Processing Metonymy: a Domain-Model Heuristic Graph 
Traversal Approach* 
Jacques Bouaud, Bruno Bachimont, Pierre Zweigenbaum 
DIAM: SIM/AP-IIP A 1)el)t (te Biomathdmatiques, Universitd Paris 6 
91, boulevard (le l'It6pital F-75634 Paris Cede.x 13 
{j b, bb, pz }(<l)biomath.jussiou.fr 
Abstract 
We address here the treatment of 
me, tonymie expressions from a knowl- 
edge representation perspe(:tive, that is, 
in the context of a text understanding 
system whi('h aims to build a (:oncep- 
tual representation from texts according 
to a domain mode, l ext)resse, d in a knowl- 
edge representation formalism. We fo- 
cus in this t)aper on the part of tile se- 
mantic analyser which deals with seman- 
tic eoml)osition. We explain how we use 
tile domain model to handle metonymy 
dynamically, and more generally, to un- 
(lerlie semantic (:omposition, using tile 
knowledge descriptions atta(:hed to ea(:h 
(:oneept of our olttology as a kind of 
eon('el)t-h;ve.l , multii)b.-role (lualia struc- 
ture. YVe rely for this on ~t heuristic 
1)ath search algorithm that exl)loits the 
gr~phic aspects of the eon(:eptual gratIhs 
formalism. The methods described have 
1)een imi)lemente<l and applie(l on French 
texts in the medical domain. 
1 Introduction 
\[\]ii(ter the eomt)ositional assulnption, senmntie 
analysis relies on the combination of the mean- 
ing representations of parts to build the meaning 
representations of a whole. However, this con> 
position often needs to call on implicit knowledge 
whi(:h helps to link the two meaning representa- 
tions. This is the (:as(*,, for institute, in metonymi(" 
expressions, where a word is used to express a 
notion closely related to its central meaning. A 
well-known stream of work addressing this t)he- 
nomellon is the Generative Lexicon theory (Puste- 
jovsky, 1991). At the heart of this theory is a lex- 
ical semantic representation called "qualia struc~ 
lure", Met(mymies are considered to correspond 
to changes in the semantic types of the words in- 
*This work has been imrtly supporte, d by the Eu- 
tel)earn project MENELAS (AIM 2023). 
volved, and the qualia structure provides the basis 
for performing type coercion in a generative, way. 
We address here, the treatment of metonymie 
expressions from a knowledge representation per- 
spe(:tive, in the context of the MENF, I,AS medi- 
cal text understanding syste, m (Zweigenbaum et 
al., 1995). One of the goals of the overall system 
is to assign stan(tar(lised, medi(:al nomenclature 
codes to the input texts (patient discharge sum- 
maries). Semantic analysis st~rts from a syntac- 
tic representation of each sentence and produces a 
conceptual representation. It is then used by sev- 
e, ral language-independent, knowledge-based com- 
ponents to perform inferences (pragmatic enrich- 
nlellt) and then code assignment (Delamarre et 
al., 1!)95). Therefore,, the, conceptual represen- 
tation outtmt by the semantic analyser nmst be 
normalised: it must ('onform to a knowledge repre- 
sentation canon in which the target nomenclature 
(:odes can lie nlal)ped. 'l'he si)eciiication of this 
canon relies on the description of a rich model of 
the domain in a knowledge representation formal- 
ism, here Conceptual Graphs (CG) (Sowa, 1984). 
We focus in this patter on the part of the se- 
mantic analyser that deals with semantic ('ore- 
position. The conceptual reI)resentation built 
must be abstracted from initial linguistic varia- 
tion, metonymy being a typical problem to be ad- 
dressed. We ext)lain how we use the domain mo(lel 
to handle metonymy, and more generally, to un- 
derlie semantic composition, using the knowledge 
descriptions attached to each concept of our ontol- 
ogy as a kind of concept-level, multiple-role qualia 
structure. The methods described have been im- 
plemented and applied to French texts. 
We first recall the problem addressed (sec- 
tion 2). Then, the pr()posed method is described 
(section 3) and illustrated on an example. We give 
some information on the imt)lementation and the 
results of the analyser (section 4), and discuss tit(,' 
relatiw', merits of the method (section 5). 
2 Metonymy and type coercion 
A (:lassical exainple of metonynly (Pustejovsky, 
1.991, It. 428ff) is 
137 
(1) John began a novel. 
where predicate 'began' expects an event as its 
second argument, so that some way must be found 
to relate the object 'novel' to an event such as 'to 
read a novel' or 'to write a novel'. In our do- 
main (coronary diseases), one often finds expres- 
sions such as 
(2) une angioplastie du segment II (an angio- 
plasty of segment II) 
(3) une angioplastie d'une artbre coronaire (an 
angioplasty of a coronary artery) 
(4) l'angioplastie de Monsieur X (the angioplasty 
of Mr X) 
(5) une angioplastie de la st6nose (an angioplasty 
of the stenosis) 
where 'angioplasty' is an action performed on a 
segment of an artery to enlarge its diameter, while 
'stenosis' is the state of an artery which has a re- 
duced diameter. These four phrases involve the 
object (or "theme") of action 'angioplasty', i.e., 
what the angioplasty operates upon. If one con- 
siders that this theme must be a physical ob- 
ject, then examples (2)-(4) conform to the selec- 
tional restrictions of 'angioplasty', while (5) vi- 
olates them. The mechanism of type coercion 
(Pustejovsky, 1991) consists in converting a word 
type into another so that semantic composition 
can work properly. (5) is then handled as a 
metonymy, where the stenosis and the stenosed 
object enter a state/thing alternation: 'stenosis' 
is turned into an 'object'. 
However, it appears that this phenomenon is de- 
pendent on the underlying types (or "sorts") un- 
der consideration. For instance in our ontology, 
'segment', 'artery', 'stenosis' and 'human' have 
four different types, and are not comparable by 
the IS-A relation, e.g. nothing can be both a seg- 
ment and an artery} This is a voluntary, method- 
ological choice (Bouaud et al., 1995), motivated by 
the fact that these objects give rise to different in- 
ferences and must not be confused by the reason- 
ing component. Additionally, in the target nor- 
malised conceptual representation, what consti- 
tutes the specific theme (in our conceptual model, 
the purported_oh j) of action 'angioplasty' must 
be precisely defined. In the context of our appli- 
cation, 'angioplasty' acts on an artery_segment, 
a physical object corresponding to a part of an 
artery, which happens not to be comparable to 
any of the four preceding themes of 'angioplasty'. 2 
Therefore, all four examples (2)-(5) must be con- 
sidered as metonymies. 
1Segment, in our ontology, corresponds to a portion 
of space, not of matter. 
2Notice, though, that these types are strongly 
linked (by relations other than IS-A) through the 
knowledge base models. The semantic analyser pre- 
cisely recovers these links thanks to the mechanism 
presented in this paper. 
To handle metonymy, Fass (1988) proposes a 
method based on a list of alternations imple- 
mented as specific metonymy rules: Part_for- 
_Whole, Container-for_Contents, etc. Sowa (1992) 
considers metonymies around the term "Prix Gon- 
court", originally introduced by Kayser (1988): 
this term undergoes different meaning shifts in 
each of seven example sentences, ranging from the 
author who won the prize to the amount of money 
received. Sowa discusses how background knowl- 
edge could help to process these metonymies, 
based on a knowledge description of what "Prix 
Goncourt" involves. 
In our system, the target conceptual representa- 
tion is defined by a domain model expressed with 
CGs. This same model constitutes the resource 
which enables the analyser to handle metonymies. 
We explain below how results similar to Puste- 
jovsky's type coercion may be obtained with a 
method based on this domain model instead of 
a qualia structure. 
3 Method 
3.1 Rationale 
The input to the semantic analyser is the syntactic 
representation of a sentence produced by a pre- 
vious large coverage syntactic analyser (B~rard- 
Dugourd et al., 1989). This representation con- 
nects words, or predicates, with grammatical rela- 
tions such as subject, object, oblique object, mod- 
ifier, etc. The output of the semantic analyser is 
a conceptual graph on which pragmatic inferences 
are performed to enrich the representation. 
In the semantic lexicon, each word points to one 
or more conceptual representations. The gram- 
matical link between two words in a sentence ex- 
presses a conceptual link between their two associ- 
ated conceptual counterparts. The task of the se- 
mantic analyser is to identify this conceptual link. 
Rather than including the knowledge needed for 
this task in the semantic lexicon, or in a specific 
rule base, the program will examine the domain 
knowledge to resolve the link. The method relies 
on a heuristic path search algorithm that exploits 
the graphic aspects of the conceptual graphs for- 
malism. 
3.2 Domain knowledge 
The main domain knowledge elements consist of 
the domain ontology (Fig. 1) which is a subsump- 
tion hierarchy of concept types (henceforth simply 
'types') and of relation types, and of a set of ref- 
erence models attached to the main types. 
The reference model of a type represents knowl- 
edge about this type as a conceptual graph 
(Fig. 2). Basically, a conceptual graph is a bi- 
partite graph with concept nodes (or concepts) 
labeled with a type plus an optional referent, and 
relation nodes labeled with relation types (Chein 
and Mugnier, 1992). A model of a given type has 
138 
Spatial  ote unetion Physie  object 
Inte,,tional O,,an e .... A. ment 
J 
Stenosis Angioplasty Velsel Artery-legment 
Artery Lad_Segment_II 
Figure 1: An extract of the domain ontology. 
an identified head concept with the same type, 
and the network of its related concepts represents 
its associated knowledge. Since types are organ- 
ised in an IS-A hierarchy, this knowledge is also 
inherited. 
Model Angioplasty(*x) is 
\[Angioplasty: *x\]- 
(pat)--,\[Human~eing:*pat\]-+ (cultural-function)---+ 
\[Medical_Sub function\] ---+ (cultural .rote) ---+ \[Patient\] 
(agt)-+ \[Human_Being:*doc\]-+ (cult ural-funetion)--, \[Medical_Subfuuction\] ---+ (cultural a'ole) --~ \[Physician\] 
(motive)-+ \[State_O f-Mind\]- (st ate_of)--~\[Human_Being:*doc\] 
(content)--+ \[Stenosis:*st 1\] % 
(purported..obj)--~ \[Artery~Segment :*as\]- 
(involves) +-\[St enosis:*st 1\] 
(involves) ~-- \[Int ernal-Stat e:*is3\] 
(par t) +- \[tIumau_Being:* pat\] % 
(descriptive..goal) +-\[Internal_State:*is3\]- 
Figure 2: An extract of reference model for type 
Angioplasty. 
3,3 Semantic lexicon 
The semantic analyser relies on a two-tier seman- 
tic lexicon: one for predicates, the other for gram- 
matical relations. Predicates map to conceptual 
graphs; most of them are reduced to one concept, 
since most of the words in the lexicon are techni- 
cal terms for which a type exists. Figure 3 reports 
some lexical entries. 
It is difficult to map grammatical relations 
to static, predefined conceptual representations, 
since their meaning in the domain depends on 
their context of use, and mostly on the predi- 
cates they link. Besides, one cannot think of 
envisioning all the possible uses of such a rela- 
tion, partly because of the use of metonymy. The 
conceptual representation of an actual grammat- 
ical link will therefore be computed dynamically 
by the semantic analyser using its context: the 
linked predicates and domain knowledge. How- 
ever, each grammatical relation may have concep- 
tual preferences for types or for conceptual rela- 
tions. These preferences are associated with the 
grammatical relation. Our grammatical relations 
include oblique complements, so that prepositions 
in our semantic lexicon are expressed under this 
second paradigm (Fig. 3). 
Entry angioplastie-f is \[Angiopiasty: *x\]. 
Entry stenose_f is \[Stenosis: *x\]. 
Entry segment-iI_f is 
\[SegmentAh*x\]- 
(relative_to)-+\[Artery\] 
(spatial.l"ole) +- \[Spatial_Object\] 
-+ (zone_of)-+ \[Artery ~egment\]. 
Gram~natieal-rel de- ¢ :prefers 
purported_obj involved_obj pat motivated_by before.state after-state rel. 
Figure 3: Some semantic lexicon entries for pred- 
icates and a grammatical relation. 
3.4 Algorithm 
Given an input triple predicate, grammatical rela- 
tion, predicate (P1; Gr; P'2), the semantic analyser 
first replaces the two predicates with their seman- 
tic entries -- two conceptual graphs. It then en- 
deavours to link them, that is, to find a concept- 
level relation between their two head concepts C1 
and C2 that, first, is compatible with the semantic 
preferences of grammaticM relation Gr, and, sec- 
ond, conforms to the representational canon made 
of the reference models. 
3.4.1 Design principle. 
The basic idea is to project the two head con- 
cepts onto the domain knowledge and find a 
plausible concept-level relation between the two. 
We implement this by heuristic graph traversal 
through the reference models and the type hierar- 
chy, looking for a chain made of concepts and con- 
ceptual relations (i.e. a linear conceptual graph), 
which could link concepts of the same types as C1 
and C2 and at the same time would satisfy the 
conceptual preferences of Gr. Semantic analysis 
then consists in solving recursively every gram- 
matical link starting from the sentence head pred- 
icate and then joining the obtained conceptual 
chains to build the conceptual representation of 
the whole sentence. We focus here only on the 
link resolution algorithm. 
3.4.2 Chain production methods. 
We consider that each predicate Pi is associated 
with the head concept Ci of a model Mi. Let Ti 
be the type of Ci. We also assume a partial order 
139 
on types. We focus here only on the strategy for 
i)roducing the set of all possible chains between Cl 
and C2. Wc can use three methods of increasing 
complexity to find chains to link C1 and C2: 
1. Concept fllsion: the two concepts may be re- 
dundant. 
If T1 < T2 or Tl > T2, then C, and 6'2 could 
be merged, and an empty chain is returned. 
2. Concept inclusion: a concept may be "in- 
cluded" in the other's model. 
(a) For every concept C' of type T' ill M1 
such that T' > T2, every path between 
Cl and C' in Mt is a returned chain. 
(b) For every concept C' of type T' in 3/& 
such that T' >_ Tt, (;very path in Mu be- 
tween C' and C9 is a returned chain. 
3. Model join: two arbitrary concepts in the two 
could be joined. 
For every pair of concepts (C\[, C~) where C~ 
of type T" is in Mi, and such that T\[ < T.~ 
or T\[ > T.~, all the paths Pathsl between C1 
and C~ in M, and Paths.2 between C~ and 
6'2 in \]1/\[2 are produced. Then, for every pair 
(Pt,P'2) in Paths1 x Paths2, the chain made 
of the two paths where last(p,) is joined to 
first(p.e) is returned. 
At this point, we are provided with all chains ex- 
tracted from the pair of models (MI, Me). 
3.4.3 Model identification. 
The models that associate knowledge to a given 
predicate P can be ranked according to their level 
of generality. The most specific model is the pred- 
icate definition in the semantic lexicon. The next 
one is the reference model associated with the type 
T of the head concept of the definition. Then, the 
following models are the reference models inher- 
ited along the ontology through supertypes of T. 
As the type hierarchy is, in our system, a tree 
(Bouaud et al., 1995), the models for a predicate 
are strictly ordered. Considering two grammati- 
cally linked predicates, the product of their mod- 
els constitutes as many model pairs that can be 
potentially used to look for possible chains. Such 
pairs are structured by a partial order based on 
the generality rank of their members, a 
3.4.4 Heuristic chain selection. 
At this stage, we are provided with all the pos- 
sibles chains between P1 and P2 extracted from 
their models. The remaining problem ix to choose 
tile most appropriate chain to substitute for Gr. 
After some experimentation, we chose the follow- 
ing scheme. The best chain ix selected accord- 
ing to five heuristic criteria: (1) satisfiability of 
aA model pair (To. 1, rn2) is more spe- 
cific than (rn\[, rn~) if max_rank(ml, m.~) is less than 
max_rank(m~, rn~), or if equal, rain_rank(m1, re.e) is 
less than min_rank(m~, m'2). 
Gr preferences; (2) most specific ,nodel pair, i.e., 
the use of most specific knowledge associated with 
words is prefered; (3) simplest chain production 
method (see 3.4.2); (4) most specific or high- 
est priority of Gr preferences; (5) shorter chain 
length. When inultiple chains remain in competi- 
tion, one is selected randomly. 
To reduce search, tile link resolntion strategy 
does not consider all possible chains, and imple- 
ments the first; two criteria directly in the chain 
production step. Chains that violate Gr prefer- 
ences are discarded, and model pairs are explored 
starting fi'om the most specific pair. 
3.5 An example 
Let us illustrate the, resolution on example (2) 
(an angioplasty of segment II). Tile inimt triple 
is (angioplastie_f;de_f;segment_iI_f). The corre- 
sponding types, Angioplasty and Segment_II, 
are not compatible and tile "fusion" inethod fails. 
The "inclusion" method also fails since no model 
for angioplastie_f includes a concept compatible 
with Segment_II, and no model for segment_ii_f 
includes a concept compatible with Angioplasty. 
However, with the "join" method, the algorithm 
identifies 6063 possible chains that satisfy the 
preferences attached to preposition des (Fig. 3). 
The selected chain uses the reference model of Angioplasty (Fig. 2) and tile definition graph for 
segment/I_f (Fig. 3) which are connected on con- 
cept trtery~qegment. The resulting conceptual 
representation .joins the two corresponding paths: 
\[Angioplasty\]-* (purported-obj)-+\[A rt cry_Segment\]. 
\[artery_Segment} ~--(zone_of)~--\[Spatial_O bjcct\] 
~ (spatial_role)-*\[Segment AI\], 
into 
\[Angioplasty\]-~ (purported-obj)-~\[Art cry_Segment 1 
~- (zone_of) ~- \[Spatial_Object\] 
-+ (spatial_role)--~\[Segment_I I\]. 
This representation reflects the fact that in the 
context of an 'angioplasty', 'segment II' is consid- 
ered from the point of view of the physical artery 
segment the angioplasty is to act upon (instead of 
the spatial notion Segment_II expresses). 
4 Implementation and results 
This analyser has been implemented on top of a 
conceptual graph processing package embedded in 
Common Lisp. In the current state, the ontol- 
ogy contains about 1,800 types and 300 relation 
types; over 500 types have their own reference 
model; the lexicon defines over 1,000 predicates 
and about 150 grammatical relations and prepo- 
sitions. The analyser correctly handles typical 
expressions found in our texts, including exam- 
ples (2)-(5) (see table 1). The complete process- 
ing chain has been tested on a set of 37 discharge 
summaries (393 sentences, 5,715 words) (Zweigen- 
baum et ~1., 1995). This corpus included devel- 
opment texts, so the results are somewhat opti- 
140 
Table 1: Conceptual rel)resentations obtained fl)r sentences (2) (5). 
(#) phrase total chains method models 
partial chains selected 
(2) 'angiot)lasty of segmenl, H' 6063 join Angiol)lasty 
\[Ailgioplasty\] ~ (imrl)or ted_oh j)-, \[At tery~%(~gnmnt \] 
\[Artery_Segment\] ~(zone_of),-\[Spatial_()l)ject\]--+ (slmt ial J'ole)-~ \[Segment J 1\] 
'segme.nt II' definition 
- (3) 'angiophtsty of a coronary artery' 2387 inclusion Angiol)lasty 
\[Angiol)lasty\]-+ (purlmr tedx)l)j)- + \[Ar t(n'y~eg ....... t\] ~-(pal't)~ .\[Coronary_Artery\] 
---(4) 'angioplasty of Mr X' 3633 inchlsion Angioplasty 
\[Angiophtstyl-,(p,,rported ml,j)-~\[Ar tery2qegme,~t\]~ - (part) +--\[llumanAteing l 
(5) 'angioi)lasty of a stenosis' 2217 
\[A ngiot)lasty\]-~ (purported. oh j) * \[hrtery~Seg ...... t\]~ -(i .... Ires) ,-\[Stenosis\] 
inclusion Angioplasty 
mistie; on the other hand, the systern is in an 
ilu:Oml/lete state of develolltnent. The test con- 
sisted in code a.ssignlne, t~t and answering a fix('.(\[ 
questionnaire, the gold standard being given by 
health (:are professionals. Overall recall and pre- 
cision were measured at /1:8 % and 63 % on the 
(:o(ling task, and 66 % and 77 % on the question- 
naire task. 
No ewfluation has been performed on 1here ba- 
sic components of the system; we can however 
provide statistics drawn from the global test for 
the semantic analyser. For 274 sentences received, 
the link resolution procedure was called on 8,749 
grammatical links and exI/lored 247,877 chains, 
with an average of 28 chains per call and 904 per 
sentence. The numbea" of paths found depends 
heavily on the richness of the lnodels used, which 
varies with the types involved, l%r instance, the 
model for type angioplasty (involved in table 1) 
is central in the domain. It is the most eoinplex 
in the knowledge base and (:ontain8 54 (:oneet)ts 
and 78 relations, which at:counts fl)r the, greater 
number of paths found in these examples. 
Ilowever, inadequate expai~sion8 are, SOlnetilnes 
made due to lack of lnodels, or to their complex 
ity, which makes the heuristic principles not se- 
lective enough. Such limitations also stem froin 
a lack of "actual" selnantic knowledge. The se- 
mantic analyser goes directly fi'om gralnmatical 
relations to concet)tua\] relations without any in- 
terme(liate selnantic ret)resentatioll. Usefll\] ilffor- 
lnatioll~ Sll(',h as the arglllnellt~tl or thelnati(: struc- 
ture of predicates (e.g. , Mel'(:uk et al. (1995), 
Pugeault et al. (1994)), could prol)a})ly overcome 
seine of its shortcomings. 
5 Discussion 
()IIC eouhl (;omtm.re this approach to a concel)t- 
based, multi-role qualia structure. The semantic 
definition of ~t word is here the reference model of 
its head concept type; each relation path starting 
fi'om the head eon(:ept of this reference model is 
similar to a qualia role, in that; it; describes one of 
the semantic facets or 1)ossible uses of the word. 
In the context of a predicate, one of the concepts 
in the reference model is selected as the incoln- 
ing point of a link from the predicate's inealfing 
representatk) n, 
The coneel)t-oriente, d domMmnlodel apl)roaeh 
advocated here hyI)othesizes that the behaviour of 
words is driven by their conceptuM ro|es in the do- 
main. This has the advantage of factoring knowl- 
edge at the conceptual level, rather than having 
to distribute it at the level of words. This knowl- 
edge can then be shared by severM words. Sharing 
even o(:(:urs across languages (e. 9. Dutch (Spyns 
and Willems, 1995)). 
Moreover, the type hierarchy Mlows concepts, 
hence words, to inherit reference models from 
more M)stract (:olmepts, thus enabling more sitar- 
ing mM modularity. The distinction between lo- 
cal information aim information inherited through 
the hierarchy in filrthermore exploited when rank- 
ing different chains between two concept types. 
Another differelme resi(tes in the way flexibility 
is obtained, in \]hlstejovsky's coercion ine(:hanism 
(Pustejovsky, 1991), the argument's semantic type 
changes for a semantic type found in one of its 
qualia. In a variant approach (Mineur and Buite- 
laar, 1995), a word has no a priori semantic type; 
it in selected at composition time among the types 
found in the qualia. In our approach, the head 
concept type associated with an argument does 
not change. The chain found between this con- 
cept and the predicate's head concept only brings 
forward internmdiate concepts and relations which 
are aetualised in th(; presence of the I)re(ticate, and 
lead to a particular representation of their lnt?an- 
ing. As a side-effect, this approach ix able to han- 
dle sentences like (6) (7): 
(6) dotm bought a h)ng nOV6`-I (Godard and ,layez, :\[993) 
(7) an aIlg'ioI)lasty of a sovere stenosis 
Since the modifier (long, sew',re) and the action 
(verb 'bought', noun 'angioplasty') require incom- 
patible types of the same noun (novel: event vs 
ot)ject, stenosis: state vs object), tyl)e changing 
via coercion cannot work on such sentences. This 
prol)lein does not occur in our approach. 
Type coercion assumes that the t)redi(:ate drives 
semantic eompositioll, and that the semantic rep- 
resentation of the argument inllst adapt to it. In 
241 
our method, both predicate and argument can 
make a step towards finding their semantic link. 
The resulting conceptual chain, as a whole, repre- 
sents both the specific facet of the argument which 
is involved in the sentence and the conceptual role 
it plays in the predicate. 
The preferences that grammatical relations as- 
sign to conceptual relations drive path selection, 
taking into account the specific syntactic context 
in which a semantic composition is to occur. This 
is crucial to let, e.g., prepositions, influence the 
choice of the conceptual link and the resolution of 
the metonymy. 
6 Conclusion 
The overall goal of the MENELAS text understand- 
ing system was to build a normalised conceptual 
representation of the input text. The aim of se- 
mantic analysis, in this context, is to build a repre- 
sentation which conforms to a domain model. We 
therefore experimented how this domain model 
could help semantic analysis to go from the flex- 
ibility of natural language to a constrained con- 
ceptual representation, a typical problem encoun- 
tered being metonymy. The approach presented 
here shows how this can be performed. It has 
been fully implemented, and used with a reason- 
able size knowledge base as a part of the MENELAS 
text understanding system. 
Metonymy processing is based on the domain 
model. Provided a new domain and task, with 
the corresponding domain model, this enables the 
generic method to adapt directly to this new do- 
main and give results that are specific to it. Build- 
ing such a domain model is generally feasible in 
sufficiently limited domains, typically, technical 
domains. Much of the strength of the method 
then hinges on the quality of the domain model: 
the concept type hierarchy and the attached ref- 
erence models must be built in a principled way 
(Bouand et al., 1995). 
References 
A. B@rard-Dugourd, J. Fargues, M.-C. Landau, 
and J.-P. Rogala. 1989. Un syst~me d'analyse 
de texte et de question/r~ponse bass sur les 
graphes conceptuels. In P. Degoulet, J.-C. 
Stephan, A. Venot, and P.-J. Yvon, editors, In- 
formatique et Gestion des Unitds de Soins, In- 
formatique et Sant~, chapter 5, pages 223-233. 
Springer-Verlag, Paris. 
Jacques Bouaud, Bruno Bachimont, Jean Charlet, 
and Pierre Zweigenbaum. 1995. Methodologi- 
cal principles for structuring an "ontology". In 
IJCAI'95 Workshop on "Basic Ontological Is- 
sues in Knowledge Sharing", August. 
M. Chein and M.-L. Mugnier. 1992. Con- 
ceptual graphs: fundamental notions. Revue 
d'InteUigence A rtificielle, 6(4):365-406. 
Denis Delamarre, Anita Burgun, Louis-Paul Seka, 
and Pierre Le Beux. 1995. Automated cod- 
ing system of patient discharge summaries us- 
ing conceptual graphs. Methods of Information 
in Medicine, 34:345-351. 
Dan Fass. 1988. Metonymy and metaphor: 
What's the difference? In Proceedings of the 
12 th COLING, pages 177-181, Budapest, Hun- 
gary. 
Danielle Godard and Jacques Jayez. 1993. To- 
wards a proper treatment of coercion phenom- 
ena. In Proceedings of the 6 th EACL, pages 
168-177, Utrecht, The Netherlands. 
Daniel Kayser. 1988. What kind of thing is a con- 
cept? Computational Intelligence, 4(2):158- 
165. 
Igor A. Mel'Suk, Andr~ Clas, and Alain Polgu~re. 
1995. Introduction ~ la lexicologie explicative et 
combinatoire. Duculot, Louvain-la-Neuve. 
Anne-Marie Mineur and Paul Buitelaar. 1995. 
A compositional treatment of polysemous ar- 
guments in categorial grammar. CLAUS Tech- 
nical Report 49, University of the Saarland. 
Also available by ftp on xxx.lanl.gov as cmp- 
lg/papers/9508/9508002. 
F. Pugeault, P. Saint-Dizier, and M.G. Mon- 
teil. 1994. Knowledge extraction from texts: 
a method for extracting predicate-argument 
structures from texts. In proc. Coling 93, Ky- 
oto. 
James Pustejovsky. 1991. Towards a generative 
lexicon. Computational Linguistic, 17(3):409- 
441. 
John F. Sowa. 1984. Conceptual Structures: 
Information Processing in Mind and Machine. 
Addison-Wesley, London. 
John F. Sown. 1992. Logical structures in the lex- 
icon. In James Pustejovsky and Sabine Bergleh 
editors, Lexical Semantics and Knowledge Rep- 
resentation, Lecture Notes in Artificial Intelli- 
gence, pages 39-60. Springer-Verlag, Paris. 
Peter Spyns and Jos L. Willems. 1995. Dutch 
medical language processing: Discussion of a 
prototype. In Robert A. Greenes, Hans E. Pe- 
terson, and Denis J. Protti, editors, Proc MED- 
INFO 95, pages 37-40, Vancouver. 
Pierre Zweigenbaum, Bruno Bachimont, Jacques 
Bouaud, Jean Charlet, and Jean-Francois 
Boisvieux. 1995. A multi-lingual architecture 
for building a normalised conceptual represen- 
tation from medical language. In Reed M. 
Gardner, editor, Proc 17th Annu Symp Com- 
puter Applications in Medical Care, New Or- 
leans, November. 
142 
