22 50 33 21 33 21 21 24 24 19 22 54 22 25

3568: 38 38 38 38 25 19 21 53 33 38 33 25 33 21 21 33 33 38 21 25 21 33 25 33 21 21 37 33  0]xS 

3569: 2675 3284 M (Scaling for Log)[42 33 33 21 21 38 37 31 24 38 25 31 45 38  0]xS 

3570: 3169 3284 M (-)S 

3571: 3194 3284 M (Linear Models.  )[45 21 38 33 33 25 31 67 38 38 33 21 29 19 31  0]xS 

3572: F1S4B Ji 

3573: 3727 3284 M (Annals of Mathematical )[46 38 38 38 21 29 31 38 21 30 62 38 21 38 33 54 38 21 21 33 38 21  0]xS 

3574: 2675 3370 M (Statistics)[38 21 38 21 21 29 21 21 33  0]xS 

3575: F3S4B Ji 

3576: 2947 3370 M (, No. 43, pp. 1470)[19 19 54 38 19 19 38 38 19 19 38 38 19 19 38 38 38  0]xS 

3577: 3495 3370 M (-)S 

3578: 3520 3370 M (1480. )[38 38 38 38 19  0]xS 

3579: 3710 3370 M ( )S 

3580: 2525 3457 M ([2])[25 37  0]xS 

3581: 2612 3457 M ( )S 

3582: 2675 3457 M (Rosenfeld, Ronald \(1994\).  \223Adaptive Statistical Language )[50 38 29 33 38 24 33 21 38 19 24 50 38 38 33 21 38 24 25 38 38 38 38 25 19 24 23 33 53 38 33 38

3583: 21 21 37 33 23 42 21 33 21 21 29 21 21 33 33 21 23 45 33 38 37 38 33 37 33  0]xS 

3584: 2675 3543 M (Modeling: A Maximum Entropy Approa)[67 38 38 33 21 21 38 37 21 25 53 25 67 33 37 21 57 38 57 24 46 38 21 25 38 38 36 24 53 38 38 25

3585: 38  0]xS 

3586: 3917 3543 M (ch,\224 Ph.D. Thesis, )[33 38 19 33 24 43 38 19 54 19 24 45 38 33 29 21 29 19  0]xS 

3587: 2675 3629 M (Carnegie Mellon University, April.)[50 33 25 38 33 37 21 33 19 67 33 21 21 38 38 19 54 38 21 37 33 25 29 21 21 36 19 19 53 38 25 21

3588: 21  0]xS 

3589: 3731 3629 M ( )S 

3590: 2525 3715 M ([3])[25 37  0]xS 

3591: 2612 3715 M ( )S 

3592: 2675 3715 M (Ney, Hermann, Essen, Ute and Kneser, Reinhard, \(1994\) )[54 33 36 19 32 54 33 25 57 33 38 38 19 32 46 29 29 33 38 19 32 54 21 33 32 33 38 38 32 54 38 33

3593: 29 33 25 19 31 50 33 21 38 38 33 25 38 19 31 25 38 38 38 38 25  0]xS 

3594: 2675 3801 M (\223On Structuring Probabilistic Dependencies in Stochastic )[33 54 38 33 42 21 25 38 33 21 38 25 21 38 37 33 43 25 38 38 33 38 21 21 21 29 21 21 33 33 54 33

3595: 38 33 38 38 33 38 33 21 33 29 33 21 38 33 42 21 38 33 38 33 29 21 21 33  0]xS 

3596: 2675 3888 M (Language Modeling,\224 in )[45 33 38 37 38 33 37 33 23 67 38 38 33 21 21 38 37 19 33 22 21 38  0]xS 

3597: F1S4B Ji 

3598: 3440 3888 M (Computer, Speech and Language)[50 38 54 38 38 21 33 29 19 22 38 38 33 33 33 38 22 38 38 38 22 42 38 38 38 38 38 38  0]xS 

3599: F3S4B Ji 

3600: 4456 3888 M (, )[19  0]xS 

3601: 2675 3974 M (Vol. 8, pp. 1)[55 38 21 19 19 38 19 19 38 38 19 19  0]xS 

3602: 3055 3974 M (-)S 

3603: 3080 3974 M (38.)[38 38  0]xS 

3604: 3175 3974 M ( )S 

3605: 2525 4060 M ([4])[25 37  0]xS 

3606: 2612 4060 M ( )S 

3607: 2675 4060 M (Brown, Peter F., Pe)[50 25 38 53 38 19 42 43 33 21 33 25 42 42 19 19 42 43  0]xS 

3608: 3335 4060 M (ter V. deSouza, Robert L. Mercer, )[21 33 25 42 55 19 42 38 33 42 38 38 33 33 19 42 50 38 38 33 25 21 42 45 19 41 67 33 25 33 33 25

3609: 19  0]xS 

3610: 2675 4146 M (Vincent J. Della Pietra and Jennifer C. Lai, \(1992\), \223Class)[55 21 38 33 33 38 21 22 29 19 22 54 33 21 21 33 22 43 21 33 21 25 33 22 33 38 38 22 29 33 38 38

3611: 21 24 33 25 22 50 19 22 45 33 21 19 22 25 38 38 38 38 25 19 21 33 50 21 33 29  0]xS 

3612: 4450 4146 M (-)S 

3613: 2675 4233 M (Based n)[50 33 29 33 38 109  0]xS 

3614: 3005 4233 M (-)S 

3615: 3030 4233 M (gram Models of Natural Language\224, )[37 25 33 57 109 67 38 38 33 21 29 109 38 24 108 54 33 21 38 25 33 21 108 45 33 38 37 38 33 37 33 33

3616: 19  0]xS 

3617: F1S4B Ji 

3618: 2675 4319 M (Computational Linguistics)[50 38 54 38 38 21 38 21 21 38 38 38 21 19 42 21 38 38 38 21 29 21 21 33  0]xS 

3619: F3S4B Ji 

3620: 3479 4319 M (, December, Vol. 18, No. 4.)[19 19 54 33 33 33 57 38 33 25 19 19 55 38 21 19 19 38 38 19 19 54 38 19 19 38  0]xS 

3621: 4314 4319 M ( )S 

3622: 2525 4405 M ([5])[25 37  0]xS 

3623: 2612 4405 M ( )S 

3624: 2675 4405 M (Goodman, Joshua, \(2000\), \223)[53 38 38 38 57 33 38 19 78 29 38 29 38 38 33 19 78 25 38 38 38 38 25 19 78  0]xS 

3625: 0.105 0 scol 3701 4405 M (Language Model Size )[45 33 38 37 38 33 37 33 77 67 38 38 33 21 77 42 21 33 33  0]xS 

3626: 2675 4491 M (Reduction by Pruning and Clustering)[50 33 38 38 33 21 21 38 38 64 38 36 64 43 25 38 38 21 38 37 64 33 38 38 64 50 21 38 29 21 33 25

3627: 21 38  0]xS 

3628: 0 0 scol 3977 4491 M (\224 ICSLP)[33 64 25 50 42 45  0]xS 

3629: 4279 4491 M (-)S 

3630: 4304 4491 M (2000, )[38 38 38 38 19  0]xS 

3631: 2675 4578 M (Beijing, October.)[50 33 21 21 21 38 37 19 19 54 33 21 38 38 33 25  0]xS 

3632: 3195 4578 M ( )S 

3633: 2525 4664 M ([6])[25 37  0]xS 

3634: 2612 4664 M ( )S 

3635: 2675 4664 M (Della Pietra, Stephen, Vincent Della Pietra and John )[54 33 21 21 33 51 43 21 33 21 25 33 19 51 42 21 33 38 38 33 38 19 50 55 21 38 33 33 38 21 50 54

3636: 33 21 21 33 50 43 21 33 21 25 33 50 33 38 38 50 29 38 38 38  0]xS 

3637: 2675 4750 M (Lafferty, 1995, \223Inducing Features of Random Fields\224, )[45 33 24 24 33 25 21 36 19 44 38 38 38 38 19 44 33 25 38 38 38 33 21 38 37 44 42 33 33 21 38 25

3638: 33 29 44 38 24 43 50 33 38 38 38 57 43 42 21 33 21 38 29 33 19  0]xS 

3639: 2675 4836 M (CMU Technical Report # C)[50 67 54 19 45 33 33 38 38 21 33 33 21 19 50 33 38 38 25 21 19 38 19  0]xS 

3640: 3510 4836 M (MU)[67  0]xS 

3641: 3631 4836 M (-)S 

3642: 3656 4836 M (CS)[50  0]xS 

3643: 3748 4836 M (-)S 

3644: 3773 4836 M (95)[38  0]xS 

3645: 3849 4836 M (-)S 

3646: 3874 4836 M (144, May.)[38 38 38 19 19 67 33 36  0]xS 

3647: 4181 4836 M ( )S 

3648: 2525 4923 M ([7])[25 37  0]xS 

3649: 2612 4923 M ( )S 

3650: 2675 4923 M (Lafferty, J. and Suhm, B. \(1996\) \223Cluster Expansions and )[45 33 24 24 33 25 21 36 19 27 29 19 27 33 38 38 27 42 38 38 57 19 27 50 19 26 25 38 38 38 38 25

3651: 26 33 50 21 38 29 21 33 25 26 46 37 38 33 38 29 21 38 38 29 26 33 38 38  0]xS 

3652: 2675 5009 M (Iterative Scaling for Maximum Entropy Language Models\224 )[25 21 33 25 33 21 21 37 33 23 42 33 33 21 21 38 37 23 24 38 25 22 67 33 37 21 57 38 57 22 46 38

3653: 21 25 38 38 36 22 45 33 38 37 38 33 37 33 22 67 38 38 33 21 29 33  0]xS 

3654: 2675 5095 M (in )[21 38  0]xS 

3655: F1S4B Ji 

3656: 2785 5095 M (Maximum Entropy and Bayesian Methods, )[62 38 33 21 54 38 54 51 46 38 21 29 38 38 33 51 38 38 38 51 46 38 33 33 29 21 38 38 51 62 33 21

3657: 38 38 38 29 19  0]xS 

3658: F3S4B Ji 

3659: 4251 5095 M (Kluwer )[54 21 38 53 33 25  0]xS 

3660: 2675 5181 M (Academic Publishers.)[53 33 33 38 33 57 21 33 19 43 38 38 21 21 29 38 33 25 29  0]xS 

3661: 3329 5181 M ( )S 

3662: 2525 5268 M ([8])[25 37  0]xS 

3663: 2612 5268 M ( )S 

3664: 2675 5268 M (Wun, J. and Khudanpur, S.  \223Efficient Training Methods )[70 38 38 19 31 29 19 31 33 38 38 30 54 38 38 38 33 38 38 38 25 19 30 42 19 30 30 33 46 24 24 21

3665: 33 21 33 38 21 30 45 25 33 21 38 21 38 37 30 67 33 21 38 38 38 29  0]xS 

3666: 2675 5354 M (f)S 

3667: 2699 5354 M (or Maximum Entropy Language Modeling\224 ICSLP)[38 25 28 67 33 37 21 57 38 57 28 46 38 21 25 38 38 36 28 45 33 38 37 38 33 37 33 28 67 38 38 33

3668: 21 21 38 37 33 28 25 50 42 45  0]xS 

3669: 4279 5354 M (-)S 

3670: 4304 5354 M (2000, )[38 38 38 38 19  0]xS 

3671: 2675 5440 M (Beijing, October.)[50 33 21 21 21 38 37 19 19 54 33 21 38 38 33 25  0]xS 

3672: 3195 5440 M ( )S 

3673: 2525 5526 M ([9])[25 37  0]xS 

3674: 2612 5526 M ( )S 

3675: 2675 5526 M (Goodman, Joshua, \(2000\), \223)[53 38 38 38 57 33 38 19 64 29 38 29 38 38 33 19 64 25 38 38 38 38 25 19 64  0]xS 

3676: 0.105 0 scol 3659 5526 M (Putting it all Together: )[43 38 21 21 21 38 37 63 21 21 63 33 21 21 63 45 38 37 33 21 38 33 25 21  0]xS 

3677: 2675 5613 M (Language Model Combination)[45 33 38 37 38 33 37 33 43 67 38 38 33 21 43 50 38 57 38 21 38 33 21 21 38  0]xS 

3678: 0 0 scol 3645 5613 M (\224 ICASSP)[33 42 25 50 53 42 42  0]xS 

3679: 3975 5613 M (-)S 

3680: 4000 5613 M (2000, Istanbul, )[38 38 38 38 19 42 25 29 21 33 38 38 38 21 19  0]xS 

3681: 2675 5699 M (June.)[29 38 38 33  0]xS 

3682: 2832 5699 M ( )S 

3683: LH

3684: (%%[Page: 4]%%) = 

3685: %%PageTrailer

3686: 

3687: %%Trailer

3688: %%BoundingBox: 12 12 600 780

3689: %%DocumentNeededResources: 

3690: %%+ font Palatino-Bold

3691: %%+ font Times-Italic

3692: %%+ font Palatino-Roman

3693: %%+ font Times-Roman

3694: %%+ font Times-Bold

3695: %%+ font Symbol

3696: %%+ font Courier

3697: %%+ font Helvetica

3698: %%+ font Helvetica-Bold

3699: %%DocumentSuppliedResources: 

3700: %%+ procset Pscript_WinNT_ErrorHandler 5.0 0

3701: %%+ procset Pscript_FatalError 5.0 0

3702: %%+ procset Pscript_Win_Basic 5.0 0

3703: %%+ procset Pscript_Win_Utils_L2 5.0 0

3704: %%+ procset Pscript_Text 5.0 0

3705: %%+ procset Pscript_Encoding256 5.0 0

3706: %%+ procset Pscript_Win_GdiObject 5.0 0

3707: %%+ procset Pscript_Win_GdiObject_L2 5.0 0

3708: %%+ procset Pscript_T42Hdr 5.0 0

3709: Pscript_WinNT_Incr dup /terminate get exec

3710: %%Pages: 4

3711: (%%[LastPage]%%) = 

3712: %%EOF

3713: %-12345X@PJL EOJ

3714: %-12345X