A Statistical Approach to the Processing of Metonymy 
Masao Utiyama, Masaki Murata, and Hitoshi Isahara 
Communications Research L~boratory, MPT, 
588-2, Iwaoka, Nishi-ku, Kobe, Hyogo 651-2492 Japal 
{mutiyam~,murat~,isahara} ~crl.go.j p 
Abstract 
This paper describes a statistical approach to 
tile interpretation of metonymy. A metonymy 
is received as an input, then its possible inter- 
p retations are ranked by al)t)lying ~ statistical 
measure. The method has been tested experi- 
mentally. It; correctly interpreted 53 out of 75 
metonymies in Jat)anese. 
1 Introduction 
Metonymy is a figure of st)eech in which tile 
name of one thing is substituted for that of 
something to which it is related. The czplicit 
tc.~m is 'the name of one thing' and the implicit 
t;c~"m is 'the name of something to which it; is 
related'. A typical examt)le of m(;tonymy is 
He read Shal(esl)eare. (1) 
'Slmkesl)(~are' is substitut(~d for 'the works of 
Shakespeare'. 'Shakest)eare' is the explicit term 
and 'works' is the implicit term. 
Metonymy is pervasive in natural . 
The correc~ treatment of lnetonylny is vital tbr 
natural  l)rocessing api)lications , es- 
1)ecially for machine translation (Kamei and 
Wakao, 19!)2; Fass, 1997). A metonymy may be 
aecel)table in a source  but unaccet)t- 
able in a target . For example, a direct 
translation of 'he read Mao', which is acceptable 
in English an(1 Japanese, is comt)letely unac- 
ceptal)le in Chinese (Kamei and Wakao, 1992). 
In such cases, the machine trmlslation system 
has to interl)ret metonynfies to generate accept- 
able translations. 
Previous approaches to processing lnetonymy 
have used hand-constructed ontologies or se- 
mantic networks (.\]?ass, 1988; Iverson and Hehn- 
reich, 1992; B(maud et al., 1996; Fass, 1997). 1 
1As for metal)her l)rocessing, I,'errari (1996) used t;ex- 
Such al)t)roaches are restricted by the knowl- 
edge bases they use, and may only be applicable 
to domain-specific tasks because the construc- 
tion of large knowledge bases could be very dif 
ficult. 
The method outlined in this I)apcr, on the 
other hand, uses cortms statistics to interpret 
metonymy, so that ~ variety of metonynfies 
can be handled without using hand-constructed 
knowledge bases. The method is quite t)romis- 
ing as shown by the exl)erimental results given 
in section 5. 
2 Recognition and Interpretation 
Two main steps, recognition and i'ntc.'q~vc- 
ration, are involved in the processing of 
metonyn~y (Fass, 1.!)97). in tile recognition st;el), 
metonylnic exl)ressions are labeled. 1111 the in- 
tel'l)r(:tation st;el) , the meanings of those ext)res- 
sions me int, eri)reted. 
Sentence (1), for examl)le, is first recognized 
as a metonymy an(t ~Shakespeare' is identified 
as the explicit term. 't'he interpretation 'works' 
is selected as an implicit term and 'Shakespeare' 
is replaced 1)y 'the works of Shakespeare'. 
A conq)rehensive survey by Fass (\]997) shows 
that the most COllllllOll metho(1 of recogniz- 
ing metonymies is by selection-restriction vio- 
lations. Whether or not statistical approaches 
can recognize metonymy as well as the selection- 
restriction violation method is an interesting 
question. Our concern here, however, is the 
interpretation of metonymy, so we leave that 
question for a future work. 
In interpretation, an implicit term (or terms) 
that is (are) related to the explicit term is (are) 
selected. The method described in this paper 
uses corpus st~tistics for interpretation. 
tual clues obtained through corl)us mmlysis tor detecting 
metal)lmrs. 
885 
This method, as applied to Japanese 
metonymies, receives a metonymy in a phrase 
of the tbnn 'Noun A Case-Marker R Predicate 
V' and returns a list of nouns ranked in or- 
der of the system's estimate of their suitability 
as interpretations of the metonylny, aSSulning 
that noun A is the explicit tenn. For exam- 
ple, given For'a a wo (accusative-case) kau (buy) 
(buy a Ford),  Vay .sya (ear), V .st .sdl  , 
r'uma (vehicle), etc. are returned, in that order. 
Tile method fbllows tile procedure outlined 
below to interpret a inetonymy. 
1. Given a metonymy in the form 'Noun A 
Case-Marker R Predicate V', nouns that 
can 1)e syntactically related to the explicit 
term A are extracted from a corpus. 
2. The extracted nouns are rmlked according 
to their appropriateness as interpretations 
of the metonymy by applying a statistical 
measure. 
The first step is discussed in section 3 and the 
second in section 4. 
3 Information Source 
\¥e use a large corpus to extract nouns which 
can be syntactically related to the exl)licit term 
of a metonylny. A large corpus is vahmble as a 
source of such nouns (Church and Hanks, 1990; 
Brown et al., 1992). 
We used Japanese noun phrases of the fornl 
A no B to extract nouns that were syntactically 
related to A. Nouns in such a syntactic relation 
are usually close semantic relatives of each other 
(Murata et al., 1999), and occur relatively infre- 
quently. We thus also used an A near B rela- 
tion, i.e. identifying tile other nouns within the 
target sentence, to extract nouns that may be 
more loosely related to A, trot occur more fre- 
quently. These two types of syntactic relation 
are treated differently by the statistical nleasure 
which we will discuss in section 4. 
The Japanese noun phrase A no B roughly 
corresponds to the English noun phrase B of A, 
lint it has a nmch broader range of usage (Kuro- 
hashi and Sakai, 1999). In fact, d no B can ex- 
press most of the possible types of semmltic re- 
lation between two nouns including metonymic 
2~Ford' is spelled qtSdo' ill Japanese. We have used 
English when we spell Japanese loan-words from English 
for the sake of readability. 
concepts such as that the name of a container 
can represent its contents and the name of an 
artist can imply an art~brnl (container for 
contents and artist for artform below).a Ex- 
amples of these and similar types of metonymic 
concepts (Lakoff and Johnson, 1980; Fass, 1997) 
are given below. 
Container for contents 
• glass no mizu (water) 
• naV  (pot) , y6 i (food) 
Artist for artform 
• Beethoven no kyoku (music) 
• Picas.so no e (painting) 
Object for user 
• ham .sandwich no kyaku (customer) 
• sax no .sO.sya (t)erformer) 
Whole tbr part 
• kuruma (car) no tirc 
• door" no knob 
These exalnt)les suggest that we can extract 
semantically related nouns by using tile A no B 
relation. 
4 Statistical Measure 
A nletonymy 'Noun A Case-Marker R, Predi- 
cate V' can be regarded as a contraction of 
'Noun A Syntactic-Relation (2 Noun B Case- 
Marker R Predicate V', where A has relation 
Q to B (Yamamoto et al., 1998). For exam- 
ple, Shakc.spcare wo yomu (read) (read Shake- 
speare) is regarded as a contraction of Shake- 
speare no .sakuhin (works) 'wo yomu (read the 
works of Shakespeare), where A=Shake.spcare, 
Q=no, B=.sakuhin, R=wo, and V=yomu. 
Given a metonymy in the fbrln A R 17, the 
appropriateness of noun B as an interpretation 
of the metonymy under the syntactic relation Q 
is defined by 
LQ(BIA,/~, V) - Pr(BIA, (2, 1~, V), (2) 
ayamamoto et al. (\]998) also used A no /3 relation 
to interpret metonymy. 
886 
where Pr(.-.) represents l)robal/ility and Q is 
either an A no B relation or an A near \]3 re- 
lation. Next;, the appropriateness of noun \]3 is 
defined by 
M(BIA, Ie, V) -nlaxLc~(BIA, l~,V ). (3) O 
We rank nouns 1)y at)plying the measure 214. 
Equation (2) can be decomposed as follows: 
LQ(!31A, R,, V) 
= Pr(BIA , Q, R,, V) 
Pr(A, Q, B, R,, V) 
Pr( A, Q, R, v) 
Pr(A, Q, 13)lh'(R, VIA, Q, Ix) 
Pr(A, Q) Pr(R, VIA, Q) 
Pr(BIA , Q)Pr(R, VIB) 
-~ er(R, v) ' (4) 
where (A, O) and {\]~,, V} are assumed to l)e in- 
del)endent of each other. 
Let f(event)1)e the frequen(:y of an cve'nt and 
Classc.s(\])) be the set of semantic (:lasses to 
which B belongs. 'l'he expressions in Equation 
(4) are then detined t)y 4 
I'r(~lA, Q) - .t'(A, Q, ~x) _ f(A, Q, ~) 
f(A, Q) ~1~ f(A, Q, 13)' (5) 
Pr(~., riB) 
IU~,I~,v) i' ' *: .1 (U, ~, V) > 0, 
.~- ~,c~cl .......... (10 Pr(l)'l(/)f(C/'R'V) 
J'US) 
otherwise, 
((0 
Pr(BIC ) - .f(13)/ICI-s.w-.XB)l j(c) (r) 
We onfitted Pr(H,, 17) fi'om Equation (4) whell 
we calculated Equation (3) in the experiment 
de, scribed in section 5 for the sake of simplicit> 
4Strictly speaking, Equation (6) does not satist\]y 
X',,e,vpr(R, vl/x) -- 1. We h~wc adopted this det- 
inition for the sake of simplicity. This simplifi- 
cation has little effect on the tilml results because 
~--;c'cc~ ........ (m Pr(l~lC)f(C,I~', V) << I will usually 
hohl. More Sol)histieated methods (M;mning ml(t 
Schiitze, 1999) of smoothing f)robability distribution 
m~y I)e I)eneticial. itowever, al)l)lying such methods 
and comparing their effects on the interpretation of 
metonymy is beyond the scope of this l)aper. 
This treatment does not alter the order of the 
nouns ranked by the syst;em because l?r(H., V) 
is a constant for a given metonymy of the form 
AR V. 
Equations (5) and (6) difl'er in their treatment 
of zero frequency nouns. In Equation (5), a 
noun B such that f(A, Q, B) = 0 will l)e ignored 
(assigned a zero probal)ility) because it is un- 
likely that such a noml will have a close relation- 
shii / with noun A. In Equation (6), on the other 
hand, a noun B such that f(B, R, V) = 0 is as- 
signed a non-zero probability. These treatments 
reflect the asymmetrical proper~y of inetonymy, 
i.e. ill a nletonylny of the form A 1{ 1~ an 
implicit term 13 will have a much tighter rela- 
tionship with the explicit term A than with the 
predicate V. Consequently, a nouil \]3 such that 
f(A,Q, B) >> 0 A f(B, JR, V) = 0 may be ap- 
propri~te as an interpretation of the metonymy. 
Therefore, a non-zero t)robat)ility should be as- 
sign(;d to Pr(l~., VI1X ) ev~,n it' I(B, 2e, V) ; (). ~ 
Equation (7) is the probability that noun J3 
occurs as a member of (::lass C. This is reduced to 
fU~) if13 is not ambiguous, i.e. IC/a,~,sc.,s,(/3)\[ = f(c) 
1. If it is ambiguous, then f(B) is distributed 
equally to all classes in Classes(B). 
The frequency of class C is ol)tained simi- 
larly: 
.f(B) (8) 
.f(c) = ~ ICl(-~c..~(13)1' 11C-.(7 
where 13 is a noun which belongs to the class C. 
Finally we derive 
f(13, ~, v) 
BqC 
(.0) 
In summary, we use the measure M as de- 
fined in Equation (3), and cah:ulated by apply- 
ing Equation (4) to Equation (9), to rank nouns 
according to their apl)ropriateness as possible 
interpretations of a metonymy. 
Example Given the statistics below, bottle we 
akeru (open) (open a bottle) will be interpreted 
5The use of Equation (6) takes into account a noun/3 
such that J'(l:~, l{, V) = 0. But, Stlch & llOtlll is usually ig- 
nored if there is another noun B' such that f(13', H., V) > 
0 be~,~,,se. Eo'~ct ....... U~)P, USIO)J'(C,~e.,V) << a < 
J'(lY, H,, V) will usually hokl. This means thai the co- 
occurrence 1)rol)al)iliW between implicit terms and verbs 
are also important in eliminating inapl)rol)riate nomls. 
887 
as described in the fbllowing t)aragraphs, assum- 
ing that cap and rcizSko (refl'igerator) are the 
candidate implicit terms. 
Statistics: 
f(bottlc, no, cap) = 1, 
f(bottlc, no, reizgko) = O, 
f(bottlc, no) = 2, 
f ( bottlc, ncar, cap) = 1, 
f (bottle, near, rciz6ko) = 2, 
f(bottlc, ncar) = 503, 
f(cap) = 478, 
f(rcizSko) = 1521, 
f(cap, wo, akcru) = 8, and 
f(rciz6ko, wo, akcru) = 23. 
f(bottlc, no, rciz6ko) = 0 indicates that bottle 
and rcizSko are not close semantic relatives of 
each other. This shows the effectiveness of us- 
ing A no B relation to filter out loosely related 
words. 
Measure: 
L,o(cap) 
Lncar(Cap) = 
Lno(reizSko) = 
Lncar ( reizS ko ) -~ 
f ( bott:le, no, cap) 
.f ( bottlc, no) 
\](ca,p, wo, a\]~c'ru) 
X 
1 8 -8.37×10 -3 , 
2 478 
f (bottle, near, cap) 
f(bottlc, near) 
f ( caI), "wo, a\]~cru) 
X 
.f ( ) 
1 8 
50--3 47-8 = 3.33 × 10 -5, 
.f ( bottlc, no, rcizSko ) 
.f ( bottlc, no) 
f ( rcizako, wo, ahcru ) × 
.f ( rcizdko 
0 23 
2 1521 
.f ( bottlc, near, rcizSko) 
f (bottlc, near) 
f(rcizSko, wo, akcru) 
X f ( rciz~ko ) 
2 23 
503 1521 - 6.01 x 1() -~, 
M(c p) 
= max{Lno(cap),Lnea.,.(cap)} 
= 8.37 x lO-3, and 
~r ( reizSko ) 
= 6.01× 10 -5 , 
where L,,o(Cap) = L,~o(Caplbo~tle, wo, akeru), 
M(c p) = M(c pl ot tz , and so o51. 
Since M > M we conclude 
that cap is a more appropriate imt)licit term 
than rcizSho. This conclusion agrees with our 
intuition. 
5 Experiment 
5.1 Material 
Metonymies Seventy-five lnetonymies were 
used in an ext)erilnent to test tile prol)osed 
lnethod. Sixty-two of them were collected from 
literature oll cognitive linguistics (Yamanashi, 
1988; Yamam~shi, 1995) and psycholinguistics 
(Kusumi, 1995) in Japanese, paying attention 
so that the types of metonymy were sufficiently 
diverse. The remaining 13 metonymies were 
direct translations of the English metonymies 
listed in (Kalnei and Wakao, 1992). These 13 
metonylnies are shown in Table 2, along with 
the results of the experiment. 
Corpus A corpus which consists of seven 
years of issues of the Mainichi Newspaper (Dora 
1991 to 1997) was used in the experiment. The 
sentences in tlle cortms were mort)hologically 
analyzed by ChaSen version 2.0b6 (Matsumoto 
et al., 1999). The corpus consists of about 153 
million words. 
Semantic Class A Japanese thesaurus, Bun- 
rui Goi-tty6 (The N~tional Language Research 
Institute, 1996), was used in the experiment. It 
has a six-layered hierarchy of abstractions and 
contains more than 55,000 nouns. A class was 
defined as a set of nouns which are classified in 
the same abstractions in the top three layers. 
The total nmnber of classes thus obtained was 
43. If a noun was not listed in the thesaurus, it 
was regarded as being in a class of its own. 
888 
5.2 Method 
'.1.11(; method we have dcseril)e,d was applied I;O 
the metonynfie, s (lescril)e,(t ill section 5.1. Tile 
1)r()eedure described 1)clew was followed in in- 
tert)rel;ing a metonynly. 
1. Given a mel,onymy of the, form :Noun A 
Case-Marker R Predicate, V', nouns re- 
\]al;e(l to A 1)y A 'n,o .1:1 relation an(l/or A 
near H relation were extra(:ix'~(l from 1;he, 
corl)us described in Se(:tion 5.\]. 
2. The exl;racted llOllllS @an(lidatcs) were 
ranked acc()rding t() the nw, asure M d(;tined 
in \]{quation (3). 
5.3 Results 
The r(;sult of at)l)lying the proi)osexl me, thod to 
our sol; of metol~ymies is summarized in 'l'alfle 
1. A reasonably good result (:an 1)e s(;cn for 
q)oi;h r(,\]ai;ions', i.e. l;he result ot)i;aincd \])y us- 
ing both A no 11 an(t d ncm" 1\] l'elal;ion~; wllen 
extracting nouus fl'onl th(' cOllmS, \[1'1~(', a(:(:u- 
ra(:y of q)ol;h re, l~tions', the ratio ()f lhe nllnil)er 
of (:orrc(:l;ly intcrl)r(;te,(1 (; t()l)-rank(;(l (:an(li(lates 
to l;he, total mmfl)er of m(',l;()nymies in ()it\]' set, 
w,,s 0.7:, (=5',Visa+22)) alld ('ol,ti(t(' l,ce 
inWwva.1 estimal;e was t)(;l;ween ().6\] an(t 0.8\].. 
\¥e regard this result as quite t)ronfising. 
Since the mc, i;onymies we used wcr(; g(m(u'a\]: 
(lomain-in(lel)(',ndca~t, on(s, l;h(~ (legr(', ~, ()f a(:cu- 
racy achi(;ve, l in this (~xp(;rim(;nt i~; likely t() t)(; 
r(',t)(',al;e(l when our me£hod is ~q)l)lie(l t() oth(;r 
genural sets ()f mel;onymies. 
'.\['~l)l(; l : tt3xl)erimental r('sults. 
I{,elal;ions used Corre(;t \¥'rong 
Both relations 53 22 
Only A 'no B 50 25 
Only A near 13 d3 32 
Tal)le 1 also shows that 'both relations' is 
more ae(:ural;e than (',il;her the result obtained 
1)y solely using the A no \]3 relation or the A 
near B relation. The use of multit)le relations 
in mel, onyn~y int(;rl)retation is I;hus seen to l)e 
1)enefieial. 
aThe correct;hess was judged by the authors. A candi- 
dat(; was judged correct when it; made sense in .Ial)anese. 
For examl)le, we rcgard(;d bet:r, cola, all(l mizu (W;d;el') 
as all (:orr(!c\[; intcrl)r(~l;ations R)r glas.s we nom, u (drink) 
(drink a glass) because lhey llla(le ,q(~llSC in some (:ontcxt. 
Table 2 shows the, results of applying the 
method to the, thirteen directly translated 
metonymies dcscril)ed in sect;ion 5.1.. Aster- 
isks (*) in the tirst (;ohlillll indicate that direct 
translation of the sentences result in unaccel)t- 
able Japanes(;. The, C's and W's in t;he sec- 
ond eohmm respectively indicate that the top- 
ranked ('andi(latcs were correct and wrong. The 
s(;nten(:es in the l;hir(t column are the original 
English metonymi(;s adol)tc, d fl'om (Kamci and 
\¥akao, t992). The Japanese llletollylllies in 
th(: form hloun ease-lnarker predi(:ate 7', in the 
fourth column, are the illputs I;o the method. 
In this ('ohunn, we and 9 a mainly r(;present 
I;he ac(:usal;ive-casc and nominative-ease, re- 
Sl)ectively. The nouns listed in the last eolmnn 
m'e the tot) three candidates, in order, according 
to the. measure M that was defined ill Equation 
(3). 
Th(,,se, l'csull;s (lemonstrate the et\[~(:tiveness of 
lhe m(',thod. '.l>n out of t;11(: 13 m(;tonynfies 
w(u'c intc, rt)rete,(l (:orre, ctly. Moreover, if we 
rcsl;ri(:t our al;l;(',nti()n to the ten nietonylHics 
i}mt m'e a(:(:Cl)tal)le, ill ,/al)anese, all l)ut one 
w(;rc, inl;('rl)r(;te(t (:orrectly. The a(:curacy was 
0.9 ---- (/)/\]0), higher than that for q)oth rela- 
tions' in Tal)le i. The reason fi)r the higher de- 
gl'ee of ac(:tlra(;y is l;\]lal; the lll(;|;Ollyllli(;s in Tal)le 
2 arc semi,what tyi)ical and relativ(;ly easy to 
int(~rl)rel; , while, the lnel;(nlynlics (:olle(:l;c(t fl'()m 
,lal)anese sour(:es included a (liversity of l;yl)es 
and wcr(~ more difficult to intext)let. 
Finally, 1;11(', efl'ecl;iv(umss of using scnlanl;i(: 
classes is discussed. The, l;op candidates ot! six 
out of the 75 metonynfies were assigned their 
al)prot)riatenc, ss by using their semantic classes, 
i.e. the wducs of 1;11o measure 114 was calculated 
with f(H,/~, V) = 0 in lgquat;ion (6). Of the, se, 
l;hrce were corrccl,. 011 l;hc, other hand, if sc- 
manl;ic class is not use(l, then three of the six 
are still COITeC|;. Here there was no lint)rove- 
merit. However, when we surveyed the results 
of the whole experiment, wc found that nouns 
for wlfich .fiB, R,, V) -- 0 often lind (:lose re- 
lationship with exl)licit terms ill m(;tonynfics 
and were al)propriate as interpretations of the 
metonynfics. We need more research betbre we 
(:an ju(lgc the etl'ectivc, ness of utilizing semantic 
classes. 
rPl'edicatcs are lemmatized. 
889 
Table 2: Results of applying the proposed lnethod to direct translations of the metonymies in 
(Kanmi and Wakao, 1992). 
Sentences Noun Case-Mm'l~er Pred. Candidates 
C Dave drank the glasses. 
C The .kettle is boiling. 
C Ile bought a Ford. 
C lie has got a Pieasso in his room. 
C Atom read Stcinbeck. 
C 
C 
W 
C 
W 
C 
Ted played J3ach. 
Ite read Mao. 
We need a couple of strong bodies 
tbr our team. 
There a r___q a lot of good heads in the 
university. 
Exxon has raised its price again. 
glass we nomu 
yakan ga waku 
Ford we kau 
Picasso we motu 
Stcinbcck we yomu 
Bach we hiku 
Mao we yomu 
karada ga hituy5 
atama ga iru 
Exxon 9 a agcru 
Washington is insensitive to the 
needs of the people. 
Washington ga musinkci 
C The T.V. said it was very crowded 
at; the festival. 
W The sign said fishing was prohibited 
here. 
T. V. 9a in 
hy&siki ga iu 
beer, cola, mizu (water) 
yu (hot water), 
oyu (hot water), 
nett5 (boiling water) 
zy@Ssya (car), best seller, 
kuruma (vehicle) 
c (painting), image, aizin (love,') 
gensaku (original work), 
mcisaku (fmnous story), 
daihySsaku (important work) 
mcnuetto (minuet), kyoku (music), 
piano 
si (poem), tyosyo (writings), 
tyosaku (writings) 
carc, ky~tsoku (rest;), 
kaigo (nursing) 
hire (person),tomodati (friend), 
bySnin (sick person) 
Nihon ( Japan) ,ziko (accident), 
kigy5 (company) 
zikanho (assistant vice-minister), 
scikai (political world), 
9ikai (Congress) 
cotnlllentgtol'~ anllOllllcer I (:~stel" 
mawari (surrmmding), 
zugara (design) 
.seibi (lnaintclmnce) 
6 Discussion 
Semantic Relation The method proposed in 
this pnper identifies implicit terms fbr tile ex- 
plicit term in a metonymy. However, it is not 
concerned with the semantic relation between 
an explicit; term and implicit term, because such 
semantic relations are not directly expressed ill 
corpora, i.e. noun phrases of the form A no 
B can be found in corpora bul; their senmntic 
relations are not. If we need such semantic re- 
lations, we must semantically analyze the noun 
phrases (Kurohashi and Sakai, 1999). 
Applicability to other s Japan- 
ese noun phrases of the form A no B are specitie 
to Japanese. The proposed method, however, 
could easily be extended to other s. For 
exmnple, in English, noun phrases B of d could 
be used to extract semantically related nouns. 
Nouns related by is-a relations or part-of re- 
lations could also be extracted from corpora 
(Hearst, 1992; Berland and Charniak, 1999). If 
such semantically related nouns are extracted, 
then they can be ranked according to the mea- 
sure M defined in Equation (3). 
Lexically based approaches Generative 
Lexicon theory (Pustejovsky, 1995) proposed 
the qualia structure which encodes semantic re- 
lations among words explicitly. It is useflfl to 
infer an implicit term of the explicit term in 
a metonymy. The proposed approach, on the 
other hand, uses corpora to infer implicit terms 
and thus sidesteps the construction of qualia 
structure. 8 
7 Conclusion 
This paper discussed a statistical approach to 
the interpretation of metonymy. The method 
tbllows the procedure described below to inter- 
pret a metonymy in Japanese: 
1. Given a metonymy of the tbrm 'Noun A 
SBriscoe et al. (1990) discusses the use o1" machine- 
readable dictionaries and corpora for acquMng lexical 
semantic information. 
890 
Case-Marker 1{ Predicate V', nouns that 
are syntactically related to the explicit 
terlll A are extracted front a corpus. 
'.2. The extracted nouns are ranked according 
to their degree of appropriateness as inter- 
pretations of the metonymy by applying a 
statistical measure. 
The method has been tested experimentally. 
Fifty-three out of seventy-five metonymies were 
correctly interpreted. This is quite a prolnis- 
ing first; step towm'd the statistical processing 
of metonymy. 

References 

Matthew Berland and Eugene Charniak. 1999. 
Finding parts in very large corpora. In A (7L- 
99, pages 57- 64. 

Jacques Bouaud, Bruno Bachimont, and Pierre 
Zwcigenbaum. 1996. Processing nletonyllly: 
a domain-model heuristic graph travcrsal 3t> 
preach. In COLINC-95, pages 137-142. 

Ted Briscoc, Ann Copestake, and Bran Bogu- 
racy. 1990. Enjoy the paper: L(;xi(:al seman- 
tics via lexicology. In COLING-90, pages 4:2-- 
4:7. 

I)(fi;cr F. l~rown, gincenl; ,l. Delia Pietra, Pe- 
ter V. deSouza, ,\]enifer C. \]~ai, m~d l/.ol)(',rl; I,. 
Mercer. 1992. Class-1)ased n-gram models of 
m~l;ur~l lmlguage. C~o'm,p'u, tat, ioruzl Li'n, guistics, 
1.8(4) :467 479. 

Kelmeth Ward Church and Patrick Hanks. 
1990. Word association norms, mutual in- 
formation, and lexicography. Uomputatio'n, al 
Lin.quistics, 16(1):22 29. 

Dan Fass. 1988. Metonymy and lnel;al)hor: 
What's the difference? In COLING-88, 
pages \]77-181. 

Dan Fass. 1997. Processin9 Mctonymy and 
Me.taph, or, volume 1 of Cont, cm.porar'y Studies 
in Cognitive Science and '\]'cch, nology. Ablcx 
Publishing Corporation. 

Steplmne Fcrrari. 1996. Using textual clues 
to improve metaphor processing. In ACL-95, 
pages 351-354. 

Marl;i A. Hearst. 1992. Automatic acquisition 
of hyponyms fi:om large text corpora. In 
COLING-92, pages 539 545. 

Eric iverson mid Stephen Helmreich. 1992. 
Metallel: An integrated approach to non- 
literal phrase interpretation. Computational 
Intelligence, 8(3):477 493. 

Shin-ichiro I(amei and Takahiro Wakao. 1992. 
Metonymy: Itcassessment, survey of accept- 
ability, and its treatment in a machine trans- 
lation system. In ACL-92, pages 309-311. 

Sadao Kurohashi and Yasuyuki Sakai. 1999. 
Semantic mmlysis of ,Japmmse noun phrases: 
A new approach to dictionary-lmsed under- 
standing. In ACL-99, pages 481 488. 

Takashi Kusumi. 1995. ttiyu-no S'yori-Katci- 
t;o lmi-Kdzfi (Pr'occssin 9 and Semantic Struc- 
ture of "\]'ropes). Kazama Pul)lisher. (in 
Jalmnese). 

George Lakoff and Mm'k Johnson. 1980. 
Meta, phors lye Live By. Chicago University 
Press. 

Christopher D. Mmming and Hinrich Schiitze, 
1999. Fou'ndations of Statistical Nat.ur(d Lan- 
guage \])recessing, chapter 6. The MIT Press. 

Yuji Matsmnoto, Akira Kitauchi, Tatsuo 
Yamashita, and Yoshitalm Hirano. 1999. 
Japanese morphological anMysis system 
ChaScn mmmal. Nara Institute of Science 
and Technology. 

Masaki Murata, Hitoshi Isalmra, and Makoto 
Nagao. 1999. IX.csolut, ion of indirect anal)hera 
in Jal)anese s(;ntcn('es using examples "X no 
Y (X of Y)". In A 6%'99 Work.shop orl, Core/" 
e.'l'(:,ncc and It.s AppIica, tio'ns, 1)ages 31 38. 

,lames l'ustejovsky. 1995. 2Yt, c Generative Lex- 
icon. 'J?he MI'I' Press. 

Tim National Language I/.ese~rch lalstitute. 
1996. Bv, nr',ui Goi-hyO Z~h,o-bav,(Th:l;o'nom, y 
of ,lapo, nc.s'e., e'nla'ulcd cditio@. (in ,Japancse). 

Atsmnu Yammnoto. Masaki Murata, and 
Makoto Nagao. 1998. Example-based 
metonymy interpretation. In \])'roe. of the 
~t,h. Annual \]lgcel;in 9 of th, c Association for 
Natural Language Prwccssing, pages 606 609. 
(in Japanese). 

Masa-aki Yamanashi. 1988. Hiyu-to \]~ikai 
('1;ropes and Understanding). Tokyo Univer- 
sity Publisher. (in Jalmnese ). 

Masa-aki Yamalmshi. 1995. Ninti Bunpa-ron 
(Cognitive Linguistics). Hitsuji Publisher. 
(ill Japanese). 
