! 
! 
Deriving Metonymic Coercions from WordNet 
Sanda M. Harabagiu 
Artificial Intelligence Center 
SRI International 
333 Ravenswood Ave, Menlo Park, CA 94025 
harabagi@ai.sri.com 
Abstract 
This paper presents a method for deriving 
metonymic coercions from the knowledge available 
in WordNet. Two different classes of metonymies are 
inferred by using (1) lexico-semantic connections be- 
tween concepts or (2) morphological cues and logical 
formulae defining lexical concepts. In both cases the 
derivation of metonymic paths is based on approx- 
imations of sortal constraints retrieved from Word- 
Net. 
This novel method of inferring coercions validates 
the related knowledge through coreference links. As 
a result, metonymic coercions are potentially useful 
for the recognition of coreferring entities in informa- 
tion extraction systems. 
1 Problem description 
The pervasive phenomenon of metonymy raises a 
problem for the interpretation of real-world texts. 
Metonymies are figures of speech in which, according 
to the literature definition from (Lakoff and John- 
son, 1990), "one entity is used to refer to another, 
that is related to it". Characteristic of a metonymic 
reading of a textual expression is the fact that the 
satisfaction of sortal constraints guides the coercion 
to related knowledge. 
The comprehensive account of the semantics of 
meaning transfers presented in (Nunberg, 1995) in- 
dicates that coercions need to be embedded in a 
conceptual and lexico-semantic space, ideally pro- 
vided by a linguistic knowledge base. Nunberg also 
notes that coercions are licensed by pragmatic cir- 
cumstances, specifically pertaining to the Gricean 
principles (Grice, 1975). 
In this paper, we revisit the notion of metonymy 
and address the computational aspects of its res- 
olution in the context of the relational seman- 
tics provided by the recently released WordNet 1.6 
lexical database (www.cogsci.princeton.edu/~wn ). 
Following the lessons learned from the WordNet- 
based inference of Gricean implicatures, reported in 
(Harabagiu et al., 1996), a novel methodology of pro- 
ducing metonymic paths was devised. 
The coercions combine WordNet relations with st'- 
142 
mantic information derived from conceptual defini- 
tions. In WordNet (Miller, 1995) synony m words 
are structured in synsets, underlying a linguistic 
concept. Every synset is associated with a gloss, 
representing a textual definition, that can be trans- 
lated in a logical form following the notation intro- 
duced in (Hobbs, 1986-1). This formalism, used in 
the implementation of TACITUS (Hobbs, 1986-2), ac- 
commodates a large variety of discourse inferences 
and, moreover, provides an elegant manner of local- 
izing ambiguities, as was shown in (Bear and Hobbs, 
1988). 
Conceptual support from linguistic knowledge 
bases was already considered in the implementation 
of several metonymy resolution systems (e.g. (Mark- 
err and Hahn, 1997), (Fass, 1991) (Hobbs. 1986-2)), 
but none of these systems provided with more in- 
ferential flexibility than the typical coercion classes 
formulated by Lakoff (Lakoff and Johnson, 1990). 
We propose here a metonymy resolution approach 
that accounts for an open class of coercions. Simi- 
larly to Nunberg (Nunberg, 1995) and more recently 
to Markert and Hahn (Markert and Hahn, 1997), we 
find metonymy and nominal reference resolution to 
be two interacting processes; therefore, the proposed 
computational model validates metonymies through 
coreference links. 
2 Classes of metonymic coercions 
Stallard proposed in (Stallard, 1993) a distinction 
between two kinds of metonymy: (1) referential 
raetonymy, in which the referent of a nominal pred- 
icate argument requires coercion and (2) predica- 
tive metonymy, featuring the coercion of the pred- 
icate usually corresponding to a verbal lexicaliza- 
tion. In his study, Stallard focuses on metonymic 
inferences required by a specific performative con- 
text, characterized by wh-questions and impera- 
tives. His formalization of referential and predica- 
tive metonymies is based on the logical form read- 
ings of utterances from the DARPA ATIS (Air Travel 
Information Service) domain (MADCOW. 1992), a 
question-answering database about commercial air 
flights, comprising questions of the form: 
I 
I 
I 
I 
I 
I 
I 
I 
i 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
(QI) Vhich wide-body jets serve dinner? 
(Q2) Which airlines fly from Boston to Denver? 
The ATIS domain is characterized by a pre- 
established formal system of categories and relations 
onto which the utterances must be mapped. In this 
domain it is known that only flights fly or serve 
meals; thus, both (Q1) and (Q2) can only be un- 
derstood metonymically. The interpretation of the 
ATT$ utterances is performed in the logical language 
imposed by the implementation in the DELPHI sys- 
tem (Bobrow et al., 1991). In this framework, a 
question is translated into a LISP-expression: 
(wh x S (and (PI x) (P2 x)) 
interpreted as a query for all members of S (the se- 
mantic class of the wh-NP) that satisfy both PI (de- 
fined as the modifiers of the wh-NP) and P2 (the 
predicate of the clause). For exemplification, (qLFi) 
and (0LF2) represent the logical form translations of 
(QI) and (Q2)t: 
(QLFI) (whx flights 
(and (exists y jets 
(and (aircraft-of x y) 
(.ide-bodyy)) 
(serve flight-of 
x meal-of dinner))) 
(QLF2) (wh x airlines 
(exists y flights 
(and (airline-of y x) 
(fly flight-of y 
orig-o~ Boston 
dest-of Denver)))) 
In the case of (QLFI), the coercion relation 
aircraft-of maps between flights and the aircrafts 
they are on, whereas in the case of (QLF2), the co- 
ercion relation airline-of translates the connec- 
tion between airlines and flights. Although both 
(QLFI) and (qLF2) have an interpolated quantifier 
for flights, which is not specified in the utterance, the 
difference comes from the position of the variable: in 
the case of (QLF1), the interpolated variable x is the 
wh-variahle, whereas in the case of (QLF2), the in- 
terpolated variable y is part of the description that 
should be returned by the query. Stallard notes that 
this is the crux of the referential/predicative distinc- 
tion of metonymies. 
In (0LFt), the NP-argument jets does not rep- 
resent the domain of the wh-variable, but flight; 
does, thus indicating a metonymic reference and 
deriving a referential metonymy reading. In con- 
trast, in (QLF2) the domain of the wh-variable co- 
incides with the semantic type from the utterance 
(i.e., airlines), but the subject-argument of the 
fly predicate is replaced by the coercion flights. 
This prompts Stallard to state that "predicative 
I (Or), (Q2), ({~LFt) and (qLF2) are borrowed from (StaJ- 
lard, 1993) 
143 
metonymy can be loosely thought of as coercion of 
a predicate place, rather than that of the argument 
NP itself". 
We argue that this definition is dependent on 
two factors: (1) the specific logical transforma- 
tion imposed by the DELPHI implementation, which 
does not cover forms of utterances other than wh- 
questions, and (2) the availability of thematic role 
relations and coercions tailored specifically for the 
ATIS domain. This characterization of referen- 
tial/predicative metonymies is not applicable when 
processing different genres of text, operating in dif- 
ferent domains and, thus needing a different knowl- 
edge representation. An example ofa metonymy res- 
olution system using a more general representation 
is reported (Markert and Hahn, 1997). 
In their model of metonymy inference, Markert 
and Hahn employ a two-tiered conceptual and se- 
mantic test: conceptual checks identify well-formed 
role chains between a pair of syntactically linked con- 
cepts, and then semantic checks distinguish whether 
these chains mirror literal or metonymic relation- 
ships. To perform these checks, they use (1) a 
concept hierarchy C with a taxonomic relation isac 
and (2) a set of relation names 7~, containing la- 
bels of all conceptual roles of the elements from C, 
hierarchically organized by isa~. For the ATIS do- 
main, we can assume C = {AIRLINE, JET, MEAL, 
PASSENGER, PILOT .... } and 7~ -- {has-aircraft, 
has-flight, property-of' serves, fly-from, fly-to .... }. 
To be able to parse texts, Markert and Hahn de- 
vised a system that grants a syntactic link between 
two concepts z and y if there is an acyclic path of 
relations ri E 7~ and concepts cj E C such that each 
ri is a conceptual role of ci-t, with ra,lge(ri)= ci 
and co = z A (c, isac. y V y isao ca). 
Following established classifications (Lakoff and 
Johnson, 1990), Markert and Hahn predefine some 
of the relations from ~ as metonymic. These rela- 
tions are {has-part, part-of, produced-by, contained- 
in, made-of}. Thus, a metonymy is recognized 
whenever one of the relations ri from a path is 
metonymic. 
In this framework, the interpretation of the re- 
lation between x =airline and y =dinner in 
(Q1) is rendered by the Path-l: c0--company 
(with z"=airline-isa-~company), rt-'has-part, ct= 
employee (with flight-attendant-isa-+ employee 
and range (serve, SUBJECT) = flight-attendant), 
r2=serve, c3=meal (with y--dinner-isa-+meal). 
Similarly, the interpretation of the conceptual relat- 
edness between r--airline and y=Boston in (Q2) 
is rendered by the Path-2: co=airline, rt=has- 
flight, cl=scheduled-flight, r2=arrive-at, c3=city 
(with y--Boston-isa-~city). 
The presence of rl=has-part in Path-I and of 
rt=has-flight in Path-2 indicates that the two 
! 
I 
! 
! 
paths correspond to metonymic readings. Relation 
has-flight is not among the predefined metonymic 
relations considered by Markert and Hahn, but 
clearly would need to be so, to classify Path-2 
as metonymic. Moreover, as a distinction on the 
predicate coercions, Path-1 contains the relation 
rz---serve, identical to the wh-predicate of (Ol), 
indicating that it is not a predicative metonymy. 
The referential metonymy from Path-1 is determined 
by the metonymic relation rl=has-part, coercing 
airline to flight attendants. In contrast, Path- 
2 has relation r2-arrive-at that coerces the wh- 
predicate fly-to from ({~2). 
Markert and Hahn do not analyze the seman- 
tics of the metonymic paths, but instead distinguish 
referential and predicative metonyrnies only in the 
anaphoric cases. Considering that expression A is a 
metonymic coercion of the concept B, they assume 
that in the predicative case A should be available 
for reference resolution, whereas in the referential 
case, only B should be so. To be able to assess 
the availability for reference resolution, they search 
for the presence of A and B in the list of forward- 
looking centers of the previous sentences (thus using 
the functional centering framework defined in (Grosz 
et al., 1995)). 
A similar path-finding methodology for deriving 
metonymies was used by the met* system (Fass, 
1991), in which connections between the sense 
frames of textual concepts are retrieved from a lex- 
icon of the size of 500 word senses. These paths 
are then classified against a small set of predefined 
metonymic inference rules, and form the grounds for 
the figurative interpretation of textual expressions. 
In met* there is no support for distinctions between 
predicative and referential metonymies, since coer- 
cions are possible from any concept in a text. The 
appeal of this implementation stems from the fact 
that it uses word sense frames, inspired by the struc- 
ture of dictionary entries and shows that the paths 
retrieved from such a knowledge representation can 
be used to identify classical forms of metonymy. 
This indicates that metonymy resolution can be per- 
formed by processing knowledge from lexical dictio- 
naries, and therefore WordNet 1.6 is a suitable can- 
didate. 
A different methodology of deriving coercions 
was implemented in TACITUS (Hobbs et al., 1993). 
Whenever sortal constraints are violated, explicit ar- 
guments are replaced with coercion variables, related 
to the explicit arguments by generic relations. The 
coercion is devised when the generic relation sub- 
sumes some predicate that is brought forward by 
the abductive interpretation of the text. In the case 
of (Q1), argument jets is replaced with a coercion 
variable L- which is expected to satisfy the subject- 
constraints of the verb serve. The abductive inter- 
144 
pretation of (Q1) brings forward a reasoning path 
showing that variable k may be coerced to any sub- 
sumer of concept person. Such a subsumer is synset 
{steward, flight attendant}, having the gloss 
(an attendant on an airplane). This gloss translates 
the generic relation between jets (a hyponym of air- 
planes) and variable k to the predicate on, cued by 
the prepositional relation attendant-on-+airplane. 
The interpretation of this prepositional relation is 
produced when it is matched against WordNet- 
based classes of prepositional attachments collected 
from large treebanks, following the methodology de- 
scribed in (Harabagiu, 1996). For this case, the on- 
prepositional relation attaches the place of work to 
the worker, thus giving meaning to the coercion of 
jets into flight attendants. 
Although the coercions derived in TACITUS do not 
distinguish the predicative or referential cases, they 
present a different method of building metonymic 
paths. By incorporating this unification-based 
mechanism of producing coercions with a lexical 
path-finder working on WordNet, a novel way of 
deriving metonymies is made possible. It has the 
advantage that it relies only on approximations of 
sortal knowledge, as indirectly available from the 
WordNet database, and it does not need full-fledged 
abductions to be able to return metonymic paths. 
3 Metonymic paths 
The process of deriving metonymic paths from 
WordNet consists of three distinct phases: (1) the 
identification of sortal constraints that need to be 
satisfied during the interpretation of nominal ex- 
pressions, (2) the retrieval of related knowledge that 
complies with the sortal restrictions, and (3) the val- 
idation of coercions against anaphoric expressions 
from the sentences following the processed sentence. 
The first two phases rely on access to semantic 
information available in (1) the relational seman- 
tic encoded in WordNet (e.g., hvpernyms, is_part, 
is_member, is_stuff" entail or pertaymym) spanning 
synsets or words encoded in the database and (2) 
the semantic of the synset definitions (known as 
glosses). To be able to have computational access 
to the gloss semantic, synset definitions have been 
translated into logical formulae inspired by nota- 
tion proposed in (Hobbs, 1986-1) and implemented 
in TAClTUS. 
Based on the davidsonian treatment of action sen- 
tences, in which events are treated as individuals, 
every gloss is transformed in a first-order predicate 
formula for which (1) verbs are mapped in predi- 
cates t,erb(e,z.y) with the convention that variable 
e represents the eventuality of that action or event 
to take place, z represents the subject of the ac- 
tion. and Y represents its object (in the case of in- 
transitive verbs, Y is not attached to a predicate. 
whereas irt the case of bitransitive verbs, y is as- 
sumed to range over both the direct and indirect 
object); (b) nouns are mapped into their lexicalized 
predicates and (c) modifiers have the same argument 
as the predicate they modify. Prepositional attach- 
ments are indicated by preposition-predicates rang- 
ing over the pair of arguments of the predicates they 
attach. For example, the gloss of synset {airline, 
airline business, airway} is (a commercial en- 
terprise that provides scheduled flights for passen- 
gers) and has the following logical form transforma- 
tion (LFT): 
\[ea~ ezptise (x) ~:co~ereial (z) &provide (e, x, y) & 
~:t li~cht (y) ~scheduled (y) ~f or ( e, p) ~passenge:r (p) \] 
Characteristic of LFTs is the fact that the gloss genus 
is always the first predicate, rendering the LFT a 
formula of the form \[genus(x)g:differentia(y)\]. 
Gloss geni are accessed repeatedly during the deriva- 
tion of metonymic paths, and thus they need to be 
easily accessible. 
The first two phases of the metonymic inference 
relies also upon lexico-semantic relations determined 
by derivational morphology, specifically the links be- 
tween verbs and their nominalizations. Relations be- 
tween verb and noun synsets that have elements with 
common morphological roots have been added to the 
database, classifying them as (a) the action nominal- 
ization; (b) the result (or object) of the action or (c) 
the agent (or subject) of the action, For example, 
verb propose and noun proposal refer to the same 
action. A nominalization(act) link was established 
between verb synset {propose, pro jet1:} and the 
sense of proposal glossed as (the act o/making a 
proposal). A second nominali=ation(result) link was 
established between the same verb synset and the 
sense of proposal glossed as (something proposed). 
Phase h Approz~mation o/ the sortal constraints. 
A nominal N is interpreted as literal or figurative 
depending on whether its sortal constraints to syn- 
tactically linked verb V are satisfied or not. Sortal 
information can be found in: 
• (i) the LFT(g), where g stands for the gloss of any 
sense i of V (hence Vi) or any of its hypernyms; 
• (ii) LFT(e). where • represents an example from 8; 
• (iii) I, FT(c). where e represent those glosses where 
and V co-occur (and the sense of Y is unknown}. 
To access the sortal constraints of V implicitly en- 
coded in WordNet, we collect all expressions from 
LFT(g) or LFT(e) such that: 
(1) they contain a predicate verbi(e~, x\[, x~), 
representing either (a) V/ or (b) one of its hypernyms 
or (c) one of the geni from the LFTs of I,~ or its 
hypernyms; 
(2} they also contain any predicate that is (a) a subject, 
(b) an object and/or {c) a prepositional attachment 
tO verbi in the same /..FT. 
When all predicates are conceptualized in the re- 
spective LFTs. such expressions have the form: 
145 
Si ~ i i i verbi(el ,xz, x~) ~ subjecti{x~) ~ objecti(x~) &: 
l'Iflprepj(e~,yj) ~ nou.j(y~)) 
The sortal information for subjects and objects of 
Vi is: 
S.bjeet,(V) = Uk • i i subjectk(zl.k ), where subjectik is: 
(a) the subject from some Si (= S~) and 
(b) is the hypernym of any other subject (from another 
S~ ok) that belongs to the same WordNet hierarchy. 
Objecti(V) = Uk, objeet~,(x~ ~, ), where objectik, is: 
' kt (a) the object from some Si (= Si ) and 
(b) is the hypernym of any other object (from another 
S~ 'ok') that belongs to the same WordNet hierarchy. 
Similarly, for each prepositional attachment de- 
termined by a preposition prep, we define the sortal 
information: 
Nouni(V ,prep) \[.Jk,, i i i = nounk,,(x t.~,,) where nounk,, is: 
(a) attached to verbi in some Si (S~") through prep and 
(b) is the hypernym of any other noun attached through 
prep (in another S~ ''gh''), when both nouns belong to 
the same WordNet hierarchy. 
Expressions S" are collected from the r.xT(¢). The 
semantic sense k of V in S~ is selected as a result of 
the fact that the similarity measure between S" and 
the collection {5; } is maximal when i = k. The sim- 
ilarity measure between two expressions is defined 
as: 
Similarity(S'~,SO = Sim(subject, S~,Si) + 
+ Sim(object.S~,,S~) + ~j Sim(prepj, S'~, Si))) 
where for role E {subject, object,{prepj} } we have: 
(i) Sire(role. S~,, S~) = 1 if the conceptualizations of 
role(S~,) and role(Si) belong to the same hierarchy 
... q.r (ii) Sire(role. S~,. Si) 0 when either role(~ ~) 
role(Si) are not defined or 
iii) Sire(role. S'~, Si) = -I otherwise. 
Finally, considering the set operator: 
St (~a $2 ={e \] e E St IJ $2 and there is no other 
e' E St LI $2 such that e' is a hyeprnym of e} 
the sorrel constraint approximations are defined as: 
rl Subject_SortdV)=Subjeeti(V) ~a \[Jq subject~, 
where subject~ is the subject from some S'~ in 
which the sense of V is i; 
r30bject.Sorti(V)=Objecti(V) ~h Uq, °bjecti¢, 
where objectS, is the object from some S~, in 
which the ~ense of V is i: 
\[3 Prep.Sort,(V,prep)=Nouni(V,prepl ~h U¢, n°uni4 ', 
where nouni¢, attaches to the sense i of V through 
prep in some S~. 
Phase IL" Sorts satis\[action and expansion o\[ co- 
ercions. Sorts satisfaction amounts to (1) the recog- 
nition of the sense of V in the text and (2) a search 
for any element from Role-Sorti(V) across all con- 
cepts semantically related ro all WordNet senses of 
N (given that t/has the same role in the text). 
The recognition of sense i of V is based also on 
the maximal value of similarity between S,, the 
expression retrieved from the r~T of the text, and 
{Sort,(V)}. where Sort,(V) is defined as: 
I 
I 
I 
I 
I 
I 
! 
I 
I 
! 
I 
I 
I 
! 
! 
! 
! 
Sorti(V) = Subject_Sort;(V) & Object_Sorti(V) & 
& I-Ii Prep_Sorti(V,prepj) 
The satisfaction of Role_Sorti(V) is a search for 
any element from this set along (1) all synonyms, (2) 
all hypernyms and (3) all ~ geni for each Word- 
Net sense of N. If this search is successful, we rule 
that N had a literal reading. Otherwise, we need 
to build metonymic paths to be able to access the 
related knowledge. Two distinct ways of deriving 
metonymic paths have been developed. 
Lezico-semantic paths. The codification of 
meronymic relations in WordNet determines the 
consideration of lexico-semantic paths composed of 
isa and at least one is_part, is_member, or is_stuff 
relations (or their reverses} as a means of deriv- 
ing coercions. Implementing 22.20% of the semantic 
connections between noun concepts as meronyms, 
WordNet 1.6 sets an acceptable level of granular- 
ity for a knowledge representation needed to derive 
metonymic information. 
Lexico-semantic metonymies retrieve concepts 
Cm E Role.Sorti(V) that (1) are linked through a 
meronymic relation to any WordNet sense of N (or 
one of its hypernyms or geni); (2) morphologically 
or idiomatically indicate meronymic relations to N, 
or (3) represent predicates from the LFr of N (or its 
hypernyms) having thematic roles for the same verb 
that is the genus of the LFr. The general form of the 
lexico-semantic paths is determined to be PathLs 
= (Co--N, rhChr2 .... ,Crn), with at least for one i, ri 
is a WordNet meronymic relation. The case when 
a triplet (Cj_t,ri,Ci) is part of an LFT extends the 
classical metonymies object-for-agent and product- 
for-producer. 
The WordNet concepts that morphologically cue 
meronymic relations are those synsets containing 
collocations of such lexemes as unit (e.g., admin- 
istrative unit, army unit), system (e.g., exhaust 
system, file system), par~ (e.g. body part, aca- 
demic department), group (e.g., jazz group, pres- 
sure group), or other words that form the same hier- 
archies as part to member. Similarly, concepts con- 
taining in their glosses idioms like ' 'a group of' ' 
or ' 'part of' ' cue meronymic relations to thegloss 
genus. 
Morpho-logical paths. Frequently, N is a nominal- 
ization ofa VN, and thus LFI'(VN) brings forward re- 
lated semantic information. Moreover. the nominal- 
ization relations indicate the role of N in the r.Fr(E~): 
a nominalization(result) link corresponds to 
an object role. whereas a nominalization(agent) 
link relates to a subject role of N in the LFT(rv~). 
This entails the interchangeability of the genus of 
r.yr(N) with verb(e, zt,z~)&noun(zi) (with i=l if 
role=subject and i=2 if role=object) when verb is 
the predicate representing l,'h and noun is the predi- 
cate for N or any of its geni (in the hierarchies of the 
146 
sense of N morphologically related to l/~v). The re- 
sulting logical form transformations are denoted as 
L~' (~). 
In this case, we can compute the Similar- 
ity(Text,LFT' (N)) and pick the Role(s) for which it 
is maximal and incorporate the corresponding LFTs 
in the most similar r.FT, (N), producing the final co- 
ercion. Morpho-logical paths are sequences of three 
kinds of steps: (1) relatedness based on morphologi- 
cal relations (i.e., N-nominalization-~VN), (2) ad- 
hoe weighted abduction based on similarity between 
text roles and the logical forms of the hypernyms 
of N (i.e. LFT' (N)-Similarity(text)~LFT(Roles)), 
and (3) unification of similar logical expression (i.e., 
LFT' (VN)-unification(LFT(text)-+Nc. 
Morpho-logic paths exploit morphological links 
and overlaps in the L~S and resolve predicative 
metonymies (by bringing into play additional ver- 
bal predicates). In contrast, lexico-semantic paths 
resolve referential metonyies. 
Phase III: Anaphora validation. Metonymic 
paths produce the expected coerced knowledge if 
they bring forward concepts that corefer with nom- 
inals from the successive sentences. There are 
three tests for the validation of coercions through 
anaphora. They determine whether there is a con- 
cept in a lexico-semantic path or a predicate in a 
morpho-logical path that (1) is identical, (2) is a 
hypernym, or (3) is a genus of one of the nonfinal 
expressions from the following sentences. 
4 A case study 
The processing associated with the derivation of 
metonymic paths is exemplified on a text presented 
in the manual defining the coreference task for the 
DARPA-sponsored MUC 7 competition (Hirshman 
and Chinchor, 1997): 
(S1) The White House sent its health care propo- 
sal to the Congress yesterday. 
($2) Senator Dole said the administration's bill 
had little chance of passing. 
The approximation of the sortal constraints of 
'¢=send from (St) determines: 
o (a) the selection of sense i=2 from the eight senses 
encoded for verb send in WordNet 1.6, due to greater 
similarity between its sorts and the roles from ($1). 
o (b) Subject~send)={{person,individual,someone}}; 
Object2(send)={ {communication}, {physical object}}; 
Prep_Sor t2 (send, to)= Prep-Sor t~ (send, from) = 
= { {person,individual,someone }. {address } }. 
The search space for the sortal constraints is de- 
termined by the LFT of ($1): 
LFT (S I) = \[White~ouse( zt )~send(el. Xt, t2 )~ 
~:proposal( x~)~heal th_care( x2)~to( el, X3)& 
~Congress (x~) ~ ye st erday( e i } \] 
I 
I 
I 
I 
I 
I 
I 
across the synonyms, hypernyms and geni of both 
WordNet senses of White House, the three senses of 
noun proposal and proper noun Congress respec- 
tively. 
The WordNet search' for the satisfaction of sortal 
constraint identifies communication as the hyper- 
nym of sense 1 and 2 of proposal and person as 
the genus of {legislature,law_makers}, the hy- 
pernym of Congress, thus indicating that in (Sl), 
proposal and Congress have literal meaning. Noun 
proposal as a nominalization is a candidate for 
morph-logic path derivations as well. The search for 
sortal constraints for the role of subject is not suc- 
cessful, requiring the inference of metonymic paths. 
A lexico-semantic path is derived, linking sense 1 
of White House to person, the genus of synset 
administrat ion: 
Pathl: llhite.Houso-isa--~govenmentAepartaent-- 
- isa-+administrat ire_trait - morphological.zue-~ 
-+admiaistration--genus-~person 
The meronymic morphological cue from Pathl is 
consistent with the meronyms encoded in Word- 
Net 1.6, because we have: 
(a) govemment-department- is_part-+admirdstration 
as an immediate inference from Path1 
(b) adminJstration-is_part--~government 
as a semantic relation encoded in WordNet 1.6 
(c) govemment-department-is_part-.+government 
as a semantic relation encoded in WordNet 1.6 
Moreover, the anaphoric validation of Pathl is 
possible because nominal administration is present 
in (S2). 
In the case ofObject2(send) a morpho-logical path 
accounts for the coerced knowledge. Nominalization 
proposal is the result of the action expressed by 
synset V'={ propose,project}. \[ntegrating predicate 
proposal as an object in the LFT(V') we obtain: 
LFT' (proposal)=present (e2, Yt, Y~)&proposal(y~ )& 
&for(e~, Ys )&consideration(ys ) 
When computing the similarity with (S1), we 
obtain Prep-Sort2(send,to) as the role candidate 
to be incorporated in LFT'(proposal). This is en- 
forced by the LFT of synset {motion,question}, 
a hyponym of sense I of proposal. The gloss of 
{motion,question} is (a proposal for action made 
to a deliberative assembly for discussion and vote), 
producing the LFT: 
LFT( motion)=proposal{p)&for(p,d)&discussion(x)& 
&vote(x)&to(p,a)&assembly(a) 
in which present(e2,yt, y~)&proposal(y~) may sub- 
stitute proposal(p)whereas discussion can be re- 
placed by its hypernym considered:ion. Further- 
more, since Congress is a hyponym of assembly 
and the filler of Prep_Sort2(send.to) in (S1). 
we can unify Lgr(motion), LFT'(proposal) and 
LgT(legislature) and obtain the final coercion as: 
147 
proposal(p)=~present (e2, zl, z~. )&proposal(x2 )& 
&for(e2, x~)&consideration(3s )&to{e2, x2)& 
&person(x~ )&make(e~, x~,/)&law(l) 
Path2 is shown in its entirety in Figure 1. The 
anaphoric validation is shown to link bill from ($2) 
to la~, since law is a genus of the first WordNet 
sense of bill. Pathg brings forward two new actions 
(indicated by e2 and ea), which accounts for its clas- 
sification as a predicative metonymy. 
\['he White House sent its health care proposal to Congress yesterday. 
t =' 
administrative_unit 
morphological J 
: ~ 
.i..~dminiaration 
i Gloss: the persons 
i who make up a 
; governing body ... 
i LFT:person(pl)& 
! make_up(e4,pl.g)& 
i government(g) 
............ 1 
{ propose.project I ! 
Gloss: present for consideration 
LFT:present(e2,y I ,y2)&for(e2.c) 
&consideration(c) 
{ moues.question } 
Gloss: a formal proposal 
for action made to a 
deliberative assembly 
for discussion and vote 
LFT:proposal(p)& forfp,d ) 
&discussion(d)&vom(d} 
&to(p,al&assemblyla) 
{ legislamre } 
Gloss: group of persons that 
make laws 
LFT:person(x2 )&make(e3,x2.1)& ! 
&law(1) i 
for(e2.c)&consideflc)&to(e2,x3 )& 
person(x3)&makefe4.x3.1 )&law(1) 
co.e/ere.ce 
I ............. 
i nator Dole said the administration's bill had little chance of passing. 
Figure h Example of metomymic paths 
The usage of the LFT of a hyponym of proposal 
is a consequence of modeling Gricean principles via 
lexical chains from WordNet. The relevance maxim 
is enforced whenever as many lexical chains as pos- 
sible can be retrieved between previously activated 
concepts and novel information. When we pro- 
cess the sort satisfaction of the object role, concept 
legislature is already activated and it satisfies the 
Prep-sort2(sead.to) const raints. The same thematic 
role is represented in the LFT of mo~:ion, thus enlarg- 
ing the number of \[exica\[ paths between proposal 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
i 
I 
I 
I 
/ 
! 
I 
I 
I 
and legislature and increasing the relevance of 
logis:l.atttre in the context of coercing knowledge 
for the object proposal. The final unification rein- 
forces the relevance of logisla~ttre, congruent with 
the coreferential link between bill and 1am 
5 Conclusions 
This paper proposes a method of deriving 
metonymic coercions by using the information avail- 
able in WordNet 1.6. The method does not presup- 
pose the availability ofsortal constraints, but rather 
builds approximations ofsortai information from the 
knowledge implemented in WordNet. It uses two dis- 
tinct approaches for knowledge coercion: one that 
relies on lexico-semantic information, another based 
on morphological links and unifying logic formulae 
(LF'rs) inferred from conceptual definitions. 
To evaluate this methodology of deriving metonymic 
coercions, a test set of 20 New York Times articles 
were parsed by FASTUS (Appelt et al., 1993) and 
used in conjunction with their coreference keys, as 
provided by the MUC test data. There were 1261 
nominal expressions, distributed in four classes as il- 
lustrated in Table 1. A percentage of 23% nominals 
were anaphoric, out of which almost 74% had literal 
meaning. In approximating the sortal constraints, 
sorts for 68.2% of nominals were returned. When the 
sorts were not satisfied by searches through Word- 
Net, metonymic paths were derived. As Table I in- 
dicates, 68.9% of these paths were lexico-semantic 
(denoted with r=-referential metonymies) and 31.1% 
were morpho-logical (denoted with p=predicative 
metonymies). 
II ,Vo~,.~l ry~ III "' Lito~, I Me~onymy II ,,Bare :~ (16\[r%i' 
r=3 p=s 
Definite ~5 (34.8%) r=', p=4 
Indefinite '6 (2.7%) r=9 p=5 \] ~: 
\[.. Proper Nouns ~}8 (45.5%) r= 11 p=2 
U-TOTAL \[1\[ 215 (74.1%) \] A9,(16.8%) H 
Table h Distribution of metonymic anaphorae 
The evaluation of path-validating anaphorae 
against coreference keys resulted in a precision rate 
of 76% and a recall of 83%. These results indicate 
that we need to experiment with different similar- 
ity measures. We also found that the metonymies 
have a significant contribution as knowledge sources 
for coreference resolution. 43% of the coreference 
keys were accounted by mere string matches. 3.1% 
by synonyms encoded in WordNet. 3.7% by hyper- 
nyms and 16.3% by links made possible through co- 
ercions. 

References 
Douglas E. Appelt..Jerry R. Hobbs..John Bear. David 
Israel. Megumi Kameyama. aJad Mabry Tyson. The 
SRI MUC-5 JV-FASTUS Information Ex:raction Sys- 
tem. In Proceedings of the Fifth Message Understand- 
ing Conference (MUC-5), 1993. 
John Bear and Jerry R. Robbs. Localizing Expressions 
of Ambiguity. In Proceedings of the Second Conference 
on Applied Natural Language Processing, Association 
for Computational Linguistics, pages 235-242, 1988. 
Richard Bobrow, Ran Ingria and David Stallard. The 
Mapping Unit Approach to Subcategorization. In 
Proceedings o\[ the Speech and Natural Language 
Workshop, 1991. 
Dan Fuss. met" : A Method of Discriminating 
Metonymy and Metaphor by Computer. In Compu- 
tational Linguistics, 17(1):49-90, 1991. 
Paul J. Grice. Logic and Conversation. In P. Cole 
and J. Morgan, editors. Syntax and Semantics Val.3: 
Speech Acts, pages 41-58. Academic Press, New York, 
~975. 
Barbara J. Grosz, Aravind K. Joshi and Scott Weinstein. 
A Framework for Modeling the Local Coherence of 
Discourse. In Computational Linguistics, 12:(3), 1995. 
Sanda ttarabaglu, Dan Moldovan and Takashi Yukawa. 
Testing Gricean Constraints on a WordNet-Based Co- 
herence Evaluation System. In Working :Votes o/the 
AAAI.96 Spring Symposium on Computational Ap- 
proaches to Interpreting and Generating Conversa- 
tional \[mplicature, pages 31-38. 1996. 
Sanda Harabagiu. An Application of WordNet to Prepo- 
sitional Att_achment. In Proceedings o/the .'~st An- 
nual Meeting of the Association for Computational 
Linguistics. ACL-96, pages 361-362, 1996. 
Lynette Hirshman and Nancy Chinchor. MUC-7 Coref- 
erence Task Definition, Version 3.0.. 1997. 
Jerry R. Hobbs. Discourse and Inference. Unpublished 
manuscript. 1986- t. 
Jerry R. Hobbs. Overview of the TACITUS Project. In 
Computational Linguistics. 12:(3). 1986-2. 
Jerry R. Hobbs. Mark E. Stickel. Douglas E. Appelt and 
Paul Martin. Interpretation as Abduction. in Artifi- 
cial Intelligence. 63:69-\[42, 1993. 
George Lakoff and Mark Johnson. Metaphors We Live 
By. Chicago University Press, Chicago. IL. 1990. 
MADCOW Committee {Hirshman. Lynette et all). 
Multi-Site Data Collection for Spoken Language Cor- 
pus. In Proceedings o\] the Speech and Natural Lan- 
guage Workshop. 1992. 
Katya Markert and Udo Hahn. On the |ntcraction of 
Metonymies and Anaphora. In Proceedings of the 15th 
International Joint Con\]erence on Artificial Intelli- 
gence (IJCAL97). Nagoya..Japan. t010-t015. 
George A. Miller. WordNet: a Lexical Database for 
English. In Commumcatwns o\[ the .4CM, Vol.38, 
No. 11:39-41. 1995. 
Geoffrey Ntmberg. Transfers of Meaning. In lournal o\[ 
Semantics. 12:109-132. 1995. 
David Stallard. Two Kinds of Metonymies. In Proceed- 
ings of the \]lst Annual Meeting of the Association 
for Computatwnal Lmquistics. A CL=q.'L pages 87-94. 
1993. 
