The Role of Cardinality in Metonymic Extensions to Nouns 
Helmut Horacek 
Universit~it des Saarlandes, FB 14 Informatik 
Postfach 1150, D-66041 Saarbrticken, Germany 
horacek@cs.uni-sb.de 
Abstract 
Meaning shifting phenomena such as metonymy have recently attracted increasing interest 
of researchers. Though these phenomena have been addressed by a variety of computa- 
tional methods, the impacts of cardinalities of metonymically related items have been 
widely ignored in all of them. Motivated by this lack of analysis, we have developed a 
method for representing expectations and knowledge about the cardinalities of metonymi- 
cally related entities and for exploiting this information to build logical forms expressing 
metonymic relations, the entities related, and their cardinalities as precisely as possible. The 
representation of lexically motivated knowledge is realized as an enhancement to Puste- 
jovsky's Generative Lexicon, and the process of building logical forms takes into account 
overwriting of default information and mismatch of cardinality requirements. Our method 
enables a precise attachment of sentence complements, and it provides better prerequisites 
for reference resolution in the context of metonymic expressions than previous approaches. 
1 Introduction 
Meaning shifting phenomena such as meto- 
nymy have recently attracted increasing interest 
of researchers. Computational methods 
addressing these phenomena aim at inferring 
implicitly represented relations, predicting 
meaning shifts of nouns or verbs, expressing 
restrictions on these meaning shifts in depen- 
dency of context- or language-specific factors, 
and facilitating reference resolution. Measu- 
rements to achieve these issues include represen- 
tation of default knowledge, and various sorts of 
dedicated inference methods and constructive 
procedures. However, the entities in the texts 
examined almost always appear in singular form 
so that issues of cardinality of metonymically 
related items have been widely ignored by the 
approaches undertaken so far. 
Motivated by this lack of analysis, we have 
examined sentences taken from the literature 
about metonymic phenomena, and we have 
modified them by varying cardinalities of the 
items appearing explicitly or implicitly, to 
analyze effects of these alternations. The results 
have inspired us to undertake extensions to 
other techniques, such as lexical representations, 
to build increasingly explicit versions of logical 
forms, and to formulate conditions on prono- 
minal accessibility, The insights gained are 
likely to improve analysis methods for meto- 
nymic expressions, especially for relating 
contextual specifications to the appropriate 
entity, and for supporting reference resolution 
to entities related implicitly. 
This paper is organized as follows. First, we 
review major computational approaches to 
metonymy. Then we illustrate the phenomena 
investigated by a number of typical examples. 
We continue by elaborating techniques to deal 
with these phenomena, that is, an enhancement 
to entries in the Generative Lexicon, and a 
procedure for building a logical form that 
appropriately represents relevant aspects of 
metonymic expressions. Finally, we discuss im- 
pacts of our analysis on pronominal resolution. 
2 Approaches to Metonymy 
Metonymy belongs to a variety of natural 
language phenomena that contribute to express- 
ing information in an effective and economic 
way. All these phenomena involve what has 
been termed 'transfers of meaning' by (Nunberg, 
1995), i.e., the meaning of some constituent 
does not correspond to what can usually be 
expected according to the syntactic and seman- 
tic environment. Metonymy, or semantic coer- 
cion, is usually defined as a figure of speech in 
which the speaker is "using one entity to refer to 
103 
another that is related to it" (Lakoff and 
Johnson 1980). For example, in the utterance 
"The ham sandwich is waiting for his check", it 
is not literally the ham sandwich, which wants to 
pay, but the person who ordered it. 
Motivations to address metonymy include 
both theoretical insights and practical appli- 
cations (cf. (MACDOW 1992) and the natural 
language database interface TEAM (Grosz et al. 
1987)). Computational approaches are mostly 
concerned with inferring implicitly expressed 
metonymic relations in English texts - (Fass 
1991), Hobbs (Hobbs et al. 1993) (Lakoff and 
Johnson 1980), (Nunberg 1993, 1995), (Puste- 
jovsky 1991), and (Wilks 1975) are prominent 
representatives. Some analyses also consider 
French (Kayser 1988), (Pustejovsky and, 
Bouillon 1995) and German (Horacek 1996). 
In his program met*, (Fass 1991) makes use 
of formal definitions of several kinds of meto- 
nymic relations; met* also allows chaining 
metonymic relations in order to fill in implicitly 
expressed knowledge. Sowa's conceptual graphs 
(Sowa 1992) are used for inserting an unknown 
relation between a concept of the type expected 
and the concept appearing on the surface, which 
is later filled on the basis of world knowledge 
accessible to the system. The TACITUS system 
(Hobbs, Martin 1987) uses similar methods for 
dealing with metonymy and for interpreting 
noun-noun components, which are considered 
special cases of reference resolution - that 
approach, which is also described in (Hobbs et 
al. 1993), treats interpretation as a uniform 
abduction process to find the best explanation 
for the observables'. These approaches work 
nicely for analyzing utterances of the kind 
considered by inserting a plausible relation to 
remove a constraint violation, and they have 
similar and characteristic properties: 
• The conditions expressing when leaving a 
metonymic relation implicit or not is 
possible are too unconstrained to cover a 
larger number of examples in several 
languages. 
• The entities involved, the real and the literal 
referent, always appear in singular form. 
There are only two approaches which in some 
aspects deviate from this characterization: 
• Pustejovsky's Generative Lexicon (Puste- 
jovsky 1991) addresses the first aspect. He 
proposes a Theory of Qualia within a 
Generative Lexicon, which enables the 
explanation of systematic polysemy. 
Applying type coercion enables one to 
arrive at cases of ordinary metonymy which 
can be grounded in terms of the semantics 
of lexemes, as well as at word senses which 
Pustejovsky has termed logical metonymy, 
like the reading of a book in the sentence 
"Mary enjoyed the book". Enhancing the 
semantic representation of a noun within 
such contexts is done by exploiting proto- 
typical knowledge derived from AGENTIVE 
or TELIC roles of the lexical entry for 
'book', which are prominent roles in the 
Qualia Structure of lexical entries for 
nouns. The accuracy of the theory has been 
extended by the incorporation of restric- 
tions on the productivity of lexemes. Parti- 
cularities of the Qualia Structure of nouns 
regulate the acceptability or unacceptability 
of leaving a metonymic relation implicit in 
context of the words engaged (McDonald, 
Busa 1994, Pustejovsky, Bouillon 1995). 
• (Stallard 1993) indirectly addresses the 
second aspect by taking into account 
scoping relations and consequences for 
pronominal reference. He has introduced a 
distinction between referential and predica- 
tive metonymy, depending on whether the 
actual or the literal argument is accessible 
for subsequent pronominal reference. This 
distinction manifests itself in different 
scope relations that hold between the actual 
and the literal argument in the corres- 
ponding logical forms. Though we do not 
agree with his usage of scoping and the 
resulting strict distinction of pronominal 
accessibility, Stallard's approach to build 
logical forms has inspired our techniques. 
Though neither Pustejovsky's nor Stallard's 
approach address the role of cardinalities, we 
show that both of them can be extended 
accordingly: the Generative Lexicon can be 
augmented to represent knowledge about cardi- 
nality information associated with the semantics 
of nominals, and techniques similar to those 
used by Stallard can be set up for building 
logical forms with more precise cardinality 
specifications of the metonymically related 
entities. But before we expose these methods in 
detail, we prepare the ground for this enterprise 
by discussing a set of sentences illustrating the 
phenomena we intend to investigate. 
104 
3 Phenomena Investigated 
For a number of metonymic relations, such as 
PRODUCER for PRODUCT ('I bought a Ford') 
and ARTIST for ARTWORK ('He plays Beet- 
hoven'), cardinalities are never a problem 
because the literal referents are expressed as 
proper names. Similar considerations apply to 
the eventualities involved in logical metonymy. 
For other metonymic relations, especially PART 
for WHOLE and ORGANIZATION for MEMBER, a 
number of complications may arise due to the 
• cardinality of the items involved in a metonymic 
expression, as the following examples demon- 
strate. Let us start with two contrastive sentences 
(1) and (2), taken from (Hobbs et al. 1993), and 
(Lakoff and Johnson 1980), respectively (an 
earlier version of the subsequent analysis is 
available in (Horacek 1994)): 
(1) The ham sandwich is waiting for his 
check. 
(2) The Boston office called. 
Following Stallard, (1) is interpreted as an 
example of referential reading, while (2) as an 
example of predicative reading: (1) can be 
rephrased more explicitly by The manx who has 
eaten a ham sandwichy is waiting for hisx check, 
while (2) in a similar Way gets expanded to The 
Boston officex represented by one of itsx 
employeesy called: These reformulations 
suggest that the man m (1) and the Boston office 
in (2) have wider scope in Stallards represen- 
tation than the ham sandwich in (1) and the 
employee in (2), which predicts pronominal 
accessibility in (la) and (2b), as opposed to (lb) 
and (2a). We challenge this analysis in its strict 
sense, but we agree with it insofar as pronominal 
references as in (lb) or (2a) are rare, but quite 
common in sentences such as (la) and (2b). 
(1) The ham sandwich is waiting for his 
chec!<. 
(1 a) He is getting impatient. 
(lb) It is25. 
* (lc) They are getting impatient. 
(2) The Boston office called. 
(2a) He was angry. 
(2b) It is our head quarter. 
(2c) They want us to organize a meeting. 
Apart from this difference, there do not seem 
to be further complications in these sentences: 
all referents involved are in singular form, as are 
the metonymic extensions. Since cardinalities of 
the entities involved are identical, conditions 
about pronominal reference depend primarily 
on pragmatic factors, which make references 
such as (la) and (2b) more common than those 
in (lb) and (2a). However, more complications 
than the analysis made so far has shown arise, 
when variations of cardinality in sentence (1) 
(see sentences (3) to (6) and their follow-ups), 
and variation of circumstances in sentence (2) 
(see the follow-ups of sentences (7) and (8)) are 
considered. For dishes made of animal ('the 
mussels'), additional complications may arise 
through interference between animals and 
persons as pronominal referents. Because we 
want to study the effects of cardinality variations 
per se, we avoid examples of this sort. 
(3) The pizzas are waiting for their checks. 
? (3a) He/she is getting impatient. 
(3b) They are getting impatient. 
(4) The fruit dumplings want(s) to pay. 
(4') The fruit dumplings is waiting for 
his/her check. 
(4a) He/she is getting impatient. 
(4") The fruit dumplings are waiting for 
their check(s). 
(4b) They are getting impatient. 
(4'") Die Fruchtknrdel warten auf ihre 
Rechnung. 
(5) The meat plate want(s) to pay. 
(5') The meat plate are waiting for their 
check(s). 
(5a) They are getting impatient. 
(5") The meat plate is waiting for his/her 
check. 
(5b) He/she is getting impatient. 
(5"') Die Schlachtplatte wartet auf ihre 
Rechnung. 
(6) Table 7 want(s) to pay. 
(6') Table 7 are waiting for their check(s). 
(6a) They are getting impatient. 
(6") Table 7 is waiting for his/her check. 
(6b) He/she is getting impatient. 
(6'") Tisch 7 wartet auf seine Rechnung. 
These sentences demonstrate that both intra- 
((1) and (3)) and intersentential ((la) and (3b)) 
prononminal reference work fine, if the literal 
referents (here, various sorts of food) and the 
real referents (here, the persons) agree in 
number. Otherwise, a variety of complications 
arise in intrasentential reference, which also 
include cross-language differences, as the 
German sentences (4"') to (6'") as opposed to 
the English sentences (4') to (6') and (4") to (6") 
demonstrate. In these sentences, a fundamental 
difference manifests itself in the way how meto- 
nymic expressions are syntactically embedded 
in these languages. In English, it is the intended 
referent that determines verb agreement and 
pronominal reference in the same sentence. 
That is, the singular or plural form of the verb 
•105 
in (4' to (6') and (4") to (6") indicates the cardi- 
nality of the intended referent, which may 
deviate from the cardinalities of the literal 
referents in each of these sentences. In (5'), the 
cardinality of the intended referent naturally 
differs from the number of the literal referent, 
assuming the meat plate is known as a dish for 
more than one person. For the fruit dumplings, 
situations with single (4') as well as with multiple 
intended referents (4") are possible, because the 
expression 'fruit dumplings' is ambiguous in the 
sense that it can refer to one plate of dumplings 
to be eaten by a single person, or to several 
plates, each for another person. Moreover, 
sentences (6') and (6") illustrate a case similar to 
sentences (4') and (4"). As a tendency, a single 
intended referent would be the default inter- 
pretation for the metonymic use of 'fruit 
dumplings', while 'table' seems to be more 
neutral with respect to the number of persons 
sharing it. As with the verb agreement, also the 
possessive pronouns are always agreed in 
.gender and number with the intended referent 
in English. In contrast to that, the verb and 
possessive pronouns in German appear in strict 
agreement with the literal referents in the corres- 
ponding sentences (4"') to (6'"). Altogether, 
English sentences of this sort are more conclu- 
sive by making cardinalities and, in case of 
singular, also gender of the intended referent 
explicit, while the German sentences are ambi- 
guous in these respects. 
Unlike with intrasentential reference, intersen- 
tential pronominal reference with number 
features deviating from the referent that is 
pronominally accessible intrasententially is 
possible also in German under a variety of 
circumstances. These include default expec- 
tations about the cardinality of the real referents 
(see (4a) and (5b) as two complementary cases), 
and may perhaps be considered felicitous in 
other cases (see (5a)), similarly to (3a). The 
remaining case, as exemplified by sentence (4b), 
is felicitous for similar reasons as the different 
cases (4') and (4"). Finally, sentences (6') to (6b) 
constitute further pieces of evidence that default 
expectations about the cardinality of metonymi- 
cally related entities (here, persons as eaters of 
kinds of food) play a certain role as to whether 
pronominal reference to metonymic expressions 
is felicitous or not, but not in a strict sense. In 
addition, (6b) is probably more common than 
(6a) as a default, due to the pragmatics of 
paying and plausibilities about table sharing, 
especially about food sharing according to 
reasonably assumed food quantities, seem to 
influence felicity in a subtle, hardly generaliz- 
able way in these sorts of sentences. For a 
member of a group as in (3b) (one of the pizza 
eaters), the pragmatics of paying (a single 
person on behalf of several ones) may even 
license the use of the singular form as in (3a). 
However, a somehow chained metonymic refer- 
ence to a group as in (lc) (to which the eater of 
the ham sandwich belongs) is certainly not 
possible. 
(7) The Boston office is represented in the 
business meeting. 
? (Ta) He/she is an expert in marketing. 
(7b) They are experts in marketing. 
(7c) They always send someone to 
important meetings. 
(8) The Boston office will meet for an 
excursion on Friday. 
* (Sa) He/she likes to walk. 
(Sb) They will make a lunch break at 2 pm. 
(8c) They like to organize social events. 
In contrast to the examples discussed so far, 
sentences in the next group ((2), (7), (8), and 
their follow-ups) involve slightly harder 
restrictions. These are, however, no cross- 
language differences, because all references in 
these sentences are intersentential. The example 
sentences demonstrate certain complications in 
metonymic uses of 'office' in contrast to 
metonymic uses of 'sandwich', which originate 
from the differences in the underlying relations 
between food and the persons eating it, as 
opposed to the office and the caller who is 
related in a more indirect or perhaps more 
pronounced way to the office than persons are 
to the food they eat. Plural pronominal refer- 
ences as in sentences (7b), (7c), (8b) and (8c) 
are felicitous, but there is a difference between 
the sets of entities the plural pronouns refer to. 
While in (7c) and (8c), the pronouns refer to the 
entire set of employees of the Boston office, 
they more plausibly refer to the representatives 
in the meeting in (7b) and to the excursion 
participants in (8b). In any case, these examples 
indicate an additional demand on :the treatment 
of cardinalities and referential accessibility of 
metonymic expressions: a distinction is to be 
made between the entities referred to metony- 
mically (here: employees of the Boston office), 
and those of its member involved in the event 
expressed by the sentence (here: the meeting 
and the excursion). For the restaurant scenario, 
these sets of persons are mostly identical except 
to those cases where one person out of a group 
of persons eating together and referred to 
metonymically is the one who intends to pay. 
(9) Which airlines serve diet food from 
Boston to New York? 
(9a) In the first class? 
106 
food(x) 
CONST = {food ingredientsl 
FORMAL = eatable(x) 
TELIC = eat(eT,y,x) 
AGENTIVE = cook(e'r,z,X) 
sandwich(x) 
CONST = {ham, bread, .. } 
FORMAL = eatable(x) 
TELIC = eat(er,y,x) 
AGENTIVE = prepare(e'T,z,x) 
pizza(x) 
CONST = {dough, tomato, ... 
FORMAL = eatable(x) 
TELIC = eat(eT,y,x) 
AGENTIVE = bake(e'r,z,x) 
Figure 1: Some 'standard' examples of Qualia Structures, for the nouns 'food', 'sandwich', and 'pizza' 
As a further aspect of metonymic expressions, 
the last set of examples demonstrates chaining 
of metonymic relations and the relevance of 
each set of items involved for the associated 
analysis. In sentence (9), the airlines are the 
literal, and the persons the real referents. 
However, relating these two entities directly by 
an employment relation is problematic, since it 
is impossible to connect the locality information 
(from Boston to New York) and the first class 
restriction to either of them. Linking this infor- 
mation to the airline would decontextualize the 
serving process, and linking it to the persons 
would make the serving process independent of 
the context of a flight and the person's working 
for the airline. Therefore, it is more appropriate 
to elaborate the relation between the airlines and 
their employees to include the implicitly 
referred flights explicitly. These flights, of 
course, are the items that are first class and go 
from Boston to New York. Note, that linking the 
locality context properly is essential for setting 
up correct database requests, at least for requests 
to databases not restricted to flights only. 
4 Expressing Lexical Knowledge 
In order to capture distinctions between the 
varying interpretations of metonymic 
expressions illustrated in the previous section, 
knowledge about the lexical items involved 
plays a crucial role. For adequately expressing 
this knowledge, we make use of entries in the 
Generative lexicon (see Figure 1). Since the 
information represented therein is insufficient 
for reasoning about cardinalities, we extend the 
entries in the Generative lexicon, prominently 
the TELIC role, by quantifier specifications. In 
the original form, the entities involved (typi- 
cally, the lexical item itself and some related 
entity) are implicitly quantified, and a typed 
event variable is used (an event may be a state 
(S), a process (P), or a transition (T)). A similar 
exploitation of taxonomic knowledge in terms 
of cardinality restrictions has been exploited for 
scope disambiguation in (Fliegner 1988). 
In the extended form (see Figures 2 and 3), 
we introduce explicit quantifiers, and we option- 
ally add restrictors to variables referred to by 
events predicates. In addition, the scoping of 
quantifiers allows the derivation of cardinalities 
- see the entries for FRUIT-DUMPLING and 
MEAT-PLATE, as contrasting examples. We 
distinguish several types of quantifiers to cover 
the cases elaborated in the previous section, in 
addition to the standard quantifiers EXIST and 
WH (the first two constitute default information, 
and the others express definitional restrictions): 
FORMAL 
TELIC 
fruit-dumpling(x) 
CONST = {dough, fruit, ... I 
= eatable(x) 
= (DEFSINGLE y 
(DEFMULTIPLE x 
(eat(er,y,x)))) 
AGENTIVE = cook(e'r,z,x) 
FORMAL 
TELIC 
meat-plate(x) 
CONST = {pork, beef, .. } 
= eatable(x) 
= (DEFSINGLE x 
(DEFMULTIPLE y 
(eat(eT,y,x)))) 
AGENTIVE = prepare(e'T,z,x) 
table(x) 
CONST = {legs, plate .... } 
FORMAL = physobj(x) 
TELIC = (DEFSlNGI E x 
(DEFMULTIPLE y 
(sit-at(eS,y,x)))) 
AGENTIVE = build(e'T,z,x) 
Figure 2: Some 'extended' examples of Qualia Structures, for special food sorts and for 'table' 
107 
• DEFSINGLE for a single object 
• DEFMULTIPLE for multiple objects 
• SINGLE for a single object 
• MULTIPLE for a multiple objects 
Figures 2 and 3 show entries in the Generative 
Lexicon with extended TELIC roles. In principle, 
the same extensions also apply to the AGENTIVE 
roles, but we do not elaborate these extensions 
because we do not make use of these roles for 
our current purpose. Figure 2, for example, 
shows that some sorts of food are associated 
with different expectations about how many 
persons typically eat them. Frt::t dumplings 
appear as sets (quantified by DEFMULTIPLE), to 
be eaten as a dish by a single person (quantified 
by DEFSINGLE). In contrast to the fruit 
dumplings, cardinality relations are inverted for 
the meat plate. Similar, but weaker default 
assumptions hold for a table in a restaurant. 
In contrast to the restaurant scenario, cardina- 
lity relations are less vague and associated with 
defaults for elementary relations in organi- 
zations. Each office, and airlines in particular, 
are supposed to employ a set of persons larger 
than one, and each person is working for one 
organization only, at least in his/her individual 
activities (hence, the quantifiers SINGLE and 
MULTIPLE in the lexical entries shown in Figure 
3). Similarly, each flight carries some set of 
people, each of which participates in one flight 
only (at the same time). 
The knowledge contained in these lexicon 
entries can be used for building logical forms 
representing metonymic expressions of the kind 
discussed in section 3. The event predicates in 
the TELIC roles (and, less frequently, in the 
AGENTIVE roles) are exploited to infer the 
relation involved, which works similarly to other 
approaches. Moreover, the new quantification 
specification yields the otherwise missing source 
of information to build an explicit logical form 
with cardinality specifications from concise 
surface expressions in a precise manner. 
5 Building Logical Forms 
Based on entries in the Generative Lexicon and 
on the context given by a sentence to be inter- 
preted, appropriate logical forms can be built 
that represent semantic relations involved more 
explicitly than this is the case with previous 
approaches. These logical forms take into 
account sorts of the entities involved and sorts 
of case frame slots in which these entities 
appear, syntactic information about number, 
and default expectations or semantic restrictions 
about the cardinality of these entities, the latter 
aspect being a new feature in our approach. In a 
nutshell, metonymic extensions are tried 
according to specifications found in the lexicon, 
as long as the sort of an NP and the sort of the 
referring case role are incompatible. In addition, 
agreement between syntactic number and 
semantic cardinality specifications is achieved, 
which may require overwriting defaults or intro- 
ducing a new set of entities as a subset of those 
already introduced. In concrete, logical forms 
are built by pursuing the procedure sketched in 
Figure 4, with step 2a performing metonymic 
extensions, and step 2c performing a final 
extension in case of a cardinality mismatch. In 
the following, we illustrate the procedure by 
some examples. Consider the sentence 
(4) The fruit dumplings wants to pay. 
office(x) 
CONST = {employees .... } 
FORMAL = organization(x) 
TELIC = (SINGLE x 
(MULTIPLE y 
PERSON 
(work(e p,y,x)))) 
AGENTIVE = establish(e'r,z,x) 
airline(x) 
CONST = {planes, office, .. } 
FORMAL = organization(x) 
TELIC = (SINGLE x 
(MULTIPLE y 
FLIGHT 
(organize 
(eT,y,x)))) 
AGENTIVE = found(e'r,z,x) 
flight(x) 
CONST = {place, source .... } 
FORMAL = location-change(x) 
TELIC = (SINGLE x 
(DEFMULTIPLE y 
PERSON 
(carry(e r,y,x)))) 
AGENTIVE = organize(e'r,z,x) 
Figure 3: Some 'extended' examples of Qualia Structures, for the nouns 'office', 'airline', and 'flight' 
108 
1. Build an initial logical form from the surface expression. 
The representation is composed as an expression of the form (Qs xE Sz <P>), with 
XE being the variable whose representation is to be extended 
(initially equal to x, denoting the literal referent), 
QE being its quantifier, and SE its sort 
(initially equal to Q and S, which are associated with the literal referent), and 
<P> being an eventually structured representation of the sentence predicate and its modifiers. 
Moreover, the conflicting restrictions are stored. Let 
SR be the sort required within the referring case frame, and 
QR the quantifier representing the associated case slot restrictions. 
2. Extend the meaning of noun phrases where a sortal incompatibility occurs. 
2a. Build a metonymically extended expression through consultation of lexical knowledge. 
Merge the partial expression (QE xE SE <P>) with the extended representation found in the 
lexical entry for SE (typically under the TELIC role), which structurally yields 
(Q~ x~ S~ (AND <P~> (Q2 x2 $2 <P2>))) - that is, the structure taken from the lexicon, with 
Q1 = QE, xl = xz, S1 = SE, and Q2 = QN, x2 = XN, $2 = SN, 
if the referent with the same sort as xE has wider scope in the lexicon entry, or 
Q2 = QE, x2 = xe, S2 = Se, and Q1 = QN, xl = Xs, S1 = SN, 
if the referent with the same sort as xz has narrower scope in the lexicon entry. 
<P> is partitioned according to sortal compatibility of its components, as follows 
if xl = xE 
then <P~> contains parts that refer to xE and are sortally compatible with S~, 
else <P2> contains parts that refer to xE and are sortally compatible with SE. 
Similarly, the remaining parts of <P> become <P2>, if xl = XE, and <P~> otherwise. 
2b. Test the compatibility of the newly inserted sort with the restrictions to be met. 
If SN (SN = $2, if S~ = SE, and SN = S1 otherwise) is again incompatible with SR, 
then repeat step 2a with Xs, SN, QN and <PN> as XE, SE, QE and, <P>, respectively 
(XN, QN and <PN> have the same index as Ss); 
else QN is overwritten by QR if QN is a default quantifier compatible with QR. 
2c. Test the cardinality compatibility of the newly inserted sort with the restrictions to be met. 
If SN is compatible with SR, but Qr~ is incompatible with QR, 
then the expression is expanded as under 2a, by a MEMBER relation between xe and xN. 
If not the whole set of entities bound to xz participates in the eventuality, 
then an expansion is performed with a SUBSET relation instead of a MEMBER relation. 
Figure 4: The procedure for building logical forms with extended metonymic relations 
The initial logical form directly built from the 
surface expression simply is 
(MULTIPLE x FRUIT-DUMPLING 
(WANT-PAY x)) 
which contains a sortal incompatibility. 
Making use of the lexical entry for 'fruit 
dumplings' and expanding the expression 
according to the TELIC role (see Figure 2) 
yields 
(SINGLE y PERSON 
(MULTIPLE x FRUIT-DUMPLING 
(AND (EAT y x) 
(WANT-PAY y)))) 
109 
• More referents than just the real and the 
literal referent may be introduced, either 
through chained metonymic extensions or 
through membership/subset insertions. 
• An additional inserted referent may provide 
a proper place to relate sentence 
complements (e.g., locality information to 
flights rather than to airlines or to persons). 
Note, that there is a scoping difference in the 
expressions underlying the phrases 'one and the 
same person eating several fruit dumplings' and 
'several persons sharing a meat plate' (the 
default interpretations), which is in contrast to 
the approach by (Stallard 1993), Moreover, the 
additional referents may not only improve the 
basis for complement attachment, but also for 
pronoun resolution. Both aspects are briefly 
discussed in the next section. 
6 Impacts on Reference Resolution 
Empirically supported by the considerable 
number of examples discussed in section 3, our 
approach is able to explain more pronominal 
references to metonymic expressions than 
others. This achievement is based on the 
following factors we have examined: 
• Reference to literal and intended referents is 
possible in an increasing number of cases. 
• Pronominal reference in plural form may 
have as antecedents distinguished sets of 
entities that are associated with a metonymic 
expression. 
• There are cross-language differences 
between German and English in the 
treatment of intersentential pronominal 
reference. 
In order to express scoping relations among sets 
properly, the logical forms representing meto- 
nymic expressions with entities of cardinality 
greater than one must deviate from Stallard's 
methods. According to (Stallard 1993), prono- 
minal reference to literal and real referents is 
regulated by their scope, which distinguishes 
referential from predicative kinds of metonymy. 
Unfortunately, this realization of metonymic 
extension is incompatible with the common use 
of scoping. However, we believe that Stallards 
distinction is in some sense artificial, because the 
felicity of pronominal reference seems to be 
more complex and influenced by other factors 
than scoping. For example, the sentence 'the 
ham sandwich is waiting for his check' can be 
followed by some information useful to a 
novice waiter: 'It costs 25.' Moreover, the 
message 'The Boston office called' can be 
followed by the remark 'He spoke angrily' in 
some plausible contexts. Hence, it does not seem 
to be referential inaccessibility which makes 
many similar examples sound odd, but the rare 
occurrence and the eventual low coherence in 
neutral contexts. For example, it is usually of 
minor interest whether the person calling on 
behalf of the Boston office is angry himself; it is 
the attitude of the responsible representatives at 
the office that is usually more interesting, since 
this is usually considered an influential factor 
regarding the content of the message. 
Given these pieces of evidence, reference reso- 
lution is essentially supported by the explicit 
logical form built through our techniques, and it 
is additionally guided as follows: 
lntrasentential reference 
Possessive pronouns always relate to the 
intended referent, which is accessible through 
the logical form. Since possessive pronouns in 
the same sentence agree in gender and number 
with the real referent in English, while they 
agree with the literal referent in German, only 
English sentences contain additional infor- 
mation about cardinality and gender of the 
intended referent. For example, the sentence 'the 
fruit dumplings is waiting for his check' carries 
the additional implication that there is one male 
person who wants to pay. 
Intersentential reference 
Reference through personal pronouns is 
possible to the literal and to the real referent, 
and to referents of the same sort but with 
possibly different cardinality as the real 
referent. Thus, all entities involved in a meto- 
nymic expression in its appearance in the 
explicit logical form are potential antecedents, 
except to some internal elements of a meto- 
nymic chain, so that a dialog memory should be 
maintained accordingly, For example, following 
the sentence 'The Boston office called', prono- 
minal reference is possible to the office (the 
literal referent), to the caller (the real referent), 
and to the people at the office (differing from 
the caller by number only). However, 'the 
flights' appearing in the extended logical form 
representing the sentence 'Which airlines serve 
diet food from New York to Boston?' are not 
pronominally accessible. 
111 
7 Conclusion 
In this paper, we have presented an approach to 
deal with cardinality aspects of metonymic 
extensions to nouns. We have discussed a variety 
of constellations with pronominal references to 
implicitly related items, sometimes associated 
with subtle conditions, including cross-language 
differences between English and German. In 
order to build explicit logical forms with cardi- 
nality specifications, we have extended entries in 
Pustejovsky's Generative Lexicon by default 
quantifier specifications, and we have described 
methods for exploiting these entries accord- 
ingly: Metonymic extensions are introduced on 
the basis of events represented in the roles of the 
Qualia structure, and member or subset relations 
are introduced on the basis of the associated 
quantifier specification. Our method for build- 
ing explicit logical forms challenges Stallard's 
distinction of predicative and referential read- 
ings of metonymic expressions: it produces dif- 
ferent scopings that reflect proper quantifier 
dominance relations rather than pronominal 
accessibility conditions, and it allows for addi- 
tional cases of pronominal reference. In addi- 
tion, our method enables a more precise attach- 
ment of contextual specifications to related 
entities, and it yields a better basis for reference 
resolution to metonyrnically related entities. 

References 
Dan Fass. 1991. met*: A Method for Discrimi- 
nating Metonymy and Metaphor by Compu- 
ter. Computational Linguistics 17(1):49-90. 
Michael Fliegner. 1988. HOKUSKOPUS - Ver- 
wendung terminologischen Wissens bei der 
Analyse von Quantorenskopus und Distribu- 
tivit/it. In Proc. of GWAI-88, pages 112-117. 
Barbara Grosz, Doug Appelt, Paul Martin, and 
Femando Pereira. 1987. TEAM: An Exper- 
iment in the Design of Transportable Natural- 
Language Interfaces. Artificial Intelligence 
32:173-243. 
Jerry Hobbs, and Paul Martin. 1987. Local 
Pragmatics. In Proc. of IJCAI-87, pages 520- 
523, Milan, Italy. 
Jerry Hobbs, Mark Stickel, Doug Appelt, and 
Paul Martin. 1993. Interpretation as Abduc- 
tion. Artificial Intelligence. 
Helmut Horacek. 1994. Some Issues in Dealing 
with Metonymy. In Proc. of KONVENS-94, 
pages 171-180, Vienna, Austria. 
Helmut Horacek. 1996. On Expressing Meto- 
nymic Relations in Multiple Languages. 
Machine Translation 11:109-158. 
Daniel Kayser. 1988. What Kind of Thing is a 
Concept. Computational Intelligence 4(2): 
158-165. 
George Lakoff, and M. Johnson. 1980. Meta- 
phors We Live By. Univ. of Chicago Press. 
MACDOW Committee (L. Hirschman et al.). 
1992. Multi-Site Data Collection for a 
Spoken Language Corpus. In Proc. of Speech 
and Natural Language Workshop. 
David McDonald, and Frederica Busa. 1994. On 
the Creative Use of Language: the Form of 
Lexical Resources. In Proc. of 7th Inter- 
national Workshop on Natural Language 
Generation, pages 81-89, Kennebunkport, 
Maine, USA. 
Geoffrey Nunberg. 1993. Transfers of Mean- 
ing. In Proc. of ACL-93, pages 191-192, 
Columbus, Ohio, USA. 
Geoffrey Nunberg. 1995. Transfers of 
Meaning. Journal of Semantics 12:109-132, 
Oxford University Press. 
James Pustejovsky. 1991. The Generative Lexi- 
con. Computational Linguistics, 17(4):409- 
441. 
James Pustejovsky, and P. Bouillon. 1995. 
Aspectual Coercion and Logical Polysemy. 
Journal of Semantics 12:133-162, Oxford 
University Press. 
John Sowa. 1992. Logical Structures in the 
Lexicon. In J. Pustejovsky, S. Bergler (eds.): 
Lexical Semantics and Knowledge Represen- 
tation, pages 39-60, Springer. 
David Stallard. 1993. Two Kinds of Metonymy. 
In Proc. of ACL-93, pages 87-94, Columbus, 
Ohio, USA. 
Yorick Wilks. 1975. An Intelligent Analyzer 
and Understander of English. Communi- 
cations ACM 18(5):264-274. 
