TIIE EVOLUTION OF MACIIINE- I RA(. I AIII, L DIC~TIONARIES 
Cheng-ming (;u(; Chan~-nin,~ lhtan,~ ,h.m #ing (hm,'4 ,/it~ Li 
( ~oiil II!lb21 ~eK;lleC Depal'! lIR:ll! 
Tsiughua Uni~0rsity, lteijing l(X)()l~l (71IINA 
I, hltli~ttut:lion 
Tile uselulness of Machine-Tractable 
Dictionaries (MTDs) ill atlCtilllatic SCllSe taggillg el 
running text ix clearly demonstrated in Tong, el ill., 
(1993). While previous work (Yarowsky, 1992: Gale, el 
al., 1092, 1903) lelies heavily on tile role of statistics, 
TOllg'S sysienl illilkes use of (Ill(: MTD as well as m<~ 
Machine Readable l)iciionarles (MRI)s) in gin example- 
based reasonillg process ill laggint~, sellsC Ilillllbers l() 
llovel words, eompollnds ~,\,ords, gilld pJllHSes in tile 
input text. The hit rate of correcl ,~,Cll:-;c lagging rtlnS ax 
high as 90%. The lewes! IlJl rtll¢ c:tcr iecordcd was 
70%. The sysieill is eotlsidercd a necessaiy inechanisin 
for the construcihm ()1 annotated Chinese \]~#ollilor C.'orlJora 
(SillCILIIC, If) t) I) I1o111 IlllllliIllL (;hilleSe tc',t . 
l'~girlier work (Wilks, et al., 1'090) identified a 
dislii101ion belwcell MRI)s and NqTJ)s. This eal×'r 
recogllizes a further distinction between two types el 
MT1)s> i.e., Type I and Type il MTDs. Type l MTI)s arc 
built ()ll the definition priiniiivcs (fi on0 pgirii0ulal 
MRD wtlerea,,-; Type I1 MTI)s are btlilt ',In the semanli0, 
prmlitives of one |lgiiciellJar ilgiCtlr:.ll IgillpIlage. The MI'D 
functitnling ill T,,mg's sysCenl belongs co die l'ype I 
category. The evolution frlmi Type I t(~ Type I1 MTI)s 
iopiesents all evoitiCiOli ()1 the i)riinitives ti~c MT1)s arc 
constrtlcted o11. Th0 atithors believe that, although it is 
leasible lo derive gi llalllral sol (11 delinilion prilnilixc~ 
fr(lill o11o parlicti\[gir MRD, a descriptive set of Selllantl0 
primitives (Wiiks, 1977) of Olle pariicuhir imlura\] 
ILingilat2e is prefcigibly deri~ed irl)ill inol'e than t)llC 
ila\[tll'aJ s(ltlrc0 Sllch :t.s :1 diction',u'y. This p.os;ilip.n 
rcpresellis gi bgickoll frolll the previous olle Ihgil ctailns 
tile dcrivalion of a natural set el senlanli0 piinlici',es 
lrolll oile partietliar MRI), parth:ulgiri)lhe Jz)#1~,l#1fHl 
I)i¢lionarv (!f ('olllettl\])orarv l','ngli~h (I ~l)O(:t",) ((\]uo. 
1989). 
Three Type l MTi)s 0olnplcJcd al TsiTI!.',hli~ 
(\]lli~,clsity will bc presented. 'l hose include MINI- 
IA)OCI); consii-tie/ed for I,oilgingin l)iciioanries, lhc 
MTi) versiou of 7he Mo&'r. (?hillese Di¢'liOllflr)~ ¢!/" 
Gem, ral ('\[lill~'A'e ('\]ltlraf'lt'rs ( M'\['I)- X I A N'I'()NG) (1 ;il, 
It)leT)> and Ihe MTI) VClSi(lll oJ" 'lTle Mll/liJhm'lional 
I)icliollarv o/ Moderll (.'hillese IVord~ (MTD 
I)\[J()(K)N(;NI",N(~) (l:eng & Zh(lu, I t)~O). 
liflmts to dcri~e natural ~;0ts ~)1 semantic 
prilnilivos lll/nl Typt; I MTDs girt Ihcn des0rihcd These 
inclll(le work (Ill detccting 0Jr0 tl\]ar dcl ill tiOllS witln n t)llC 
parliciilgir MT1), work on colnplllhlg the coinpaiil~iliiy 
I/I word SellSCS beCw0cfi IW() IIW~noJillgtlal di0tiongirics, 
and the most recent step at der{vmg a ilatura\] set o! 
Chincse senlaniic primilives Into1 Typ0 i M'I'Ds (11 
(Jhinese. The work ~as done on ~tii1.4 worksigiiions 
with the (7 eoinF, tlter lallgUgige. 
2. I)qsign ~!i!d (\]onslrueiion o\[Type I IVITDs 
The tteSiOll gilld CO\[lslrtl0lillll of M'I'Ds ill 
general are discussed in detail by \arii)us authors ill 
~tlO (1094). l)escribed hi this papei-al-e three Type 1 
MTDs in two languages. The Iirst ~lne is the 1,:nglish 
MINI-I,I)OC/:, designed alld collstructed by Cheng- 
ruing Guo I\[or I.ongmans. Ihe second tile MTD- 
X\[ANTONG collstructed by nlachine by Xiang Tong 
('l'Ullg, ct ai., 1993), aud tile tllild tile MTD- 
I)UOGONGN\[,:NG by Jullq)ing Gong ((}Ollg & \[ \[uang. 
1991) by halld. 
2. I. MINI-LI)O(~E 
The discussion prcsentcd bcimv represents tl 
gcllera\[ dcsc£iptioi! of file nlost esscllCiZl\[ poinJs in the 
design llllt\] collSllilCliotl o1 MINI I,I)OC\] {. For gi 
detailed Ltclst)tlllt 4)F I{IC plt/ict;{ the lGtdcl is iclcllcd to 
Lilt MINIqJ)OCi'~ ,syMcnl d()cunlcntati,,)ll UpOll 
1 ,ollglll arts' publicatioil. 
As is well klloWli, \[,Oliglllall Dictionary ot 
('.()llJClllpoi0.1} Hnglish (LI)OCI.:,) h;_ls ill (7onlrotled 
Vocabulal b (CV) o\[ aiy, iti{ 2,0()0 wolds. Th0sc 2,0()0 CV 
words tire tlsed i!/ every word sense dclinitk;ll of the 
clltire 13)O(~1'L It was found that Lt stibset el aboul 
4,(XXt C.V word senses arc used to define (hc 2,0(/0 CV 
words lheinselves ((hie, I999). MINI-IJ)OCI{ is a 
diciiomn) of lliese 4,000 dciinitiou ln imiti~es ~i 
IA)OCF.. /n other words, MINI .EliOt't{ hns aboul 
,4_000 eill.ries el word SellSeS. eacll ellilV el ;l Wol'd scilse 
beiiig a dcfiniti(m pritnitive of I J)O(\]l~. 
MINI 1.l)()(71,; differs from IA)OCI:, in the 
lollil ot lilt dl:l'iliililiil text. 'vVIicieas iJ)OCi:, ~ofd 
s(:'nses are, del ined by woids~ MINI -I ,O()CI ,\] dciinitiolls 
;.ire giVCil ill !f~r()i'{.I sellses, i.e., each ~AOltl iil the 
deliniti~m text is already disambiguated, mid has a scnsc 
IItilllbCI tlCtgi.Cciled ;.t\[ tile end, The scC t~i the CV ~old 
senses tised in the definilion text o\[ MINI I.DO(!I:, 
thlls ~.itllili Ihc set of total MI NI-I DOCt < sense chilies. 
When non-(~V words, or phrases 17~rmed either of (IV 
vvoitls, 11Oll-(3V ~A,()IIIS, ()1 a COltibi(laiit)u t)f (,*\ iaud iit)ll- 
(~7 words are lOtllld in MINI-LDOCI:, SellSe 
tteiinititnls, gill of fort is made it) disalnbigu;xtc tile noli- 
(TV word or phrase, arid r0dllce its definition as fl)tilld irl 
I,DOCI{ Ill the definition i)liiniti~cs (if 1,1lOCi:,. 'i'hi.~ 
restll!s ill embedded defh!ilioll, which SO!llelitlles tIHIS 
down throe or four Icvels. Ill nlost egiscs, lhe embedded 
dcfinifion bollo##l.~ olH to dc\[iililioll piimitivcs ~iihiil 
throe levels of einl~,dded delinilions. 
One hilpllitant nioti~e bchiud th0 design mid 
o.)llsh'tletion of MINI-I .t)OCI{ is ils expected role as at! 
inlportant, although apparemly meagre, sour0e el lexical 
and knowledge daia in a piocess called bool~lrtq~pi,g, 
i.e., tt) convelt, m steps, lhe rest of I .I)OCI:, into a Type 
I MTD of the MINI-I .DOCI'~ kind. 
MINI-IJ)OCI'~ has bcell tx/inpletcd: and is now 
availablc to both academic alld comnlercial uscrs. 
Besides its possible use an the iltl01CllS o1' Hnglish 
lex{utlil, its COliStl LiCLiC}II et)iislittiCes a lieCt;ssztl) fiisl 
stop in Belleialillg ~i natuYal set oF semantic primitives 
Ior tile I{nglish Igin\[,tiag0. As illenti,,/lled above, the 
seillantic immitivcs fillill tile basic units lcu the 
COtlSiiilCtiotl (If Type II MTDs ill Ihe s;.tille way 
del'inition lninlili~cs lOlifl tilt: basic tlniis itll lh:.ll ~)l 
Type ! MT! Is. 
1231 
2.2. MTD-XIANTONG 
MTD-XIANTONG is the MTD version of 
The Modern Chinese Diclionary af General Chinese 
Characters (Fu, 1997). XIANTONG ix lhe shortened 
name for tht: original Chinese dit:titmaty, it contains 
abont 10,000 monosylhtbic Chinese words in the fo:m 
of t:haraetcrs with more than 50,000 muhisyllabic 
Chinese words and phrases as examples. The meanings 
of eve D' monosyllable Chinese word are det'incd in the 
sense definitions given under the entry word in 
qncstion. However, the mtfltisyllabie words and phrases 
given as examples following each word sense definition 
do not have accompanying definitions. Each word sense 
definilion is separated irom its examples by a colon. 
An example is given bt:low: 
, 1o I I 
3o I 
40 
The construction of MTD-XiANTONG 
proceeds in two steps. At Step I. more than 50,000 
multisyllabie example words and phrases are 
automatically sense tagged. At Step 2 the definition 
text of every word sense of all head words are 
disambiguatcd, attaching a sense nunaber to every word 
in the sense definitions. Note that the MTD version of 
XIANTONG conlains entries of words in frequent use 
only. Mon,,)syllabic Chint:se words that are not in 
frequent use (2,193 of them) stay out of MTD- 
XIANTONG. Word sense cntrics in MTD-XIANTONG 
total around 15,0(X) with about 50,000 examples of 
compound words and phrases. All words in the 
definition texts and all exanlples are taggccd with wt)rd 
sense numbers. 
The algorithm used in automatically sense 
tagging the nmltisyllabic words in XIANTONG is 
fairly, simple. It takes advantage of one important fact 
ctmct:rning the making of tht: oliginul Chinest: 
dictionary, i.e., repeated appearance of the same example 
word and phrase under diflk:rent head words, or morn 
exactly, under different sense definitions of different 
head words. In fact, all head words in XIANTONG are 
single Chinese characters. For e, xample a bisyllabic 
Chinese word made up of two Chinese characters may 
appear twice as an example under the two composite 
characters, each time as an exmnple of a particular word 
sense of one o\[ the two vharaetcrs being defined. For 
example, tile wt)ld ~.,~. 0ecover) appears under both 
,g~ and ~ ,)~._ appears under ~ as an example 
m Sense 1 (p.275). and under ~__ as an example in 
Sense 1 (p. 151), t~x~. Hence .,~ A01 '~t~_A()I. By 
checking through all example words and phrases in the 
way they arc cross-referrt:d, all words that art: t:ross- 
referred as examples in the dictionary are automatically 
tagged with appropriate sense numbers. This process 
pr¢xluced the bulk of MTD entries which form the core 
of the reasoning system in the atttomatic sense tagging 
of running Chmest: text (Tong, ct al., 1993). 
At Step 2 the antomatic tagging of sense 
numbers to each word in the definition text proceeds in 
the ~me way as described in Tong, et al. (1993). 
Efforts are now being made to blend in human 
judgement to the making el a m(ne accurate version of 
MTD-XIANTONG. 
2.3. MTD-DUOGONGNENG 
MTI)-I)UOGONGNHNG is the MTD version 
of The Mult!funclional I)icli(marv o/' Modern Chittexe 
Words (lSeng & Zhou, 19019). f)UOGONGN\[~NG is 
the shoitt:ned ltanm: itn tht: oHgimtl Chmt:sc dictionaD. 
It c,antains ab,.;,it 7,000 most f,-equenfiy used ,:,.,rds ip, 
everyday Chinese 
MTD I)).JOGONGNIEN(\] ix prodnced by 
manually tagging word sense numbers to each 'word in 
the definition text of evmy sense cnhy undm t:Vel 3 h(.:ad 
word of the entire dictionary. Unlike MTD-X1ANTONG 
which uses ils men definilmn primilives MTD- 
DUOGONGNEN(; uses the definition primitive of 
another Type 1 MTD, i.e., MT1)-X1ANTONG. Hence the 
tagged sense numbers in MTD-DUOGONGNENG agree 
with the syste,n of sense ,;umbers ,;f MTD- 
XIANTONG. This paves lhc way Ior lhe ~,eneration of 
a set o des¢_ iptive semantic p "i i1 t Yes II'O~1 nlole !han 
one MTDs. 
in tht: corpus of tht: delinition texts t)l MTD- 
DUOGONGNENG, !0,003 :vord types appear with 
5~g73 definilion primilives Tim 5,g73 definition 
primitives are the word senses of 2,574 monosyllabic 
Chinese words in the form of Chinese t:haraetcrs from 
MTD-XIANTONG. Tht: total numbt:l of word senses 
~given in MTD-XIANTONG of Ihese 2,574 Chinese 
characters ::!Flqot nts tO 9,363. 
3. Evolution from Type t to Type n MtTDs 
The absolnle advanlage el Type II MTI)s over 
Type I MTI)s is an empirical question. However, the 
advantage of a descriptive set of semantic primitives 
over a set of definition primiiivcs derived from ont: 
~;,~,,Io data source should be obvious smcc the formt:r 
em, ers all possible sonree dala of one nattmd langnat, e 
Using a Type It MTI) on parallel machines, 
especially when a descriptive set of senaantic primitives 
ale used zts n6croit:atu~es of an Artificial Nt:ttla\] 
Network (ANN) svslenl eonslilules ;In inlereslin\[~ 
research to both ANN and MTDs. lqlrlher slill, Type 
11 MTDs may hel t ) revive the dated notion of turning 
symbolic proct:ssing of natulal languages into 
mathematical computation by using "t l'uncti,m that tarns 
Type II MTD dcfinitums based on semantic lmmitives 
into nlathemalieal representations. 
3.1. Front Coneeptual EquivMents to Circular 
Definitions 
Circular definilic)ns may cxisl wilhin 
definitions primitives derived from a dictionary. For 
instance, tx~th hotlonH and basel may fall into a set of 
definition primitives dcri~t:d \[\[oln LI)OCE. \[\[ox~e~m, 
they are locked in :, -tefiniti,.m cite!e: 
bollolttl: Ihe base 011 whieh somellun,. 
stands; the lowest part, inside or outside 
basel: the bottom ol something 
Desirably, a descriptive sel of semantic nri n I yes ix free 
of such circular definitions between members of the set. 
Karen Sparck-Jones(1967) conducted an in-depth stndy 
on circular definitions in ttnglish dictionzuies. Similal 
study on Chinese dictionaries did not come t,-ue ur, ti! 
reeenlly when Chinese MRDs are availaNe. Below are 
some resnlts from a simihu- study on MTD- 
XIANTONG. 
The basic idea is iol Circuh, r l)gfiniiion.s to 
emerge out of sets: '.'.1 (?o::,,'eptual li'quivalenL;'. The :;o- 
called Com'erptttal I,i'quivalettl~' refer to a set of related 
conceptual entities derivt:d from the sense definitions of 
a MTD. The so-called Circular IJefinilions refer to 
1232 
de\[+inition circles tound in a sct oI (;om'eptual 
l+~quiwde,tL~. A stud} which takes rmmt)syllal)ic-x~,ord 
concepts an search units reveals that MTI3 XIANTONG 
has 399 sets of conceptual equivalents mmmg 15,273 
sense definitions. The 3g +.) sets of conceptual 
equivalents contain 1,716 sense dcIinitions which 
arnount to 11.23% of the total. 2g strict ci+culm 
definitions are spotted among the 389 sets of conceptual 
equivalents. Sic tim following for an example: 
u. conceptual equivalents: 
K >,0~ >: ~,<it ~{ A~!,I 
!J:\] A0t ~# t¢02 I< ,\0t 
b. circular delinititm 
~_.A0t <- - .)~ A01 . \[U~/\0l 
3.2. Sense Coml!alil~ Ilelween Monolin.gu'al 
Chinese Ilietionaries 
Sense compatibility computation Bas studies 
in hlyrd, ct al. (1987) ats a mapping ploblenl between 
dictionaries. The mapping was seen as u ~'wnmelric 
binary rehtlion belweclI tw() ddlcrcnt dictiomu its. It* 
lhe p{csellt Stlld 3 Ol1 tilt: deri:'alion of at deseripti~ c set 
of semantic primitives, sense compatibility computation 
ix SCCll ~:ts ;.tit illll:,t)ltal/t clcnlcllt ill merging data from 
multiple dictionary sources. 
T~ao dictionaries XIANTON(I and New 
Modern Chinese l)icliomlrv (henceforth XIANHAN) 
(Wan, ct al., 1985) are used as data. r'~o sense 
definitions illldcr Ihc same head word in Ihc two 
dictiomu its arc comt)atiblc if: 
a. tilt" \['&'o definilion texts are identical: 
b one definition text is inchtded as part ~+l the 
other; 
c. all segrncllls of one definititm tcxt scr.,al'ated 
by punctuations al;q;.Cal ill the other 
definition text; 
d. more than half of tim cxatnplc woids and 
phrases gi*.cn after one dcfiution text appcat as 
examples afler lhe other; 
e. both ha,,c h)ng definition texts, and the tirst 
ten words are identical. 
Results show at total ot 5g,g+~ ol all delinitiott lcxls 
mc compatible with dctaih:d split as lollows: 
a. 23(~: I). 12.5(~ e. 1().5~ 
d. 6.g% e. (¢~ 
On the basis (}i such lind*haS, XIANTONG 
and XIANI IAN arc mcrged into one dicti{mary with 
75,572 wt~rd sense delinilions. 
3.3. From l}efiniliotl Primitives to Semaltlic 
prindtives 
The set of dole*nit*on primitt~es ,}1 a \['ype l 
MTl) close on Ihe MTI). A descriptive set of semantic 
primitives close on all data of tree natural language. The 
distinction between descriplive sctnantic primitives and 
pres('ripffve semantic i}rinailivcs is given in Wilks 
(/977). l{ssentially, unlike prcscri\])tivc scmanlic 
Dimitives which ate prescribed, st} to spcnk, I} 3 inan, 
descriptive semantic primili'+es are derived froln natural 
sources such as a dietionary. This paper represents a 
backoff from (hm (1989) where it was ctaitncd that a 
natural set of semantic primitives had been derived frDm 
tJ)OCi:+. It in now believed that although it is feasible 
to derive a set of definition primitives from {me 
particular MR/) or M'FD, a descriplive set ~i semantic 
prumtB,'cs ttuc ol one natural hmguage Is prclerably 
dciivcd \[ix)n-t iilOi-c \[Jtlili oiic ii;_iLLli+iil SOLirCe Stlch tt~ a 
dietionary. 
What is presented belov+ tCl)lcsents a first step 
to the generation ¢)l a set of Chinese semantic primitives 
1lOill tl sol o1 dolinllion prilniilVCS derived llOl'll one 
particular Chinese dictionary, i.e., MTD- 
1)/K)G()NGNHNG. A hunch set of selllantie primitives, 
~ hich is u subset of the dclinition D imiti,,cs dmi'ecd 
from MTI)-I)UOGONGNb;NG, is applied to the sense 
delinitions ol MTD-XIANT()NG ill an elh}rt to \[illd 
out hou, nlany word SCllSC eilliies Catll t3c accol#lled.l},r, 
i.e. defined, within five defining cycles. 
As given above the definition primitB, es lOulld 
in MTI)-t)UOGON(INEN(TI total 5,,R73. 5,796 of them 
are senses ol words Ihat are kilown as most \[reqtlelltly 
used Chinese words..'1,417 (tilt of these 5,796 de.finition 
primitives appear m{',re than ,.,rice in MTI) 
I3UOGONGNI);NG sense dciinitmns. These 4,417 
dclmition primitives are chosen as the imtial hunch set 
of semantic primitives. On the other Itand, MTD- 
X1ANrONC, was chosen as the first other dictiotuuy to 
tr3, out these 4,417 defining pMmitivcs ftom MTD 
DUOGONGNI'~N(I. A subset of 15,273 sense 
delinitions ol MrD..XIANTONG paicipate in the 
generation processs. This stlbscl, 7,617 strong, bchmg 
t~+ the illost Irequcntly-used words in contelllporttly 
Chinesc. At the lust de'lining cycle, 3,804 ~vt.lld SlHISCN 
"UC defined by the 4,417 hunch set of semantic 
or naB, is N(}te that every tlnm at word sense ix dclined, 
it joins the ttui-lch set immediately to define more Bord 
senses m the txcxt det'hdtlg cycle. At the second 
dclining cscic, till additio\[lul _'J_ 6 wold SCliSes ale 
defined. At the third defining cycle, another 144 word 
senses arc dclincd. At Ihe lourlh delinlng cycle, 27 more 
word scnses arc defined. AI the lasl dclirdtq, cycle d 
additi(~tml v,'cwd senses me dclit~cd+ At the end of the 
Iifth defining cyclc, 3,256 word senses mmam 
undefined. A complexity of reasons contribute to the 
latltlrc. ()tic tnajor one among them ts date to incorrect 
and inconq)lctc data. As discussed above, MTD- 
XIANTONG was produced by machmeJ Woads tagged 
~x.ith iX)0 or MIS indicutc thztt the system was umtl)le tl) 
tag tile correct sense ntnllbef either because the word 
sense is llt}\[ IOtltld ill Ihc dictionary or bccatlSC Ihc 
STy'stCl+;q CallFIOt collie tip with the correct sense nubll-lel 
during the reasoning process, fiflorts arc being made to 
\ el*l 3 the datu cxhausti\.ely by hmnan judgement. 
Rererenee~ 
.cng: Z (7., mid ,~,hotl, X-.I, ( \[9N9)0 i't/e MHlli/Hl*Cliollal 
l)iclionarv o\]" Modern (.'hinese. International 
Culture Puhlishil+.g (:orporation: 't\]cijing+ 
I:u, X..l+. (1937), Moiler** Chinese l)iclionarx+ o/ 
(;cneral Chim:se ('haracter.L Fot'cign 
I ,anguagc Teaching and Research Pubtishmp 
t lotlsC: Bcijing. 
Gale. W. A., Chttrch, K. W., and Yarowsky, 1). (1992). 
Work O*1 slatislical methods fol + word sense 
disambiguation. Working Noiex.lor AAAI Fall 
S.vmpoMmn ()n Frol)at)ilislic Approa<he.~ to 
N::lurql !.at,guaget pp. 54-.60. 
(hde, W. A., Church, K. W., and Yan~wsky, D. ( 1093).A 
naclhod for disambiguatmg word senses in a 
large corpus. (7ottll)uler~ aml t:lumanilie.v_ 
1233 
Gong, J-P., and Huang, C N. (1991), A pmjcct ol 
statistics & analysis of Chinese morphen-~cs 
and senses, Proceedings of NI,PRS-Ol, pp. 
108-112. 
Guo, C-M, (1989), Conslrttcling a Machine I'raclable 
Dictionary ,/'rom Longman Dictionary q/" 
Contemporary tinglih. Phi) desertaiion, 
New Mexico State Uni~,crsity, 
G no, C-M. (Etl.). (1994), Machine 7~'aclable 
Dictionaries: Design and Construction. 
A blex: Nom¢ood, NJ. 
Sinclare, J. (1991), Monitor corpora. Corpus, 
Corn:ordain'e, Collocaliotl. Oxford University 
Press. pp. 24-26. 
Sparek-Jones, K_ (19(,7) Dictionary Cir¢4es. System 
Development Corfx~ration. 
Tong, X., Huang, C-N., and Ouo, C-M. (1993). 
Example-based sense tagging of running 
Chinese text. Paper read at the First 
lntclnational Wtukshup on Vcly l~argc 
Corpora. Columbus, OH. 
Wan, Q-Z., Min, I:-L., Yu, X-M., and Wu, T-Z. (1985), 
New Modern Chinese Dictionary. Xinhua 
Publishing l louse: I3eijing. 
Wilks, Y. A. (1977). Good and bad arguments abx~ut 
semantic primitives. Communication and 
Cognilion¢ 10, pp, 182-221. 
Wilks, Y. A., l;'ass, D. C., Guo, C-M., McDonald, J. E., 
Plate, T., and l~rian, B. M. (1990), Providing 
machine-tractable dictionary tools. Machine 
1"ranshtliott, 5, |7|). cy.)_ 154. 
Yarowsky, D. (lg~)2).Word sense disambiguation using 
statistical models of Roget's categories lraincd 
on large c~)rF~ )re, COl.IN(; - 92. 
1234 
