THE NATURE OF NEAR-SYNONYMIC RELATIONS 
Chrysanne DiMarco 
Department of (Jotnt)uter Sci(',\]~ce 
University o\[ W~terloo 
Wa, te\]'\[oo, ()Jfl;;~rio, C,;ma,da N21, a(ll 
(:d i m ar(:o(~'~logos.ll w;t t eft oo.(:,~ 
1 Introduction 
The words gawp, gaze, ~md stare all <lenote a, ldnd 
of prolonge(I look: they a.re nc¢u'-syno'nyms, or 
plcsionyms \[Cruse 1986\]. \[h)wever, a.s we learn 
from their indivi(lu;tl enl.ries in the Oxford ad- 
vanced learner's dictionary ( OALD; fourth edi- 
rio:n, 198f))~ to ga.ze is to look long a.nd steadily; 
to st;axe is to (Io this with the eyes wide open; 
a.nd gawping has the additiom~l requirement tha.t 
the ~ct I)e impolite or stupid. In recent work \[l)i-- 
Marco, llirst, m~d Stede 1()93; I)iM~u'co and IIirst 
1993\], we ddress 1;1l(; problen~ of representing the 
lexica.l \['ea, tures theft distinguish groups of ne++r- 
synonynts. 
Our lexical tim.tures for (lil/erei~Lia.tion are not 
intended to be any kind of l)rimit;ives for (le(:om-- 
l)ositiomfl semantics: they iir~'(~ \]lOt being used to 
rel>resent whole meanings, t)ut r+~ther l;o rep\]'esent 
d{\[.'fi'~vnec.~ between \]ne,~nings. These dift'erences 
between plesionyms (-a,, I)e sh,~des of (tenota, Lion 
or (:onnota,tion, or emt)ha.ses on dift'e\]'ent (:Oral)O- 
nen t s o \[" t he meani n g. 
Our eveutuM goM is ~ represent;a.tio\], for a lex- 
icon in which sem~mti(" and stylistic distiuctions 
ca,Jl I)e l/lade bel;ween syn(tttylns a,n(l plesionyms, 
I)oth within a.nd ~(;ross la.ngua.ges, \['or the purpose 
of lexica.1 choit:e in natural l,~ngua,ge gener,~tion 
a, nd machine tra.nslation. The na,ture o\[ these dis- 
tinctions suggests tha.t timy can be viewed a.s rc- 
lations I)etween nea,r-synot~yms. In this l)a.per, we 
undertake ~ study of the ch~u'~cteristics of ne~> 
synonymic rel;~tions as a. stel) towa:r<ls a. knowl 
edge rel)resent,~tiou for lexi(:M discrimination. 
2 Previous research 
As a first st;eli , whi(;h we des(:ribed in \[l)iMar(:o, 
llirst, ~u,d Ste<le 1993\], we carried out a. stu<ly o1' 
dictiona,ry usage notes in order to compile a. list. of 
the kinds of dime, nsions flint axe used frequently 
~s (lenotat;iw; or connot~tive dil\[erentiae. We i)r(,- 
duced zt l>relimina, ry list; of 26 (lenota.tional dimen- 
sions and 12 eonnota,tive dimensions (including a 
few l;h~t we aAded from the discussion on lexi(:a,l 
a.sl)e(:ts t)y Viua.y m~d l)a.rl)eltmt \[1958\]). (This 
set is not, yet; complete 1)1' definitive, of course, 
bul; we h;we mamtge(t to in(:lu(le ~ fairly (:ompre 
Imnsive selection of the most common diffexences 
between ne~r-synonyms.) Some of the dimensions 
~re sin@e l)in~u'y choices; others are continuous. 
We show ;~ representative sa, ml)le in Table 1. l)3a(:h 
line of the ta£le shows ~ dimension of differentia- 
lion followed by ex~unple sentences in which two 
i)lesionyms vary a.long th,~t dimension. 
3 Chafl\]u and Herrmalan 
3.1. Basic theory 
(\]ha,\[fin ++nd llerrma.n n \[1988\] lt~+ve provMed ++ the,- 
oreti(:a,1 ~q)pro~(:h for accounting for sem~ntie re 
l~ttions th~tt we will apply to near-synonymic rely> 
t, ions. They desert be ~ systematic study o\[' the na- 
ture o\[" semmltic relations, beginning with a, (:a,ta,- 
h)gue o\[ the relatlion properties that an a.dequ~te 
theory o{" semantic rel~,tio:ns should explain, tbl- 
s lowed I)y ~ list; of s~mlp\[e rel~Ltions. P} e, e re\]a,- 
tio:ns (e.g., synonymity, pscudoantonym) a~re then 
broke,~ down into relation elen\]ents (e.g, symmel~ 
rical position, locative inclusion), whit:h a.re for- 
ma.lly &'fined. (A al\[in and \[lerrmmm's study (:ul- 
mina,tes with ~m explanation of how this rela.tiot> 
ele/netlt ~tt)l)roa.ch (:~m be used to a.ccount lot each 
of the rela.tion propertie, s. We will undert~ke 
a, siniib~r kind o\[' stutly in l)rOl)osing ~ the, oret- 
iea, I ,to:count o\[' near-synonymic relations. I\[ow- 
ever, unlike (\]hattin ~md llerrmann, who begml 
with rea,dily recognizable semantic relations ~md 
then detined relation elements, we :find that in 
our study of ne~r-synt)nyms, it is more al)prot)ri- 
ate Lo begin with tit(; rela,tion elements, which are 
more e~Lsily identitied, a,nd then move on to the 
construction of the relations, which a.re more (lif 
\[i(:ult 1;o detine. 
We will begin by ex~n\]itfing four properties tha.t 
(;ha.flin a.nd \]lerrma.un believe ~ny theory of se- 
ma.ntic rel~ttions should account for ~md we will 
show ttmt these prol)erties a, re a,\[so relewmt t() 
any theory o\[' ne,;tr-syn(tuymi(; rel~ions, t 
1 Chaflin and Iterrmann iuclud.e relation discrimination, but 
as mlr whole study is of lexical differentiae, all our relation 
l)r.pert, ies have some|hing to do with discrimination. They also 
in(:lu(|e r,:lation w~rification, but a demonstration of this pro I) 
ert.y wouM inwdve psychological testing, which we |rove not yet 
.ln(lerl,;I.kell. 
691 
DENOTAT1ONAI, DIMENSIONS 
Intentional/accidental: 
She {stared at \[ glimpsed} him through 
the window. 
Continuous/intermittent: 
Wine {seeped I dripped} from the barrel. 
Immediate/iterative: 
She {struck I beat} the drum. 
Sudden/gradual: 
The boy {shot\[ edged} across the road. 
Degree: 
We often have {mist \] JOg} along the 
coast. 
CONNOTATIVE DIMENSIONS 
Formal/informal: 
lie was {inebriated\] drunk}. 
Abstract/concrete: 
The {erTor \[ blunder} cost him dearly. 
Pejoratlve/favorable: 
That suit mMves you look {skinny\[ 
slim}. 
Forceful/weak: 
The building was completely 
{destroyed l ruined} by the bomb. 
Table 1: Examples of features that dictionary usage notes adduce in word differentiation (adapted fl:om 
\[l)iMarco and Hirst 1993\]). 
Relation comparison. The primary property 
is relation comparison: pairs of near-synonynts 
can be COlnpared and judged as more, or less, 
similar to each other than others. For example, 
there is something similar in the relationship be- 
tween stingy/frugal and between j'at/plump. In 
each case, the first word (stingy, fat) is pejorative 
while the second (frugal, plump) has a nuance of 
being admirable or attractive. This relationship 
would not be maintained if, for example, we re- 
placed fat/plump by rotund/plump. 
Relation expressions. The second relation 
property is relation czpressions, which refers to 
people's ahility to use comnmn words and phrases 
to express near-synonymic relations. For exam- 
pie, mistake and era'or hoth refhr to something 
done incorrectly or improperly, but mistake is 
more general than erro% ~ceording to the usage 
,tot(; in the OALD. 
Relation complexity. 2 The property of rela- 
tion complexity refers to the need to represent 
different relations between the same pair of near- 
synonyms, on more than one level of complexity; 
we nee(l to t)e able tso inchlde nuances that are 
relevant to a given situation and ignore others. 
Relation creativity. Chaflin and llerrmann 
ohserve that "the production and recognition of' 
relations is a creative ability", so that the re 
lation between two words "can he readily iden- 
tified although the reader may never have con- 
sidered the relation of these particular terms be- 
tbre" \[p. 292\]. We wilt show tha, t relation ere- 
ativit 9 is equally necessary to a theory of near- 
synonymic relations, l'br example, the relation of 
2Chatfin and'Jlerrtnann \[1988\] use the somewhat misleading 
terln relation ambiguity, bug we believe it is ntore accurate and 
less confusing to use the term relation complexi@. 
arrange/organize 3 can be recognized as one that 
contrasts correctness with functionality, and we 
might then detect this same relationship for other 
pairs of near-synonyms (e.g., trim/shave). 
In summing up the importance of these rela- 
tion properties to a theory of semantic relations, 
Chaffm and Herrmann state that "these diverse 
phenomena must be explained by theories of re- 
lations" and "we will tind that in order to ex- 
plain relations it is necessary to assume that rela- 
tions are normally composed of more primitive el- 
ements that account for their characteristics and 
for people's abilities to make judgments about 
them" \[p. 292\]. We, believe these observations are 
equally true of theories of plesionymic relations 
and we will show that a relation-element theory 
of near-synonymy will account for these relation 
properties. 
3.2 Theoretical assumptions 
In developing their theory of semantic rela- 
tions, Chaflin and Herrmann make the following 
)'e, presentationa l assumptions \[paraphrased from 
pp. 293-294\]: 
• A relation R, between 1;wo concepts x and y is 
composed of a set o\[" dyadic reb~tion eh,,ments (&,...,&): 
~:l~?j --~ (E~, ..., i,;, )4 
• I~.elation elements may be hierarchically or- 
ganized so that the presence of one element 
depends on the presence of another, o1' ele- 
ments may be independent of one attother. 
In the following representation, independent 
3 "Arrange is to put in a pleasing or correct order ... Or- 
ganize is to put into a working system" (fi'om the usage note 
in tile OALI)). 
4 This notation should be read as "the relation I~ decomposes 
to the relation elements ... ", 
692 
(21(2n:l(2nts ;~r(2 sel)~r~t(2\[1 I)y (:olnm~s ~i(I (l(1- 
tmnde:nt elements a,\[)\[)(2a, l: ill \[mr(2nth(2ses fol- 
lowing th(2 (;lement tha.t th(2y d(2p(2nd on: 
ll.eb~tions In~ w shin'(2 one or more (21ements. 
The greater the i)rol)orlJon of (21ei~l(21lts l;wo 
rel~tio:ns ha,v(2 in (x)\]YIMOll, the It|OF(2, shni\[a,r 
t\]l(2y ~r('. 
Two (;xa, nipl(2s of (lhldIin a,nd l\[(2rl'lnaain's 
s(2.ma, rll;ie relations a r(2 ,Wl~o~ym, il@ a, tld p,s, cls- 
doantonym, which they delin(2 in t(2rrl\[s of th(2 fol 
lowing sets of r(21a.tiol:l (2hmmnts: \[~ 
synonymity: inters(2.ction (inch,shin (I,ila,t(2r~d)) 
paeudoa?ztonym: dhn(2nsion (bipola, r, (;OlillO|;a,- 
live) 
We will ada,pt th(2se r(2l)r(!s(2n t a, 
tiona;l a,ssun\[ptions 'l 1,0 our study of I)l(2sionyuiy 
a, lld |is(2 tii(2ni ill (;otisti'u(;tiiig lt(~;u'-syl\[ollytrii(', r(2- 
bttions \['l:OPd the r(2hl,I;hm (2l(2iilelits to t)(2 (tefi\[i(2(i 
below. 
4 The relation elements 
Ch~dlh/a,nd lIerrm~nn define a, s(2t o\[' relation c,h> 
mcnl,v of which s(2n:lanl, ie rel~tions am(2 eonipos(!\[I. 
These relation (2\[(2til(2iiis a,ro d(2scrih(2d a,s "(;le 
inents tha, t the l'(2l~tions h~(l ill commoli ~ll(I (21(> 
ments theft distinguistl(2d t he r(;l~tions from (2a, ch 
oth('r" \[p. 301\]. We ol)serv(2 tlia,t, for oitr pill'- 
l)OS(2s~ a, i'el~:~tion e\](21Ii(211t is a, (tonol,~tiona,I or (:eli 
nota,tiv(2 li;at\[u'(2 tha,t is pa,rt (or a, ll) of a (les\[::\['il)- 
lion ()f~ I\[(2~+U ' .SyilOtlylilig i:cla,tion; a,ll(i ti(211(;(2 i\](2{l,l'- 
synonymi(: r(;l~tions can be dif\[or(2ntial;(2(l t)y |,h(;se 
va;rious (2\](2nlellts. '\['hlls~ gi V(211 t h is ol)s(2rvlt|,iOll, 
we ca, n consider our f(2gtttrcs or (li\[r(2r('nti~tion, 
~ts illustr~tted in Ta, lil(2 1, to be exa, nlt)l(2s of the 
i'(2l~tion (2\](!lIlel\[tS tha,t (:()tlipOS(2> ~uid \[\[isl;inguish~ 
n(2~r-synony:nlic rela,tions. 
~(2 stipli|~t(1 that; our I\[(la, l'-sy\[Iollylili(; r(21a£ion 
oleIIlellts a,r(2 ilnita, ry, 1;lilt| is~ th(2y r(21)r(2s(2ut l(2x 
i(:~d r(2h~tionships tliat need I\[Ot })(2 (l(2('x)nlpos(2d 
a, ny :fl:u'th(2r. Wtlil(2 we exl)(;(:t |ha, I; i'(2lation (21- 
Oill(2:lll~S will b(2 l~ligua~g(;-ill(l(21)(2nd(2lil; > th(2 d(2gree 
I;0 which rela~tion (2\](2m(2nts need to t)(2, decomposed 
niight dit\]hr fro\[n la,ngm~g(2 to la,llgUltge: We will 
t'(;I\]n(2 the (2\[etni;nt to (2x++(:l, ly the 1(2v(21 (l\[' distinc- 
tioi\] 11(2(;(2sSa,l'y for th(2 l)l(2sionynis 0\[' t}le l~ctiglla,ge> 
:-I~ll(i rio furth(;r. 
"SNore ilia| Challin mid lleri'iliiilitl IA,eat Hylioliyitiy its ~t siilgt<', 
sl;Iil~+lltic l'+|al,lon, while we ave ilil,(!rcs~ed lit tim lil&ily dilferent, 
ll(~itt' SyltOliylnic relkitlollsllii)s that c~ui eKisL 
6I,ack of space precludes a full explanld,ion of t, hese relation 
elelnent sl>\['llCl~lll'(~s! \])II~D il~ iS llOl, nOC(~SSDA'y for undersl,&iiding {lie 
work we wi\]l pi'esenl.. 
7(~ha|Iin and Iierrmann ,Mso make processing assumpl.ions, 
inchldhlg one th~tt rel~tes to psyc\]lological verification of their 
re|&l~ions; we do li(l|, iisc I~l).()st: ;tssuull)tions ill I, his plqmr. 
5 The relations 
We will work through several exa, mpl(2s, showing 
how neaA'-syltotlyltli(; l'(2\]~l, tiollS (;~1,i1 I)(2 construct(2xl 
from r(2.h~tion (2lem(2nts. All the word d(2scrii)tions 
in th(2 (2xampl(2s b(210w will I)(2 taken from usa+t,;(2. 
not(2s in the OA LD. Our first ex+mlt)l(2 is the r(2.p- 
r(2s(2nt~d, ion of the distinction betw(2en ask and be- 
SCCC\]~: 
\[Ask\] is tile most usua,l ~md ilffortmd 
wor(I ... besee,:h \[is\] strong;(2r a,|ld Irlol(2 
form M t h~m beg. 
Fron/this usa,g(2 |lot(2, a, lld our owt\[ na,tiv(2 speld~(2r 
knowledge., we identify the i'(2h~l;ion (11(2m(2nts tha, t 
distinguish (2.~u:h word: 
a,~i~: g(2n(2r~d; inforniaJ 
bc,~eceh: fm'raa,l; fore(2fill 
This rtota, tion (:m~ be r(2;~d a,s "a:;k is move g(2n 
(2ral ~md l\[XOl'(2 iiil'orma,l than bc,seeeh; beseech is 
mot(2 rorma l a, nd inor(2 rhetorically forceful tha,n 
a,~k." We construct th(2 r(,la,tion b(2tw(2(21J a,vA: a,.d 
bc.sccch by taldng t he eom pl(2m(2nt of th(2i r r(2sl)(2c 
tiv(2 roh~tion-(21(2m(2ut 8tl'tl(:t\[lr(2s ;\]btld t\]l(~ll indic:at 
ing d(21)(2tM(2n(2i(2s b(2tw(2(2.tl l>h(2 r(2sulting rela, IJon 
(2l(2\]ll(2\[tts: 
..~;/bc.~,.~<.h: (r;(2,,,',~l (ro,.,,ml> ro,.,:(2r,,IjO) 
This i'ehl~tion sl,at(2s |;l:\[~t a.s'k is less forloa.\] mid 
less for(:(2l'ul tha.tl /Jc,';ccc/z. (The. ji suhs(;L'ipt on 
forma, l a,n(I forceful should I)(! read a,s i:ndicating; 
1;h(2 (lir(2(:l, ion 0\[" tim :r(21a, tionshi I) I)(2tw(2(2n word .;7 
a, nd word i.) This nota, tion a,lso shows tliitt th(2 
l'eh/,tion ('h;In(2nts formal ~tll(l Jbrccful a,r(2 i)oth d(2 
l)(2n(hmt O\[l th(2 doniii\[~nt (21ein(2nt .qe'ncral. 
(\]li~\[tiu mid lt(2rrn:iann show tha,t t'or s(2illa, l\[- 
tie r(2la,tiolis, the sa, in(2 r(2J a,tiOl\[ (',a,ti hoi(l b(2l, w(2en 
11101'(2 I;ha,\[i Oi1(2 pa, ir of WOl'(ts. As this 11(2,xt ex- 
mnl)l(2 for lit, lit, a,nd (:~uu:iated shows, tile s~mw 
plcsiow!lmic r(21a, tiol\[ (in this (:a.s(2, th(2 r(21~tion b(2- 
twe(2n ask~beseech,) can hold t)(2tw(2(2n lilOr(2 tha,:n 
one pair o\[ words. Tim OALD u s~g(2 llOt(2 for I,h,i'~ 
a,n(t c'maciatcd descril)(2s them a,s follows: 
Thin is th(2 most genera,\] word. it ma,y 
I)(2 n(2gativ(2, suggesting w(2M(ii(2.ss or lax:l~ 
.. 11 • of ti(2a, lth . Emaciated indicates a, s(2-- 
rious condition r(2sulting from sl;a,rva, tion. 
We identify IJh(2 rela,I;ion (2\[(2ilteilts that distinguish 
tlzi'~z a,n(\] cJmaciagcd: 
thim g(2n(!ra, I 
emaciated: l'ornl;d; fol'c(2ful 
We tM~(2 the eoniplein(2tit 0\[' tti(2s(2 r(2la,tion (2l(2.- 
nq(2nts a,n(I indica, t(2 tim (l(21)(2nde:nci(2s: 
U~in/emacial, cd: (gencrM (rorm~dji, fo,'e(2fulji)) 
A Irior(2 (:Oml)i(2x (2xa, lnl)\](2 is tii(2 r(21a3;ion I)(2- 
tw(2(2,1t qum'rcl a,nd 7"ow. The OA LI) lisa,l{(2 lie|(2 
describes th(2ir distinguish|hi>; fe;~tl\[r(2s ~ts fidlows: 
69.7 
A quarrel is a sharp, often angry, ex- 
change of words between people ... A 
row is angry and may involve shouting, 
usually tbr a short time ... A row can 
also take place between public figures or 
organizations. 
\[\['here are two ways we can construct the relation 
between quarrel and row, depending on whether 
the argument is between people or inanimate or- 
ganizations: 
quarrel/~vw: 
(Ibrceful, fbrlnalji, emotionalji(vectorialjl)) 
quarrel/row: 
(forceful, \[nan\[mat ell (forlnalji)) 
The tirst relation states that row is more formal 
and more emotional; quarrel is more forceflll. It 
also indicates that the greater emotion of a row 
is linked to a difference in scale, the vectorial el- 
ement, which in this case refers to the diff>rent 
lengths of time of a quar~vl and a row. The sec- 
ond relation notes that a row can inw)lve inan- 
intate entities but, if it does, then the effect is 
more formal. Thus, we can have different rela- 
tions between plesionyms, depending on the dif- 
ferent usages of the words. 
By following the same kind of approach, we can 
(:onstruct relations for some other pairs of near- 
synonyms: 
flvwn/,qrirnaee: 
(general (formalji, torcefn\]jl) 
mistake/blunder: 
(general (formalji, fornefi~lji, carelessji)) 
fat/plump: 
(general (forcefifl (politqi(attractivejl)))) 
We observe that the same or similar relation 
(:an hold between different 
pairs of near-synonyms, for example, ask/beseech, 
thin/ernaeiated, and ffow~@rirnace. This is mta.l- 
ogous to the case of semantic relations, which, 
as Chaffin and Herrmann note, are readily recog- 
nizable and nameable. Near-synonymic relations 
cannot be so easily labelled, but we can still see 
that some basic set of relations might be defined 
and could be used to construct new relations, l?k)r 
example, we showed that the relation between ask 
and beseech couhl be represented by the following 
structure: 
ask/beseech: 
(geuera\] (forxnal.ii, forcefulji)) 
We saw how this basic relation could also apply 
to thin/emaciated and fl'own/grimace; this sug-. 
gests that, lor lexical-choice processing, we will 
want to keep a catalogue of existing relations from 
which new relations could be built. Another pair 
of near-synonyms, mistake and blunder, share the 
same distinctions, except that blunder is often the 
result of carelessness (OALD). So we add to the 
existing specification to obtain the following rela- 
tion: 
re\[sick, c/blunder: 
(general (formal\[l, fo:rcefulj~, carelessji)) 
Imstty, dependencies can lead to quite com- 
plicated relations, as iu the case of fat/plu'mp, 
where the distinction of politeness (intpoliteness) 
is related to different dependencies for each near- 
synonynu the nuances of force and impoliteness 
are interdependent, as are those of politeness and 
a.ttrantiveness. 
6 The relation properties 
In Section 3.11, we set out a list of relation 1)roper- 
ties that any theory of i,ear-synonymic relations 
should be able to account lot'. In tlhis section, 
we discuss how a relation-element approach ad- 
dresses these issues. 
Relation comparison. By breaking down the 
relations between ptesionyms into relation ele- 
ments, we can obtain a finer degree ot' discrim- 
ination between similar words for the task of lex- 
ical choice in generation. As we discuss in \[l)i- 
Marco, \[first, and Stede 1993\], many of the se- 
mantic distinctions between plesionyms do not 
\]end themselvns to neat, taxonomic differentia- 
tion; ratlher, they are fuzzy, with plesionyms of- 
ten having an area of overlap. For exa,mple, the 
boundary between forest and wood is vague, and 
there are some situations it, which either word 
might be equally appropriate. The i)roblem is 
compounded when we are dealing with more than 
one language, for the %veakpoint' between small 
and large tracts of trees is different for differ- 
ent languages. For multilingual generation, we 
can compare plesionyms in different languages in 
terms of their different elelnent structures, so that 
it shouhl be easier to choose the particular word 
iu a particular language that tits a given situation. 
Relation expressions. We have seen thai; o\[: 
ten the distinctions between near-synonyms need 
to be expressed using common words and phrases. 
But we have shown that there are ways of ex- 
pressing relations using fairly common vocabu- 
lary to represent these distinctions. The ease of 
relation identilic,~tion may contribute towards re- 
lation veritication: we ('an anticipate that psycho- 
logical tests, of the soN, Chaffin and IIerrmann 
carried out lbr semantic relations, could be used 
to verify our relations and relation elements, as we 
can meaningfidly and precisely represent the sub- 
jects' intuitions about the distinctions between 
n ear- syn onyms. 
694 
Relation complexity. R,elations may need to 
be de.scribed ~1; more= than one hwel of c<)ml)lex- 
ity, so that the distinctions between two words 
may be identified in more than one way. We have 
shown how a relation-elenteut approacll allows us 
to detine difl'erent relation structures for the same 
1)air of neaa: synonyms (e.g., quarrel/row). 
Relation creativity. We have noted in pre- 
vious work \[DiMarco, \]lirst, and S1;ede 1993\] 
that the rel)resentation of 1;he distinctions be= 
tween near-synonyms would seeln to reqllire a 
constrained, bul; not finite, vocalmlary. With a 
relation-element approa<-h, we have seen how a 
hasi(: set of relations might be constru(:te(I; new 
relation eleme.n.ts nlay be a<lde(\[, but we may be 
able to incorporate them into existing relations, 
so that tim ('al;alogue of relations need not grow 
uncontrollably. Tbus, we (;an \])roduee new re- 
lations by elaborating on existing, well-known 
relations or by concatenating existiug relations 
\[p. 3221. 
7 hnplementing near-synonymic 
relations 
We are currently invesl, igating (lilferent systems 
for implementing a relational theory o\[ near: 
synonymy. The first system that we al'e looking 
at is WordNe(; \[Miller et al :1990), whi<;h s('ems 
p~rti(;uhtrly relew~nt as words are organized both 
by semantic relations and by "synsets" (synonym 
sets). 
WordNet contains delinitions of uomls, verbs, 
and adjectives; for now, we are COileeiitroot 
ing on the reI>resenta,tion of adje(:tiwtl near- 
synonyms. In keel)log with the l)hilosol)hy of 
WordNet, we envisage the use <)f a i)ointer 
for each type (>\[" near-synonymic relation in 
our cal;alogue, so that we might tel)resent 
the relations betwee.n plesionyms as follows: 
rl : (genera\[ (favourableji)) 
r2 : (general (fbl:tnaljl, forcefitl.ii)) 
r3: (general (favon,'abh b forcefulji)) 
Currently, the coding of a, synse.t of adjectives 
wouhl look as lbllows in Word Net: 
{ thin, slender, erase.fated, thin1, & } 
where "thinl, &" indicates that members of this 
syltset are related to the '(:<)n(:e.l>t' thin/ I>y the 
simila,rity relation. 
We can imagine iml)<)siug additional structure 
on a synset and malting use of a catah)gue of near- 
synonymic relations to obtain the f<)llowing cod- 
ing: 
{ \[thin, sh;nder, rl\], \[thin, eros.elated, r21, 
\[shin<let, emaciated, r3\], thin1, & }s 
8In WordNet, square brackets are used to indicate a lexical 
While such a representation of near synonymic 
relations would be very easy and natural in Word. 
Net, it relies on the solution of a uunll)er of chah 
lenging probh~ms, speeili('ally, how to generate a 
comt>lete set of near-synonymic rehLtion elenl(utts, 
and how to define a constrained and reusa, ble cat- 
aJogue of nea.r- synonymic relations. 
8 Conclusion 
Our evelltlla\] g()a,\[ is a knowledge representati,..>n 
h)r the <Iiscriminati(>ll of near-synonylns. We \]|ave 
1;aken a step towards such a, rel>rcsentation I)y 
investigating the nature of near-synonymic re 
lations. We have set out Ibm" properties that 
any theory of nea, r--synonymic relations sh<mhI ad- 
dress, and we have shown how relation-element 
theory gives us a way of representing distinctions 
I)etween near:synonyms that a.ccounts for these 
properties. We are cui:rently investigating the 
ret>resentation of our relational api)roach to near- 
synonymy in the WordNet system. 
A cknowh~dgmne.nts 
We tha.nk (Iraeme Ilirst for reading and giving incisive 
(:ommcnts (m earlier drafts of the paper. Wc also thank 
the anonymons referees for snggesting an extension of this 
work. We acknowledge the support of the Natural Sciences 
and Engineering tlx~sear<:h (Jounci| ot Canada. 
R,~,,fer orl('.(~s 
|)iMarco, Chrysanne and Hirst, (;rae.me. "Usage notes 
as the b;rsis for a representation of near-synonymy for 
lexical choice." Proceedings, Ninth Annual Conff.rencc 
of IIw University of Watc,'loo (Tcnlre for the New Ox- 
fln'd English Dictionary and Text t£cse.a~:h, Ox\[ord, 
September 1993, 33 43. 
i)iMareo, (\]hrys~tnne; 11irst, (h'a.cme; and Stede, Man\[red. 
"The senm.nti<: a.nd stylistic dilfcrentiatiolL o\[ synonyms 
and ucar synonyms." l'rocecdings, AAA \[ Spring Sym- 
posium ou \]\]uihlin9 Lexicons .for Machine 7Yanslation, 
.qtanfor</, March 1993. 
Cha\[lin, l/oger ~md Ilerrmanu, I)onglas ,I, "The m~ture of 
semantic r(~l+tl.ions: A comparison of two a.ppi:oa(:h',:s." 
in: \]~clalional models of the lexicon, Martha W. Evens 
(editor), Cambridge University Press, 1988, 288 334. 
(',ruse, I).A. Lexieal semantics. Canfln:idge University 
I'ress, 1986. 
Miller, ({eorge A.; BeekwiLh, Richard; Fellbaum, Chris- 
tia.ne; (Iross, I)erck; and Mifler, Katherine. "l!'iw~ pa 
pers on Wordnet." CSI, Report 43, July 1990, (k)gni= 
tive S(:i(mee l,aboratory, Princeton University. 
O:dord advanced learner's dictionary of current English, 
fourth edition. Oxford University Press, 7198!). 
Vinay, J.P.a.nd I)arbclnet, 21. Stylistique comparde du 
francais el dc l'anglais, l~eauehe.min lt&h :1958. 
rclatlmt between words, rather than a scmmltlc relation between 
concepts. 
69'; 
