Automatic Recognition of Verbal Polysemy 
Fumiyo FUKUMOTO * Jun'ichi TSUJII l 
Centre for Computational Linguistics, UMIST 
P.O.Box 88, Manchester M60 1QD, United Kingdom 
E-mail fukmnoto(c~ccl.umist.ac.ul~ tsujii@('cl.umist.ac.uk l 
Abstract 
Polysemy is one of the major causes of difficulties in se- 
lnantic clustering of words in a corpus. In this paper, 
we first; give a definition of polysemy from the view- 
point of clustering and then, b~rsed on this definition, 
we propose a clustering method which reeognises ver- 
bal 1)olysemics from a textual corpus. The results of 
experiments denmnstrate the effectiveness of the pro- 
I)osed method. 
1 Introduction 
'\]?here has 1)een quite a h)t of research concerned with 
automatic clustering of semantically similar words or 
automatic recognition of colloc~rtions among them from 
eorl)ort~ \[Church, 1OVl\], \[Hindle, 1991\], \[Smadja, 1991\]. 
Most of this work is based on similarity measures de- 
rived fl'om the distrilmtion of woMs in corpora. How- 
ever, the Nets that a single word does have more than 
one meaning and that the distribution of a word in a 
corpus is a mixture of usages of different meanings of 
the same word often hamper such atteml~ts. 
The meaning of a word depends on the domain in 
which it is used; the sitme word c'an be use(l differently 
in different dolnains. It is also often the ease theft a 
word which is l/olysemous in general is not l)olysemous 
in a r(,strieted subject domain. In general, restriction 
of tllc subject domain makes the t)roblenl of 1)olysemy 
less l)rol)lematie. However, even in texts fronl a re- 
stricted domain such as Wall Street Journal l, one en- 
eount.ers quite a large nulnber of l)olyselnous words, in 
particular, unlike nouns, verbs are often i/olys(mwus 
ev(,n in a restricted subject domain. 
Because polysemous verbs are usually also high- 
frequency verbs, their treal:ment is crucial ill actual 
applications. Furthermore, beeause of their high- 
frequen(:y, polysemous verbs tend to have a harmflfl in- 
th,ence on the senlantic ehtstering of l/ollns, \])eeallSO se- 
mantic clustering of nollns is usually 1)eribrmed based 
on th(.ir eollo('ational 1)ehaviour with verbs. 
* I.'UKUMOTO i.~ now at Department of Ele(:trical Engineer- 
ing mM (~omputcr Sciencc, Faculty of EngineerilJg, Yamanashi 
UaivcrMty. E-mail fukumoto~skyc.esi,yamalmshi.ac.jp 
t Wall Street ,lo'~tr'nal was prepared by ACi,(Associalkm for 
(~omputational IAt~gMstics' Data Collection Initi~ttivu) in \[99l. 
Although polysemy is said to be widespread in lan- 
guage, the definition of polysemy is highly subjective. 
Polysemy can only be recognised by hunmn intuition 
and different linguists often identify a different number 
of senses in the same word. In this paper, we first give a 
definition of polysemy fl'om the viewpoint of clustering, 
and propose an overlapping clustering method which 
automatically reeognises polysemous words. The re- 
sults of experiments are also given to demonstrate the 
effectiveness of our method. 
2 Related Work 
Although there have been several attempts to extract 
semantically similar words from a, given corpus, few 
studies seriously deal with the problenl of 1)olysemy; 
of these, even fewer are based on real texts. 
The techniques developed by Zernik \[Zernik, 1991\] 
and Brown \[Brown, 1991\] seem to cope with the dis- 
crimination of polysemy and 1)e ll~Lse(l on real texts. 
Zernik used monolingual texts which consist of about 
1 nfillion words tagged by 1)art-of-spee(:h. I~Iis method 
associates ca(-h word se.nse of a polysemous woM with 
a set of its co-occurring words. If a word has sew 
eral senses, then the word is assoeiated with several 
different sets of co-occurring words, each of which cor- 
responds to one of the senses of the word. The linfita- 
tion of Zernik's method, however, is that it solely re- 
lies on human intuition for identifying different senses 
of a word, i.e. the human editor tlas to determine, by 
her/his intuition, how many seilses a word has, and 
then identii~y the sets of co-occurring words (signa- 
t.lcres) that correspond to the different senses. 
Brown used bilingual texts, which consist of \]2 mil- 
lion words. The results of Brown's technique, when al)- 
plied to a French-English nmchine transb~tion system, 
seems to show its eflbctiveness and validity. However, 
as he admits, the at)preach is linfited because it can 
only assign at most two senses to a word. More seri- 
ously, 1)olysemy is defined in terms of translation, i.e. 
only when a word is Lranslated into two different words 
in a target language, it is recogniscd as polysemous. 
The apllroach can bc used only when a large 1)aral - 
lel corpus is awdhtble. Furthermore, individual senses 
thus identified (1(1 not neeessarily constitute single se- 
mantic units in the monolingual domain to which 1)lau- 
sible semantic prollertics (i.e. semantic rest;rictions, 
762 
colhlcations, etc.) can lie associated. 
The defects of these two methods show that it is cru- 
cial to have an N)pr()l)riate detinition of polyscmy in 
terms of distributimml 1)char|ours of words in mono- 
lingual texts. The approach proposed iu this paper 
focuses on this problem. Like Brown's apl)roach , our 
al)proach ad(lpts ;L rebttivist.ic vicw of polysclny. That 
is, ~ word is rccognised as l)olysenmus in terms of other 
relai.ed words..\[{owever, while Brown's al)l)roach idcn- 
tilies polysemous words in terms of rela~ted words of 
ram|her lmigui~gc, we. use semantically similar words 
of the same llmguage to identify polysemous words. 
Whether a word is polysemous or nol; depends on 
whether i~ set of other, semanti('Mly similar words ex- 
ists whose distrilmtional 1)eh~viours correspond to it 
sitbset of the distributionM behaviour of the word. 
Because tile distributional beiu~viour of it word is 
character|seal 1)y its co-occurring words, the t)rocess of 
identifying such subsets essentially correslmuds to 1.he 
1)rocess llcrformed manually by {:he hmnan edil.or in 
Zernik's approach. 
The experilmm~s in this p~ller use a corlluS &llllO- 
tal;ed only 1)y 1)art-ofsl)eech 1)ut not structurally an- 
notltl;cd. Howev(% the clustering algoritlm b which m> 
t(nna.ti(:ally recognises l)olysemous words, only ;~ssmnes 
that w(irds are semanl;ic~lly ch~tracterised by a vector 
ill a.n 't>(linmltsional space so that i{: c;tn 1)e al)l)lie.d to 
any data sa.tisf'ying this condition. 
3 Polysemy in Context 
'\]'he l)asic assumption of this w(irk is the stone as that 
Inade in pr(wious COl'pus-t)ased al)tn'oach(~s , i.(', SOlll.gtll- 
tically simib/r words appeiu' ill ~t similar (xmtext. Se- 
nmnl.ically simihu" verbs, for example, co-oc(:ur with 
the s~mm n(mns. The following sentences from the Wall 
Street Journal corpus show the t)oint: 
(s\]) New York Times said it offered to buy the 
shares of 1lop radio corl). 
(s2) tie may sell more shares in the Ollen market; 
or in 1)rive|to translu;tions. 
I1. is intuitively ol/vious that buy and sell are sema.nti- 
tally feb|ted and that the semantic ('loseness (if these 
two verbs is ,nanifest(xl lly the fact that they ('o-oc('ur 
Wll,h 1,1l(~ SitlllP ll/lllll ,sh&\['(!s. "igVo (!all l;hillk (If |ill ?b 
(timcnsional space, (~iL(:li dimension of wllich is associ- 
ated wilh a speciiic noun aml in whi(:h ~ vm'b is as- 
signed a. vector whose value of the i-th dimension is 
the wdue of mutual information (mu in short) \[Chur('h, 
1991\] between the verb and the noun assigned to the 
i-th axis. If the 1)iu~i(: assumpti(m is correct, then se- 
mlmfic~dly similm' verbs form it cluster in 1:he Sl)ace, 
and t:herefore, sta.tistical clustering ~flgorithms can be 
~q)iilied to verb vectors in order to discover semantic 
classes of verbs. 
Ih)w(,ver, this strltigh(;forw~trd method is often ha,ln- 
pered by the existence, of 1)olysenmus words. The fol- 
lowing s(mtences show potysemous usages of t~rke. 
(s3) In the past, however, coke has typically 
taken a minority stake in such ventures. 
(s3') Guber and peters tried to buy a stake in 
lllgill in 1988. " 
(s4) That process of sort, ing out specifies is 
likely to take time. 
(s4') We spent a lot of time and money ill lmild- 
ing onr grou t) of sta.tions. 
(sS) Peol)le |ire queuing at the door to take Ills 
llroducl~ l/u| he dtlesn't have tile working 
capit.M to m~d~e the thing. 
(s5') Goodyear used i~twood trade credits lo 
olltltin, chemi(:;ds mid other products ;rod 
services in the U.S. 
We can nl~d(e the following obserwttions. 
1. take and buy in (s3) ,md (s3'), take and spend ill 
(s4) and (s,I'), t~tke and obt,6n in (s5) and (s5') 
co-occm' with the noun sl.ake, time ~tnd product, 
respectively, mid the verbs of each of these pairs 
\]utve almost the stone SPllSO. 
2. While certain usages of tttke have senses similm' to 
buy, spend, ~tnd obt~tin, these three specific v(~x'l)s 
h~tve distinct, senses and we hardly see synonymy 
itmong these verbs. 
In the space spanned by the three axes, each ass()- 
ci~tted with stake, tim(', a.nd product, t.~tke does not 
constitute a clust.er with aaly of the three wu'bs, take 
co-occurs with the three iiO/lltS iLll(| hits high "m,u v;-tll|es 
with t.heni, while \]lily, spend lind obtain have high m,u 
values only with one of the three nmms. Therefore, I.he 
(1.istaIK:c8 \[)el;WOelt take mid these three verbs are large 
&lid the synonymy of fake with them (lislq)petu's. 
\[n order to c~tpture the synonylny of ttflu, with the 
three verbs correctly, oHe has to deconipose the vector 
assiglled to take into three COlllpon()lit, v(~(Ttol'S, e~tch of 
which corresponds to the three distinct usages of take. 
The decomposition of a vector into i~ set of its cOral)O- 
nent vectors requires i~ l)roller det:onqlosition of con- 
text in wlfich the wor(l occurs. Figure 1 shows tlw de- 
(:onq)osition of the verb take in the thl'ee-dimensional 
spaces, takel, take2, iul(l take3 iLre the (:OmliO- 
nent ve(:tors which ('olh~ctively ('onslitute the vector 
assigned to take. 
For the sltke of si,nplMty, we assume in |he ~d)ove 
t.hi~t tile three nouns chlu'~rcterise the contexl.s where 
the ver\]) la.k(~ o(:cttrs ;in(l, a,t 1.he slmm time, each of 
l.lwm ch;u'acterises n distinct usltge of take. IIowcver, 
ill iL ~j(?llcra\[ situ,%tion, ;~ \[l(llys(!IilO~lS V(~rll (:o-o(:(:ltrs 
with a bu'ge groull of nouns and one has 1;o divide the 
gl'Olt 1) of llOllliS inl;o it set of sullgroups, each of which 
correctly chm'acterises the context for a stlecific sense 
of the polysenmus word. The Mgorithm has to be able 
to determine when the cont.ext of & word should be 
divided and how. 
There m'e clustering algorithlns, called o,oe, rlappin, 9 
cluste'rinf! \[Jardhw, 1991\], which allow ml entity t.(/ I)e- 
763 
time 
take2 / 
\[-.. 
1 
1 
stako buy 
spend 
take ~, 
I I 
take3 
*" J/'~'-~ product 
obtain 
\]:'igure 1: The decoml)osition of the verl) take 
long to more than one cluster. However, these algo- 
rithms assume that ewm an entity which belongs to 
more than one clusters is still a single entity. An en- 
tity behmgs to several clusters because it can be seen 
from several different viewpoints. 'rite same entity, for 
example, egg, can be seen as food, like bread, and as 
ingredients-of-food, like flour, at the same time. 
However, as we saw in the above, polyselnous verbs 
can be captured more naturally by seeing them as mul- 
tiple entities, which hal)pen to take the same surface 
form. takel, take2 and take3 are distinct; entities 
(we (:all them hypothetical verbs in the following) with 
which different sets of nouns co-occur, and with which, 
therefore~ ditferent contexts are associated. 
Therefore, unlike standard overlapping clustering al- 
gorithms, our algorithm explicitly introduces new en- 
tities when an entity is judged polysemous and asso- 
ciates them with contexts which are subcontexts of 
the context of the original entity. Our algorithm has 
two basic operations, splittin9 and lumping. Splitting 
means to divide a polysemous verb into two hypothet- 
ical wwbs and lumping means to combine two hypo- 
theticai verbs to make one verb out of them. 
4 Measuring the Compactness 
of a Group of Verbs 
The algorithm should decide when a verb has to he 
split into two hypothetical verbs. The decision is based 
on a measure of the sel-ilan~;ic compactness of a group 
of verbs. The semantic compactness of a group of verbs 
is a measure which shows the degree of dispersion of 
the group in an n-dimensional space. The compactness 
of a group of verbs, VG= {vl, v2, ..., v,~}, is defined 
as follows. 
1. Let vi be one of the verbs v,, • .., v,,, and a vector 
assigned to vi be (vii, " ", vm). Each vij(1 < j <_ 
n) is computed by the following formula. 
vii = mu(vi, t,j) if'mu(vi,n5) >_ a, 
0 otherwise (1) 
IIere, mu(vi, n j) is the vahle of mutual informa- 
e tion defined in \[Chur Jr, 1991\] between t~i and nj. 
c~ is a threshold value given in advance. 
2. The centre of gravity of a group of verbs, vl, • •., 
v,, is the mean vector of the vectors assigned to 
the verbs~ which is used to eompute the disper- 
sions of the individual verbs in the group. The 
(:entre of gravity ~ = (gt,'", g~), and the length 
of it I 9 \[, are defined as follows. 
(,¢1,. • • ¢~) = ~'i~," " -- vi,,) ' H?, 
i=l i=1 
tl Wt 
(2) 
3. The dispersion, disp(vl,...,~4~), indicates the 
compaetness of a group and is defined ~ts: 
disp(vl," " , v.~) = 
i=1 j=\[ 
4. Let us think of two clusters of verbs, A and B, 
which have the same degree of dispersions. If I g I 
of A is larger than that of B, the absolute vMue 
of mu calculated for A is larger than that of \]3. 
This means that the absolute probabilities of co- 
occurrences of each notln and the verbs of A is 
larger than those of B; zus a result, A shouhl be 
judged to be semantically more compact than B. 
Therefore, the dispersion of (3) is amrmalised ms: 
5. 
disp(v~,. . ., vm) 
~t,i.~.>o,<,,~,..., ~,,~) = I~1 (4) 
disp,~o,, of (4) is prolmrdonal to the number of 
verbs. This means that a cluster of a greater 
number of verbs tends to be judged to be less 
compact than those, of a smaller number of verbs. 
Therefore, the dispersion of (4) should be fl~rther 
normalised to compensatc the effect of the num- 
ber of verbs in a group. This normalisation is 
done by least square estimation. The result is (5), 
which will be used to measure the COml)aetness of 
a group of verbs. 
C'o,,,(,,~, . . . , <,~) = ,lisv,,,.( ,~,,. . . , ,,,,, ) 
,8 * m - 7 (/3 = 0.964, 7 = 0.495) is a coetfi- 
eient that is eml)irically determined by least square 
estimation 2. 
In the following, we use (5) as tlle wdue which shows 
the coml)actncss of a groul). A group with a smaller 
value of (5) is judged semantically more compact. 
2In this case, we set a' in (\[) equa\]s \[o 2{.0. 
764 
5 Clustering Method 
ht lhis st,::lion, wc ltr:,s:,nl our clustering algorilhln. 
Wc first ('xplain the :)pert:ions of splittin+l Hlld hvmping. 
Th('n, we show th(, flow of the algorithm and Cxltlain 
how the whoh' algorit,hm worl:s. 
5.1 Th(', Basic Idea 
Tlw clust:ering algorithm prolmsed in this imlwr I)('- 
longs to the ovorlapl>ing tyl):,. Tlw L¢I,. (1; ::: 1,?.,3,...) 
mvthod, prol)OSe(\[ I)y .Iardim', is (tilt, of th(, typical over- 
btppine; chtstt,t'ing algorithms \[Jardino, 1991\]. The os-. 
scntial dill'('rence l>etwoen ()Ill: algorithlu and tlw lit. 
ut('thod is thai out' algorilhnt txpli('itly introdtu,vs a 
(+ou(lit:ion when an cnt:it:y (+t verb) should It(, sl)li.I aim 
assigned t.o several clust, ers. In ottr method, wlu,tlwr }t 
VOI'\]) I! h}ls I;W() SOllS(!s 01' llOt i?~ judged I> 3' COml)aring 
tlw SOlll;-LltLic ('Olll\])a('l;1H'ss wthws of groups of V(,l'\])s {,0 
It:' produced. Thai: is, (hero art' possil>ililios of creating 
tit(' following three clustvrs: 
{,,'t, ,'~ }, { (,".,, ,'~ } ((it 
{t.. (,,,, ,,~ } (7) 
ultero +'I att:(l c,.,: ht (6) aro new, hyltOl}wtic:tl Vt,l'I)s 
whi(h corl'CSl)Ond 1o Iwo disfiu('t sensc>~ <tt' lilt' st:no 
yet'It, c. These Ix\::) n(,wly int:rodu('vd verl)s are sup 
I)Os('d Io al)l>ear in dilh'l'eUt c:mlexls. Their COllteXIS 
are }wtuMly hyl)othcsis(,d lt,v (\[ividing the sot of :~otttts 
1hat ('n-o('('ur with th:' v(u'l) c into l\vo distinct s:'ts 
()\[ nottt+s. This division of the ('()II{:('Xl O\[ th(, origimd 
v(u'l) +' is hyl)oth(,sis:,d has('d on ttw s('l of nottns thai 
('o o(+('ltl'S with Wl and lh(' set of no:ins that ('o-.o('('ltl'S 
wilh w:+. 
5.2 ,gplil, ti?t 9 and L'.,mpi?~g 
Tho Olt(,rations of splilti'n:/and l'umpin~l art' d('lin('d }is 
I'ollows: 
1. I)'un('tion split(v i, vp, t,q) r(q.llrlts *'(i arid i';~. 
+'i is a vvrb whose COOl'ditmte in an tl-(linwllnioltal 
Sl)a('o i> (v/i, "" ", t'i,). +'ct aud v,J arc hypotho 
sisod verbs whose com'dinatcs in tit(' ii.:liltU,ttsional 
space are ma(\[o h'om tit(' (oordinah,s of It:(' orig- 
inal v(,rl) +'i by dividing Ill(, set o\[ nOUllS that ('o- 
occur wltlt I'i into two distinct sets. Tho division 
is math'in terms of two sets of nouns: ore'is the 
sol of nouns which co-c)<'('ur wit h ci,, and the ot her 
is tit(' set of nouns which co-occttr wit h QI' 
splil(vi, +'~,, +',() ::: (f',, ~'J) 
u,h('rc ('om(l'i, vq) < (7'olll(l'i, "1;) (8) 
I'(I :z: 
t'(I I 
I'(i. 2 
S.I. U(Ij :: 
I'(t tt 
vij if "l,J :/ 0 
(1 ot hel'wis(, 
'2. 
vJ -: 
I',,\]l 
i'\]~ 2 
• sA. I',',¢j = 
0 it" (v, n = 0 and 
q,a -y: O) 
ci.i olherwise 
Not:' /hat il' lit(, noun associated wilh the dilm'n.- 
sion j wld('h vo-o('('urs \vilh c i also ('o-o('(.urs with 
Itoth o1 cp and c,i, Ihc valu('s o\[ lit:' ,\] tit dinwnsimt 
o1" ~'r~ and (,1. (V(L/ and vJi), art, tit(' same value, 
i.:'. the vaJm' ol' the '~ ~tr l>ol,ween thv ltOIltl ~tssol'i- 
at (,tl wil It t lw jilt dinwnsio(t and el. \]'~url:llerltlol:e. 
if I lw noun associated wil It I he dimensiolt j, which 
('o oc('l(rs with (,:, (loo,q llOt ('o-or':'(iv with \]):tilt v r 
and v,/, the vahu, of the 7tLtl, })t't\V('P(I {\]1:) ttOIt(( }IS'. 
sociated with the ./-tit (timcnsion and vl is set to 
111(' values o\[ tit(' .j4h dinwnuion of eft. it:re, wv 
call this value lit:' surplus value. \Ve l'Oca\]l that 
lit(' COml>a('tn('ss value of a groult of t'i and +',t is 
snmlh,r than thai: of +,; aud f,p. This nwans thai 
the \[ornwr is more cotnl)a('l Ihan the laltcr. If Lhe 
surphts vahw is (tot sot to l)oth c(~ and c J, tit(, 
group of c.t aim +', t is more('Omlm('t than that of 
v(i ;LIi(I v v. '\['hcrefor(,, ill ordor lo lit}d((' UrI }/1l(l 
+'/3 as symmetrical as possibh', tit(' surplus vaha, 
is set 1o eft. 
\]:'un:'liou l, mp(l'(t, i,i~) has the opltosite ('tl'e:'t of 
tit(' \[uncliOll splil(v i, I'p, uq), i.e. it uwrges (!(~ and 
v,). Function lump(col, vfl) returns *~i. 
l,mp(c~l, c,3) -- It; (9) 
I' i 
I'll 
U i' 2 
S.I. Vii 
UiH 
+'+~.i + c Jj it" cr~j-Tkl',~ j 
~,rt) olherwiso 
5.3 Flow of the Algorithm 
(:Hven a group of xerl)s, th, vu, "", c,,. the algorithm 
prodm'es a svt of somantic clusters, which are ordered 
iu Iho a~,ceueting oMer of thvh' senmntic coral)at:hess 
values. 1\[' +'i is non-.l)Olysemoum it lwlongs to at least 
(tit(, o\[ tilt' l'Osltltatlt Sellla.llti(' ('htst(,rH. If it is l)olyse - 
mous, the algorithm splits it inlo several hypolhetical 
verbs and each o\[' Ihom h¢longs lo at h'ast one of tlw 
soluatttic chlstcrs. The lhtw of lit:, algorithut is shuwn 
hi Figurc 2. 
As shown in Figure 2. tit(' algorithm is COml>osed 
of throe pro('odures: MakeqnitiabCluster-Set, Make-. 
Temporary-Cluster.-Set and Recognition-of-lOolysemy. 
1. Make-Initial-Cluster Set 
Tlw procedure Make+Initial-Cluster-Set l)rmh:('es 
all possibh' pairs o\[ verbs in Ill:' input with thcir 
sclnantic ('oltt\]t}Wlll:,ss values. Tho resull: is a llst 
765 
begin 
do M a ke-lnitiaI-Cluster-Set 
for i (1 < i < ,,,I,,,-Ii~ 2 
do Make-Temporary-CR~ster-Set ; 
if ,t set of ('lust(u's whi('h is r(qriev{,d I) 3 
Ma ke-Temporary-CI uster-Set {,xisl s 
then do Recognition-of-Polysemy: 
end_if 
store the newly obtail}(,d ('htsl('r ; 
if the n(,wly ol}taine(: chtstt,r ('(}ntains 
all the v('r\])s in input 
then exit front the loop ; 
end_if 
end_ibr 
end 
Figure 2: Th{' flow of th(' algorithm 
2. 
3. 
of pairs wl.i('h aro s(srt{'(l lit the ascon(ling or(l('r 
of their s(mlanti(' ('on}pa(:tnt,ss v;th}os. 'Fh(' list 
is called IC.S (Initial C.lusl(,r Set). 1CS contains 
,,(,,- 1) pairs. In th(, :F()I/-lo()l I in lho algorithm, 2 
a l)air (sf v('rlss is retri(,v{'d fronl ICS, (}n{' at ('a('h 
itt,ration, mM l)ass{,(1 to th(, next two pr(}('(,dur(,s. 
Make Temporary-Cluster-Set 
The l)roc(,(hu'(, tM((,s two argulll('llts: 'fit(, first ar- 
glllll('llt is a pair (1t' verbs froul ICS an(l the s('('- 
on(l on(' in a set (}f ('hlst('rs (C(!S - Crt'at('(l (?\]llS- 
t('l.' Sot). CCS C()llsists (5:\[ the ('lltsl('l's whi('h ll~tV(' 
I)(,en ('r(!~tt(!(l st) far. \Vh('n th{' algorithn} t(~:'mi - 
ll;Ltt's, CCS is th(, outllut of th{, algorithm. Make- 
Temporary-Cluster-Set :'t!lri(,vt,s tit(, (.htsl(,rs frolll 
CCS which ('ontain (me of th(' vcrl)s of th{, first 
argum(,nt (;t t)air f'r()m ICS). Th(~ ('htstt'rs thus l'O- 
tri('ve(l fr(}ln CCS al'(' 15asse(l to tit(, nexI l)r()('('(llll'O 
l"(/\]' further ('onsi(lt'ralion. If th(,r(, is n() CCS whi(.h 
('()itt~tilts oil(' (1t' th(' v(u'lls of a pair fronl IC!S, a pair 
of v('rbs from ICS in stored in CCS as a n{'wly ob- 
tain{'d {'lusi (u'. 
Recognition-of- Polysemy 
This procc(lure, which recogniscs a polysemous 
vt, r\]~, also tal,:(,s two ~trgult~('nts: th(' pair (}f v('rl/s 
from ICS and a set of chlst('rs :'('tri(,v('d l/y Make- 
Temporary-Cluster-Set. 
W(' r('('all the dist'ussi(m lit s(,('li()lt 5.1. Let {t', 
~t'l} I){' th(' pair of v(,rl)s frolll IC.S ~tlt(l { i,, ~t'2 } 1}0 
(5:1(, (5i' the ('lust(,rs (5t' the se('(/n(I &rglllll(,llt, i.r. the 
('lllS{:(!rs so f;u" ol)lain(,d whicl, (:onta.in (me ()f the 
V('I*\])S, ~! ill the p~Lir. We have to (l('t(u'n}ilw wh('ther 
t11(' \,orb v has two s{,nses, which ('(irr{,sllon(ls t(i 
u,, and w2, resltcctiv(qy. This is {l('t(!rlni:wd 1)y 
('Oml)a.ring the sont~tltli(' C(llnpa('tn('ss values ()f the 
thr('t' (liff{u'ent ('lust('rs shown in (6) and (7). Th{' 
,splitting fun(tion (8) is a l)l)lied to I,, aq, and u,2 
~tn(1 1)rothw('(l newly hyl)oth('ti('al v(u'lls, *q and 1,2. 
Tilt' l.wm,ping function (9) is al)pliod to vt and u2 
and lU~t\]:('s on(' verb ~, ()ut of th('m. If both of th(' 
S('lllallti(' ('(lllll)a('tll('SS vahl('S of ('a(:h sot sh(swl| ill 
(6) are smalh,r lhall :-i set shown ill (7), the srts (6) 
a.r(' s('h'('te(1, (}th(,rwis(,, (7) is scl(,('t(,(1 avd stored 
i,, CCS as a newly ol)taiu('d ('lnst(,r. 
If Bh(' newly ol)tain(,(l ('luster (lo(,s not contain all th(, 
verbs il} input, the n(,xt p~tir ()f v(,rl)s is l ak{!l} front lOS. 
~tll(\[ th("ii th(' whole 1)ro('css is l:Ol){'al('(l. 
6 Experiments 
We ha.re ('ondu('t('(l two OXl)orinl(,nls. The first ex- 
periment is ('on('ern{'d with the ('lust(wing te('hniqu(, 
~tn(l with verifying the eff(,t't ()f the l)r()l)t)s(,(l me/hod. 
The s('('oltd ('Xl/erilllOllt is ('Oll(hl('to(l to SOO h(}w vari(lU.S 
1)+trt-{sf-slle('('h 1)&il's ;tfl'{'('{ the (qust(!ring resntts. 
6.1 Data tbr the Experiments 
5\['h( ' ('orl)us we have us('(I is th{' Wall Str(¢et Jo'ur- 
'~tal whi('h consists of 2,878,688 o('('urr(,nc(,s of part-of- 
spet'('h taggo(1 wor(ls \[Chur('h, 1991\], 73,225 (liffor(mt 
woi'(ls. \]}'l'Olll this ('orl)lls, \vo (sbtain('(l 5,9,10,193 wor(l 
pairs in ~t window siz(' of 5 words, 2,743,974 (lillk,r(mt 
w()rd pairs. 
2{3 groups of v('rlss wet(' used in |lw ('Xl)orin~cnt:s, \]08 
verb tokons with 56 dif\[i'rcl}t original forn\] of verbs. 
~ti'li('s(, {gr(511lSS ('01ltailt i0 diff(u'rnt l)olyst,lnOUs \'orbs. 
Th(' groups of v('rlls are divid('(l into two diff{'r('nt 
tylst's, "tyl)('\]' and "@1}('2"; %yl)c\] ' is a sel: of v(,r\])s ('Oil- 
raining ()nr or mort, l/olys{m:ous v(,tbs, mt(I "tyl)('2' (loos 
not ('ontain any l)o|ys(,mous verbs. Ea('h group is co:n- 
1)os('(I of 3 to 10 ditf(,r(,nt v('rl/s. 'Fh(, seh,(:tion of v('rl)s 
of 't.yl)('l' was mad(' witll th(, illl('lltiolI of pro('{'ssing 
v(,rbs with wi(l(, usages, as i(h'}ltiti('(l in the Collins (li('- 
ti(511ary and thesaUlUlS \[~XI('\[,('o(I, 199l\]. Tht'n, a llltlll- 
I}or of syn(snyms of the (:ht)s(m verbs w('rc st'h't't('d from 
th(' th('sam'us. Thr ('hlst(?l'int, ~ analysis is al)lllie(1 to 
{,a('ll grtsu l} :sf v(,rbs. Tim SktllS{' ('OPI/US and tile gro}tl)S 
of verbs ~tt'(' uso(1 throughout th(, (,Xl)orin:t,nts. 
6.2 Experiment-I 
Ill \]~\]xD(~l'illl(}llt~\[, w(' llS('(i voFb-ll()llll pairs, i.e. w(' as- 
$51111{' all /t-(lilllOllSiOlt~tl Sl)a.('(,. ilt whi('h ~t verb is a,s- 
signed ~ reeler whos(' valu(' of the /-th (\[iln('nsiol} is 
Ill(' v,qhlo of mtt bi'lwe(!n tit(! vor}) a.lld the llOllll ~ts 
signed to the i-th axis. This ix l)ocauso, in tilt, small 
window sizes, Ill(, s(,}nantic relationshil)s between these 
two wet(Is mighl be quit(' strong, OSl)ecially those be- 
tween ~t verb and its object whir'i: l/elunits the eff{,ctivo 
re('og,fition of vorlml lmlysomy. The inflected forms of 
tl,c sam(, llOltllS ~tll(l vorI)s art' troat(,d ~ts single units. 
For oxa.lnl)l(,, "lilll(!'(lt()llll~ singular) an(l 'tiillOS'(noun, 
plural) are tl'o:%l od ,~ts sil}gh' milts. Wc obtained 228,665 
diD'rent vor})-nolln pairs from 2,7,13,974 and Dr)ill 
766 
tht's(', we seh,('ted 6,768 different vcrl)-liOllli pairs, 70:1 
dit\[(!rcnt w'rl)s alld 1,79(5 llolitis Oil condili<)u lhat fre- 
qllell('i('s a,lld 7//,'//, SI, Y(" llOl; hlw (,'V,,, > 5, .I Ill,r, .q) ~> 3) 
t</ pet'ntit ~L relial)le si:atis(i<'al analysis ;-/lilt tls('(l ~li('lll 
in lht, cxl>erhnent :l. Thc results are shown iu Tabh' 1. 
Tal)h, 1: '\].'he resuh.~ of \];\]Xl)erinl('ni-I 
_. -_ L--A~-~~:'~l\[ 'or! ('orr('('t ill('orr('('l~\] 
t,,,,,, - V t 
~,,,ll(%) ~>0()- fs(00.,) s(a(is) II__': .... l~S(';:)A __: I 
hi Tal)h' ;1~ 'groul>' uieaiis the nundler </t' each group, 
ly\])c\] and t.yl)e2; ~('(11'l'('Cl;' lll('&llS thc llllllt\])('l' ()\[' <gl't/ltl)S 
of verl)s which are <'lustcrt,({ c<)rrc('tly: "in('orrc('i" 
means lhai. they are not.. Figure 3 shows t!acL s:-/lll- 
ill( ` of Ihe results, i.e. tyt)el-c<)rrect, tyl)o2-correct. 
tyl)el-incorrect, a.n(I type2-incorrect. \];\]a('h valu(' 
iu Figure 3 shows the vahte of 111(' .SClllStilli(' ('Ollil)a('l;- 
tl('SS ()\[ ,h, g~l'Oll\]) ()\[ verbs. 
in lqgltre "3, under the heading tyl)el-correct, we 
uan set, thai 'lake' is re('ogn\]sed ns a p(ll)'SCltlOliS v0rb 
siJl(\[ lias lhre(' (liff('rent S('ltS('.'-J, 's|)('ltd', "btly', ali(I 
'ol>i:ain'. \[11 "/ similar way, "close' has two diffcrl,ul 
SOllSOS, 'olld' all(| 'opel1' S/lid s<,nianlically cl()sc v(ubs 
Stile grolll)('(\[ t.(){~(!th(,r. LTli(h'r Ih(' h('st(lilll{ type2- 
correct s(,nlanti('ally similar v('rbs are groupc(l l(/- 
gcther, ll<>wcver, un(ler l:he heading typel-incorrect 
'lcavt" is incorre('l:ly re<'og:iised as a n<)li-I)olysenious 
vcrl/; also under the heading tyl)e2-incorrect "('onlc" 
is in<'orrcctly re('ogl)ised as ;i l>olyscnl<>us verl/. 
6.3 Experiment-II 
Wt, have ('<>ndu('ted an exlwriuwtit; using t lw various 
i)arts=t)i~slwc('h sh(iwt) in Tal)lt' 2. 
'l'alllc 2: The tyl)e all(\[ the nundier <if pairs 
--i~:,~(o i~,.~5 ,, \]-7 .,-:, \]_: _ .J 
t~(;h:t--\,(-,r-17 --- 2r10,Ta2 T 07~7 ~-7;.<~ - \] 
v('rh-advcrb 23,248 / 1,200 L 2<'~ I a?-0\[ 
,<|,'(,rl.v<,,'ll I1 a :140t 007 / }.9,:I, I 
29 658 3 197 \[,:3:38 ,)8 verl,-preposition J~ ~ .,_ ~\[,338J__3~ d 
\[n Tabh, 2, x-y shows the t3'lle of 1)arl>ol:sl)ec('h l)air 
of .c and y in this order, wher(' ,raml y art' qlw I)art - 
of-sllet'('h ()\[' ill(' words. "pair(l)' shows lht' numl)cr 
of difli'rcnl: 1)art of-sl/ee('h pairs frmn 2,713,974 and 
"l)air(2)' shows th(' nuntber of different lmrbof-sl)ee('h 
l/ah's t)n ('ondition l:hat frequencies and 're, it re't. N,r~/ > 
5. m ,(.c, y) > 3; .r and y show the ntunber (if different 
word. We used Lhesc ill E:,:l)erinwnt-II. The r('suh s are 
shown in Table 3. 
:) IIOle, N~:¢; is tilt' l/tllllh(!l' o\[' loia\[ ('o OC'{'lll'i'i~ll( "t''~ o\[lht' WC>l'tls 
3' illld ,I\] ill I, tliS order ill ii ~qndow. 
Tal>le 3: The results of EXl>e)'hncnt-Ii 
."-9 ~ cori'e,'l(%) \] ili,'o,'rect(CX) \] 
ilOllll-V('l'\]) 11 \] \] 
,',,rb .<:,.,1:, tl / 2:(s0.s) 1 
.,t.t,r+.v.t..) II 20(rr.0) 
verl)-l)r('l)osil to:t_; .IL 5(')3,,~.,0) _ • __20(77'0)- 
7 Discussion 
In l':Xl)erinwnt L de,',crihe(1 ill the 1)r('vious seciiou, 18 
()Ill of 26 groups of verlL~ art, aualysed <'ort'ecily antl 
the percentage attains 60.2 I/ in all. flow(we,', as 
shown i,t Table I, there arc 8 {>Troul)a which could nol 
i)C II('('OglliS('(l ('(I\]'I'('('i\]~L 'i'll(' t'i'\[Ol'S ;ll'(' classificd into 
t,wo iyl)e~,: t \[. Error.~ of recot, nilion of imlysclnoUS 
vcrl)s as nonqmlysemous ones; and '2. \]';rmrs of re('og-. 
uiliou of IIOII-\]:.OIVsI'IlH)IIS VOI'\])S lit-; 1)O13"5Cl11OllS Oll('S, 
The IlllIlI1)(T Of gl'Olll/S classified iltlO each error type 
is \[ and 7, l'eSpeclivtqy. 'FIw causc o\[ Ihese crrol'S is 
lhaI, ('o-oct'l|rl'itlg ltOHllS shared by Iwo verbs sccm tt> 
1/(, slanled ill these data. For exanq)h', (/l)s(q'vill(p, tilt' 
('t)l'l)llS, W(' ('Sill St'(' thai "h'avc' }l.~ls sl.I \]('~lHt \[\VO %('IIS('S. 
'l'('tir(" all(l 'l't'ltlaill'. The Following scnt('n('es arc I'roln 
the W.,ll b'tvt:ct ,h~lvcl~,al. 
(s6) I,:aplat: l('fl his jol_) al warncr-laml)erl. 
(s6') A1)oui 12 ':/ hay(' rqtireAl front a full-time 
.jol,. 
(sT) '\['hey can even h,avt, a sticky l)rob|('m, in 
the \[orni of higher brokerage conuuissions. 
(sT') but l'Cmain a. serious llro!flen A. 
l\[t)wevcr, tyl)el-incorrecf, ill l"it~;ure 3 shmv<~ that 
'leave is incorre('l},'~ r('cognised as a II(/iI-I)OI.V,qCiLI(/IIS 
verb. This error wa,~ caust,d I)y Ihc fa('t \[hat lhl' vahlc 
(/\[ tilt' S('ltKlllti(" ('Olll|/S/('lltt'SS of "r(,tirc' an(l "l't'lilaill' 
was sntalh'r I ban t hal of any oI h('l' l>air of words illi(l 
1/3 th(' |'acl thai Ill(' ('ardil:alily o\['a stq (IJ' li()lltlS whic\]: 
('o-oc('ur with "rt,lire" alt(I 'l'('iilaill" is larger thsl.ll IIH/I 
(1\[ Silly olht'r pair of words. \Ve 1)rovisionally ('on('htd(' 
f3tat, the use of verb-noun l)ail"s a\]on(' is 110| allpropriat(' 
for all the groups o\[ Vcl'l)s. 
hi Experintcnl-ll. th(, overall resulls are not as good 
a.s lhose of \];\]xperilnent l. }\]owever, wc could observe 
so111(, inte)'esting charat'le)'istics, uamely, some groups 
whi('h could lJol be anal3sed co)'recdy \]13 ' using verb 
noun pairs could lw analysed correctly 1 G" using verb. 
adverb pairs or vt'rb-i)rvl)osiliol) pairs. The rcsulls 
show that 3 itll| o\[ 8 grt)ups such as tyi)el-incorrect 
iu Figure 3 whi('h were incorrect in E.':l)crinwnt-I coul(| 
1)/' analysed corre('lly I)y using vcrll adverll pairs. Also, 
a.n t)the)' 3 groul)S su('h as type2-incorrect could bt' 
analysed ('orr(,('lly I)y using vcrl)-prel)osiiiou pairs. \~,c 
• I ~V(' lit) till{ ('ont;idor here getieral Ol'l'Ol'S o1' Selllgtllt, i(: cllts- 
ters, i.e. the cast, (d' Iwo verbs which are illll <,(,tmttgical/y clo<,c 
lint m'<' jml~4ed to ((lllSliltlJ(, ;t <-,t'lll}lllJit' el|isle\[', ll, ecause lifts 
kind o\['crl'()r did not occur ill Ihe cllt'te)ll oxpclillll'llls. 
767 
0.700 0.800 0.900 typel-correct 
I 1 I , 
end end 0,738 0,831 
clo~el --1 
c\]oSe<close 2 0,795 882 I-- 
open -- open J O. 905 
spend -- spend ~_~( 
_~ takel -- 
take take2 0,8~ .905 
t.ake3 ~m~ 0,912 
buy -- buy 
obtain-- obtain \] ~0.915 
cancel- canoe\]. 918 
rJde~ ride 
solve-- solve 
typel-lncorrect 0,600 0.700 0,800 0.900 
_l I I I 
I leave % 
retire I t O. " 
I remain -__I__ 
& • 0.774 
borrow \]__ 
lend 
type2-correct 0.500 0.700 0.900 
I I I, 
C Olll(? --- 
begin - 
0.579 
ncreas( ~ 
r edtlc:e 
/ 0 ~:~25 
\] buy 
0. 753 
typo2-incorrect 0 . 600 C. 800 0 . 900 
-- I I I , 0. 370 
I fee\]. --fee\] ~ ~ 0 9~ 4 x ;::m:l. 
. - 0,928 
87- come -I-- come2 -- ---:-.~'~ I I 0 8 
" " ' .93 
X ..... T <ore ......... , 1 
go -- go 
sound -- C'olln(~- 
Figure 3: The results of ~he clusterillg analysis 
can therefore exl)ect that w(, may l)e abh' to ol)tain 
more ac('ur;tte (:lusters by merging tiles(, thr(,e kinds 
of part-of-speech 1)airs into one larg(,r set. Because 
~hese three difhwent 1)airs show distinct chara('t('ris~i('s 
of contexts in which a verb al)l)eacs. \'Ve have b(,(,n 
(ondu('dng more experiments (m these. 
8 Conclusion 
We have given a defilfition of polysemy from the view- 
l)oint of ('hLstering, ~tlt(l 1)rol)osed an overlttl)l)ilLg ('hLs- 
t('riug method which a utoma, tically recognises verbal 
1)olysemies from a textual corpus. The signifi('anl fca.- 
lure of our al)t)roach is that every Sel)~u'at(' meanhig 
of a word is recognised in terms of olh(u' words tlmt 
al)l)ea.r in the corl)us. \'Vh(,ther a word is polysenLous 
or not del)ends on whetlwr a set of oth(,r words ex- 
isfs whose lts~tge ('OlT(~Sl)Oll(Ls to OllP of {h(' lll('t/l\[ilt~S 
of a. 1)olys(qLIOllS word. As ~t l'Pslllt, ()ILl' lll(qhod ('all 
avoid human intuition in tlt(, judgelnent of distinct 
VCOI'(~ lltCallhlgS ~tll(| t, hlls, ~/lllllall illt('rv(qltlOlL. 
The results of the exl)eriments demonstrate the ap- 
1)li('ability of autolnad(' medmd of recognition of pof 
},'b('lllOllS vPrl)s. WP lntve ('OLLdlL('t0d lllOro ('x1)Pl'illl(qlls 
by ('halLging t)al'alneters su('h as the threshold values 
for fl'e(luen(Jes (N<,;) and m.'lt (mu(x,!/))in oI'dPI' tO 
see how these l)a.ranletel'S affect the l)erfornlml('( ~ of the 
('lustering algorithm. We have also extended our tech- 
ld(lUe to the disambiguation of word senses. Vfe hol)e 
lo report these results soon. 
References 
Church, I<<uuwd~ \\'. el al.. 1991. f-sing Stadsti<'s in Lex 
hal Analysis L~ J:ic,d o.cq,,isitio'n: E:~:ploiting ov,-lh~.c rr> 
~o',rces to bu, ihl a. le:~:ico'n. (Zcrnik Uri (cd.)), 1)~ges 115 
164. London. Lawrel,ce \]'2rll)amn Associates, 
Hindle Donald. 1990. Noun classification from predicate- 
argunwnt 'q\[l'lt('i,ll'CS. Iu 28~t~ A'n,n,.a.l Meeting o.f the 
Association for Comp'l~tationa.l Li'n,gu'i,qtics, pages 268 
275. Pittshm'gh, Ibumsyh,ania, Association fer Compu- 
~ationai Linguistics, N'\]~orrJsiewn..New Jersey. 
Smadja, Frank A.. 1993. Retrieving Collocations front 
Texl: Xtract. In Com, l~,.tati(m, ad Liwpn,~tic,% pages 143- 
177, 19 (1993) 1. 
Zernik Uri. 1991. '\].'rainl vs. Trah,2: Tagging Word S('nses 
in Corlms, Le:cica, I a.cq,.i,~itio'.: E:~Tdalti.,g on-6n, r,- 
.~o',.t'ccs to build (z l,:~:icon, Uri Zoufik(Ed.). l)ag('s 91-112. 
London. Lawrence Erlbannx Associates. 
Browll. Pet('r F. (,~ al.. 1991. V\'erd-S(,nse \])isand)iguatiol~ 
Using Statistical Melhods. In lb'oc, of the 29th Ann.,at 
Meet,\[rig o,f the A UL, lmg('s 26-t-270, 
Jardin(' N. and Sibson R,. 1968. The constrlwt,ion of hi- 
erarchic and nolPhieral'chic clasMih'allens, \[n Comp,.t(, 
dou, rna, l, lmgCs :\[77-184, 
Chur('h~ Kenneth W.. 1988. A stochastic parts 1)rogl'Hllltll(' 
and nora1 l)hraser for miresh'h't('d text. In Proc. of the 
2nd conference (m. a.pplied n,a~ttra.l lo.'n.yu.,.g~: proce,~,~in 9, 
pages 136-143. 
McLcod, William T.. 1987. The new Collin,, &ctio'm<~/ and 
the.qa'u.r,.~ i'n on(: vol,.m,. London. Harl)erCollins Pub- 
lishers. 
768 
