On The Semantic Interpretation of Nominals 
James PUSTEJOVSKY Peter G. ANICK 
.Department of Computer Science 
Brandeis University 
Waltham, MA 0225~ 
617-736-2709 
jamesp ~by'andcis @csnet 
Abstract 
In this paper wc examine a subset of polyscmous elements, 
the logical structure of nominals, and argue that maw 
cases of polysemy have well-defined calculi, which interact 
with the grmnmar in predictable and determinate ways for 
disambiguation. These calculi constitute part of the lexicai 
organization of the grammar and contribute to the lexical 
semantics of a word. The lexieal system of the grammar 
is distinct from the conceptual representation associated 
with a lcxieal item, where polysemy is less constrained by 
grarmnar. We propose a structured' semantic representa- 
tion, the Lexical Conceptual Paradigm (LCP) which groups 
nouns into paradigmatic classes exhibiting like behavior. 
1. Introduction 
Much of the work on polysemy has tended to con- 
fuse the nature of word meaning by labeling many different 
types of ambiguity as the same phenomenon. In the ex- 
treme case, every possible lexieai semantic distinction is a 
case of polysemy and must have its own conceptual repre- 
sentation. In such a theory, various features are introduced 
to distinguish one sense from another~ but the meanings 
m'e all part of the same homogeneous conceptual space. 
In this paper, we argue that there are various types 
of polysemy, some more accessible to grammatical phenom- 
ena than others. We use this as a principled distinguishing 
characteristic of polysemous types. We distinguish two sys- 
tems that together comprise the meaning of lexieal items, 
the lexical system and the conceptual system. In particu- 
lar, if there are certai n grammatical characteri,~tics that are 
affected by a polysenmus relation between concepts, then 
we will say that this relation is overtly part of the lexical 
51B 
Artificial Intelligence Applications Group 
Digital Equipment Corporation 
Marlboro, MA. 01752 
617-490-8120 
anick@aiag, dee @decwrl. dec. corn 
and 
Computer Science Department 
Brandeis University 
system. 1 The mQor point to be argued here is that poly- 
semy is not a. single phenomenon operating over all lexical 
items with equM force and frequency. Rather, Lhere are 
subsystems in the lexicon which exhibit a restricted range 
of polysemous behavior. Each subsystem is determined 
by a representation called a Lexieal Conceptual Paradigm 
(LCP), which groups elements into classes with similar be- 
havior. We limit ourselves in this paper to cases of poly- 
semy involving nominals. 
We will proceed as follows. In Section Two we 
examine the different types of polysemous nominals. In 
particular, we look at a classification of relational nouns, 
paying particular attention to those exhibiting polysemous 
behavior. Then we examine the semantic interpretations 
possible for artifactual nominals such as book and record, 
and their associated polysemous behavior resulting from 
certain syntactic contexts. In Section Three we introduce 
a framework ibr representing this information in the lex- 
icon. We argue that certain semantic information asso- 
ciated with nomina~ is more priviliged and accessible to 
syntactic selection than other knowledge associated with 
the word. We term these privileged properties the hidden- 
event roles of a nominal, and they form part of the Lexical 
System of semantic analysis. All other information forms 
part of the more traditional notion of a conceptual space, 
what we term simply the Conceptual System. Finally, in 
Section Foul" we look at the computational implications of 
1 We will discuss neither lexical ambiguity for verb~ trot holnonymy. Our poinL 
here is to narrow iwl on the finer sel~antic distinctions within a smaller set of lexica\] 
items. For general discussion on iauues in lexical ambiguity, however, see /Boguraev 
1979/, \]Hit~t 1987/,/H~we~ 1977/, and/Wilks 1975\]'. See/StMlard 1987/for a related 
approach to polyaemy. 
2 As this group of nominals has been ~tudied extensively in the linguiBtic 
literature (cf. for example, /Anderson 1979/, /Puatejovsky 1984/, /Willimm~ 1985/, 
/Saflr 1987/, /Moor$gat 1985/), we will diecuss it here only briefly, 
suc~h a lexic:d organization. The distinction is an important 
one for computational reasons. First, it establishes clear 
criteria for partitioning the semantic information associ- 
ated with a word; this will bear directly on the decisions 
made by a parser in order to disambiguate lexical items 
during a parse. Secondly, it affects the planning strategies 
necessm3r for lexicM selection in the process of language 
generation. Finally, it relates to the question of how to 
enter multiple word senses for a lexical item. According to 
this view, those words containing logical ambiguities carry 
these sensc.,~ in the same entry. 
2. The litelational Structure of Nominals 
One reason to studythe semantics of relational nom- 
inals is that they exhibit polysemy in very well-defined and 
structured ways. For example, nominalizations such as ar- 
rival, destruction, and solution are ambiguons between the 
event aorniltal reading and the resultative reading, as illus- 
trated in (1) through (3)Y 
(1) a. We witnessed the city's destruction. Event 
b. The destruction was widespread throughout the 
city. Resnltative 
(2) a. Mary's arriva ! is expected to be at 3:00 pm. 
Event 
b. Mary's arrival means that she gets the couch. 
Res~tltative 
(3) a. John's solution to the problem took 20 minutes. 
Event 
b. This solution is too difficult to understand. Re- 
sultative 
It seems natural to ascribe the relational senses to nomi- 
nalizations ;~s being some projection of their related verbal 
argument si;ructure. There axe many nominals, however, 
which are oot nominalizations, yet seem to refer to rela- 
tions in their meaning. Classic examples include nouns 
such as picture and story, a The difference is that their 
relational structures are implicit, while the relation in a 
nomhtalization is explicit. 
Let us now turn to the class of dependent nomi- 
nals. If the denotation of one sense of a lexical item is 
"depende:ct" on another, then that dependency is part of 
the semantic representation of that lexical item. A famil- 
iar example is father or mother, where the relational nature 
'must be p~xt of the semmatics of the noun. 4 We term these 
Primitive relational nominals. Perhaps not as familiar are 
nouns such as blemish, wart, scratch, cut, etc. In each of 
thesd cases, the object is evaluated with respect to another 
object, and in fact it is hard to imagine the dependent ob- 
ject existing in isolation. For this reason, these will be 
called Primitive Figure-Ground nominals. The object it- 
self is the figure to a necessary ground object. 
A related class of lexical items includes nouns such 
as window and door. These are not simply Primitive Figure- 
Ground nominals, for notice that there is a hidden argu- 
ment present that relational nouns such as wart do not 
have. 
(4) a. Plastic windows are on sale at Lechmere. 
b. The cat crawled through the window. 
(5) a. John painted the door blue. 
b. John walked through the door quickly. 
As pointed out in \]Lakoff 1987/, as well as/Talmy 1975/, 
the nouns here are polysemous, since the window referred 
to can be the opening as well as the object. Thus, the hidk. 
den argument mentioned is the open space that is left as 
a result of the absence of the window. We term this argu- 
ment the inverted figure, and the noun class itself Double 
Figure-Ground Nominals. 5 In Section Three, it will bc 
shown that the polysemy illustrated in (4) and (5) is very 
different from word senses not making reference to either 
the figure or inverted figure. 
The next class of relational nominals are those which 
enter into a specified relation; namely, the class of artifacts. 
An artifact is, by definition, an object associated with a 
particular activity or event; for example, cigarettes are for 
smoking, books are for reading, etc. Because of these as- 
sociated activities, this class of nouns exhibit polysemous 
behavior. 
(6) a. This record weighs an ounce. 
b. This record lasts 35 minutes. 
(7) a. The book has a red cover. 
b. This book will talCe a week (to read). 
(8) a. These cigarettes are longer thaal the normal size. 
b. His cigarette is bothering me. 
3 See /Pustejovsky 1984/, /Safir 1987/, and /WilliamB 1985/ for further dis- 
cu6slon. 
4 /Partee and Kamp 1986/discusses the semantics of such nominals. 
5 Another type of dependent nomilml is that seen with 'lock ~ and ~key'. These 
objects do exi.t independent of the ground object it is associated with through its 
function, but still allow selection for this dependency; for example, "key to the lock" 
and "10ck for the door". 
519 
The polysemy in (6) arises becmme of the possible 
reference to the event of playing the record as well as the 
record by itself. Similarly, in (7) the book itself or the event 
of reading the book can be referred to by the nominal. 
Finally, the difference in (8) points to the cigarette as an 
object with attributes versus an object in the context of 
being smoked. ~ 
In this section we have presented five types of re- 
lational nominals (nominalizations, primitive relationals, 
primitive figure-grounds, double figure-grounds, and arti- 
factual nominals), showing how they exhibit subtle but 
productive polysemous behavior. In the next section, we 
outline our approach to polysemy and preserit an explana- 
tion of these lexieal mnbiguities in terms of a richer semam 
tic representation. 
3. The Theory of Lexical Organization 
Unlike many previous approaches to word meaning, 
we distinguish the logical lexieal semantics of a word from 
its deeper, conceptual denotation. .t We term these the 
L-system and U-system, respectively. 
The L-system is the particuhu- organization that 
the lexicon assumes independently of the conceptual 
system. Only semantic information tlmt is somehow 
reflected in the syntax is represented here. 
The C-system is the organization of the concepts 
themselves and not the language. This is the struc- 
ture of the me~mings, and as such, would be rep- 
resented as a semantic network or radial category 
structures (Cf./Hayes 1977/,/Lakoff 1987/). 
l)br each of the nominal types described in the pre- 
vious sections, we give lexical representations which allow 
us to capture their polysemous behavior. Consider first 
the Double Figure-Ground Nominals in (4) and (5). The 
lexical semantics of such nouns as window and door must 
refer t,o the three arguments mentioned by the implicit re- 
lation, the figure, ground, and inverted-figure. Assuming a 
first-order, partial-decomposition model of lexical seman- 
tics, as given in /Pustejovsky 1987/, the translations for 
this class would be as follows: 
6 It should be pointed out here that these are not case~ of metonymy. In 
metonymy, the ability of one referring expression to stand in for another object is very 
uneonstrained~ and fails to conform to any strict 9ondiLiolm on flllblltitutioIl t unlike t|le 
ea~ea mentioned above, 
520 
(10) doo,. =. ~,~y\[~,.ti/~(~) ^ Vig.,.4~) ^ G,'o~-,4c) A 
rig~,.~,(y) ^ ~(~, go.tl,...(~, y))\]. 
where the three.-place relation is explicitly represented as 
a conjunction of functions over the argmnents~ c is a co,_~- 
stant~ Figurei is the inverted-figure, and ~, is a metalogical 
operator indicating the purpose of the object being defined 
(el. /Pustejovsky in '1,reparation~). The important thixa5 
to note about this representation is that it predicates two 
distinct types of information over two different but rein,ted 
arguments, x and y; inanely, that a door~ for example 4 is 
defined in terms of both the concrete object whi(:h is arti- 
Net (the figure), and the space in the absence of this object 
(the inverted-figure). Thus, differeiit matrix predications 
will foreground different subexpressions in the semantics 
ibr the norm. For example, in (Sa), the artifact reading is 
selected, picking out the figure, while in (5b) the inverted- 
figure is selected. 
(5) a. John painted the door blue. 
b. John walked through the doo'r quickly. 
We argue that these are the only two types of polysemy tbr 
these nominals that are lexieally motivated, and that o~her 
apparent eases of polysemy are .simply inferences associ- 
ated with the conceptual representation of the object. This 
can be seen in sentence such as (1l), taken from /Lakoff 
1987/. 
(11) a. The window is rotting. (Reference to the wood) 
b. The kids broke the window. (Ii.eibrenee to the 
pane) 
These are not reMly polysemous in our sense, since the ar- 
gument structure of the verb selecting the nominal does 
not specify how the noun is to be interpreted. That is, 
although both readings select for the ~u'tifact interpreta- 
tlon, only inferences in the conceptual system, and not the 
lexieal semantics, tells us that rotting is predic~ted of tlhe 
wooden part of the window, s 
Another importmlt aspect to the representations in 
(9) m~d (10) is the expression introduced by the operator 
~r. '\['his is an exmnple of a hidden-event 'role denoted by a 
nominal expressiom I a the case of door, the hiddcn-evex_,t 
............. -8--~;\/ie~tructure of ~he conceptual ~ya~em will permit such infvx'cnce~ in a nut- 
m,al wetv, depending on the eyatem'~ colamon~len~e model of physics ~xnd agaric0, gee, 
/Hobbs et al 1986/for aL mlggeative ~pl)rot~eh to such i~auea. 
is go.through(w, y), a pointer to a particular lexical item, its 
argumelg; structure, and its selectional properties. 
The richer representations in (9) and (10) now pro- 
rides us ~vith a mechanism for capturing some interesting 
and subtle lexical distinctions in the artifact nominal class. 
For nouns such as record, book, and bulb, we assign the fol- 
lowing semantics: 9 
(12) record .=~ ~z3e\[arti/aet(x) ^ 7r(x, play(e,z))\]. 1° 
(13) book=c. ,\x3e\[artifact(x) A r(x, read(e,w, x))\]. 
(14) bulb ~ ~x3e\[arti/aet(z) A ~r(x,illuminate(e,z))\]. 
By explicitly referencing the event that the object is asso- 
ciated wiLh we can solve several puzzles. First, notice that 
when the event readings of record and book are selected, 
they differ in their aspectual interpretations. 
(15) a. This record lasts 35 mimttes. 
b. This book takes a week (to read). 
c. *This record takes 35 minntes. 
d. ?This book lasts a week. 
This is certainly surprising if no reference is made to the 
type of egent referenced by the object. Within the calcu- 
lus of aspect outlined in/Pustejovsky 1987/, play(z) and 
read(x,y) fall into different event-types, activity and ac.- 
complishment, respectively, and license different temporal 
predicates. So, it is not surprising that lexical semantic in- 
formation is accessible to such processes in the grammar. 
Another interesting application of the notion of hid- 
den event (or h-event) comes from evaluative predicates. 11 
\])'or example, consider the differences between the (a) and 
(b) NPs below: 
(16) a. a vinyl record: ~P3x3e\[arti/aet(x) A vinyl(x) A 
~(~,playCe,~)) ^ P(z)\]. 
b. ~ long record )tP3x3e\[arti/aet(x) A r~(x,play(e,x)) A 
Ion.~(~) ^ P(x)\]. 
(17) a. an opaque bulb ~P3x3e\[arti/act(z) A opaque(z) h 
~r(x, illuminate(e, z)) A P(z)\]. 
b. ~ bright bulb ~x3e\[arti/aet(~) ^ ~r(x, illuminate(e, x)) /~ 
b,.igh~(e) ^ p(x)\]. 
In (16b), long is a property that only the playing of the 
record has, while in (17b), bright is a property that only 
the state of illumination for bulbs has (el. /Jackendoff 
1983/for multiple senses of lexical items). By adopting a 
semantics that makes reference to events, just as with nom- 
inalization:b 12 we can begin to understand how to analyze 
evaluative predicates. Nmnely, in the cases above, we can 
distinguish the types of attribution as predication over an 
individual variable, the artifact ((16a) and (17a)), or over 
an event variable, the hidden event ((165) and (175)). 
The structures given in (12)-(14) are all examples 
of artifactual objects. The general abstraction for these 
individuals is the concept of an object made for a particular 
use: 
,\x3e\[arlifaet(x) A lr(x,~\[e,x\])\] 
where c~ is some predicate. Such a structure we will term a 
Lexical Conceptual Paradigm (LCP). We view these nom- 
inals as exhibiting paradigmatic behavior (much like the 
inflectional endings for verb classes) for the following rea- 
son: a paradigm acts as both an abstraction, in that it 
defines classes, as well as a structured object, with a clus- 
tering of different behaviors. When an object is assigned 
to a particular paradigm, it assumes the set of behaviors 
characteristic of that entire class. So it is with such ar- 
tifactual nominals. There are many such subsystems in 
the lexicon, each with their own internM consistency rep- 
resented by unique paradigms. This idc~ is explored in 
detail in/Pustejovsky and Anick 1988/. 
4. Computational Implications of Lexical 
Organization 
In this section we discuss the relevance and implica- 
tions of the above analysis of nominal semantics for com- 
putationM purposes. We will first look at the effect that 
richer semantic representations have on lexical selection in 
the process of language generation. Secondly, we outline 
how the problem of lexical disambiguation is facilitated by 
the use of such lexical entries in the process of analysis. 
Imagine a linguistic generation system in the ser- 
vice of a fairly rich semantic knowledge base and planner. 
Lexieal selection can be defined as the mapping from such 
9 We follow /Davidson, 1980\] and the extensions in /Parsons, 1985/for our 
representation of an event variable for the verb. See/Pustejovsky 1987/for discussion. 
10 Notice that the sense of 'play' here is the ergative reading and not the agen- 
tire. 
11 /ParteeandKamp, 1986/ provide a raodel-theoretic interpretation of evdu. 
ative a4jeetives, making use of pL'ototypc theory. Our analysis, howeverp follows mare 
closely that position taken in /tligginbotham 1986/. 
12 For a discussion of the event/resultative distinction in naminalizations, see 
/Pustejovsky 19~/and /Pustejovsky 1997/. 
521 
a knowledge base to a linguistic generation system. Appro- 
priate word choice is, of course, a function of numerous fac- 
tors and considerations (/Ward 1988/, /Pustejovsky and 
Nirenburg 1987/,/Danlos 1987/), but, in any case, the se- 
lection process makes sense only if the incoming semantic 
representation provides for there to be a distinction that 
is later reflected as a lexical distinction. 
Consider now an implicitly relational nominal such 
as cigarette. The lexical representation for such a noun is 
given as (18). 
(18) ~iga~eue ~ ~Je\[~tffa~t(z) A ~(z, smok~(e, w, ~))\]. 
One of the properties of such nominals is that they may 
denote the situation or event that the object is embedded 
within (in this way, it seems to stand in a metonymic rela- 
tion to the event). Thus, if given an underlying semantic 
form such as that in (19), the generator could produce, 
anmng others, the linguistic forms shown in (20). 
(19) 3e13e23z\[smoke(el,x) A cigarette(z) A bother(e2,el,john)\] 
(20) a. The cigarette's smoking bothered John. 
b. The smoking of the cigarette bothered John. 
c. The cigarette bothered John. 
The interesting thing to notice here is that (20c) 
makes use of the nominal without explicit mention of the 
predicate smoke. In general, such paraphrase classes as 
that in (20) involve the hidden event associated with an 
object. Furthermore, such classes exist mainly for nomi- 
nals that are artifacts and have such lexical semantic rep- 
resentations. A similar decision procedure is at work in 
the examples in (21), this time in matrix object position. 
(21) a. John enjoyed the cigarette. (h-event = smoke) 
b. Mary enjoyed the book. (h-event = read) 
c. Bill hated the movie. (h-event = watch) 
For each nominal appearing as head in the object position 
in these examples, there is an associated hidden event that 
allows for the appropriate default presupposition for just 
what activity was enjoyed for each object. Thus, it is clear 
how the representation here provides for choice points in 
the process of lexical selection. 
As an indication of how lexical disambiguation can 
be facilitated with the richer nominal representations pre- 
sented here, consider the multiple uses of a polysemous 
noun such as lisp, illustrated in (22). 
522 
(22) a. John is running lisp. 
b. Mary has entered lisp. 
c. John knows lisp well. 
d. Mary is writing lisp. 
Now, it might be argued that these senses are all slight 
variants of one central sense for the nominal, perhaps that 
of lisp as a language. Yet what we know about lisp that 
makes it different from another language, say FORTRAN, 
is that it is an environment as well as a language. We 
can think of the preference rules in/Boguraev's 1979/lex- 
ical disambiguation system as being elements of strnctured 
packets, where the above senses are all logically related in 
one lexical representation of lisp; that is, its Lexical Con. 
ceptual Paradigm. These word senses compete as a group 
with other unrelated meanings. Then, finer discrimina- 
tions are aclfieved by reference to the internal structure of 
the LCP for that word. We explore such a technique in 
/Pustejovsky and Anick 1988/and compare our approach 
to/Hirst 1987/and other lexical disambiguation systems. 
5. Conclusion and Future Research 
In this paper we have argued against the treatment 
of polysemy as a relation between meanings in a homo- 
geneous concept space. Rather, we argue, that tbr nom- 
inals at least, there are well-defined subsystems of logi- 
cally structured ambiguities arising from determinate and 
finite sources. We suggest that the incorporation of hidden 
events in the lexical entry of a nominal provides a handle 
for linking to other grammatical systems such as aspect 
mad thematic roles, as well as exposing the behavior of 
other relationships, such as figure-ground, which also have 
grammatical reflexes. The utility of such a representation 
in NLU systems is illustrated by the applications to such 
tasks as lexical selection in naturM language generation 
as well as to word sense disambiguation tasks and appli- 
cations in machine translation. For example~ we are ex- 
amining the subtle range of grammaticalizations involved 
in German nominalization patterns using an event-based 
semantics; i.e. assessing the appropriateness of one nom- 
inal expression over another. Finally, it is interesting to 
speculate on what the relationship between LCP-theory 
and acquisition and metaphor is. We are exploring this in 
current extensions to this work. 
References

Boguraev, Braxmmr Konstantinov, "Automatic Resolution 
of Linguistic Ambiguities", Univemity of Cambridge PhD, 
1979. 

Danlos, Lawrence, The Linguistic Basis of Text Genera- 
tion, CaIabridge Univeristy Press, 1987. 

Davidsoi~, Donald Essays on Actions and Events, Claren- 
don Press, Oxford, 1980. 

Fillmore~ Charles, "Construction Grammar," presented at 
Symposium on Lexical Semantics, Stanford University, Au- 
gust 3, 1987. 

Finin, Timothy Wilking, "The Semantic Interpretation of 
Nomin~l Compounds", University of Illinois PhD, 1980. 

Hayes, :\['hilip, "On Semantic Nets, Frmnes, and Associa- 
tions," Proceedings of 5th IJCAI~ Cambridge, MA, 1977. 

Higginbotham, James, "On Semantics," Linguistic Inquiry, 
1986. 

Hirst, Graeme, Semantic Interpretation and the Resolution 
of Ambiguity, Cambridge University Press, Cambridge, 1987. 

Hobbs, Jerry, William CrofL Todd Davies, Douglas Ed- 
wards, ~,,nd Kenneth Laws, "Commonsensc Metaphysics 
and Lexical Semantics", in Proceedings of ~$th Conference 
o~ the A.~sociation for Computational Linguistics, Columbia 
U., New York, 1986. 

Jackendoff, Ray, Semantics and Cognition, MIT Press, Cam 
bridge, MA. 1983. 

Lakoff, George, Women, Firc~ Dangerous Things, Univer- 
sity of Chicago Press, Chicago, 1987. 

.t~rsons, Terence, "Underlying Events in the Logical Anal- 
ysis of English", in E. LePore and B. McLaughlin (Eds.), 
Actions and Event.a: Perspectives on the Philosphy o~ Don- 
ald Davidson, Basil Blackwell, Oxford. 

ParSee, Bsxbm'a, and Hans Kmnp, "Prototype Theory/' 
Ms°, UMASS, Amherst, MA. i986. 

Pustejov.~ky, James, Studies in Generalized Binding, Ph.D. 
Univcrsity of Massachusetts, GLSA, 1984. 

Pustejov.,~ky, James, "An Event Structure for Lexical Se- 
mantics": submitted to Computational Linguistics. 

Pustejovsky, James, "Lexical Organization and Concep- 
tual Structure" in preparation. 

Pustejovsky, Jmnes and Peter Anick, "The Semantic In- 
terpretation of Nominals", Brandeis University Computer 
Science Technical Report, 1988. 

Pustejovsky, James and Nirenburg, Sergei, "Lexical Selec- 
tion in the Process of Language Generation", in Proceed- 
ings ~5th Association for Oomputational Linguistics, 1987. 

Stallard, David, "The Logical Analysis of Lexical Ambigu- 
ity", in Proceedings of ~5th Meeting of the Association for 
Computational Linguistics, Stanford, 1987. 

Talmy, Leon0xd, "Lexicalization Patterns'. Semantic Struc- 
ture in Lexical Forms", in Language Typology and Syntac- 
tic Description, T. Shopen (ed.). Cambridge University 
Press, 1985. 

Ward, Nigel, "Issues ill Word Choice", in Proceedings of 
COLING-88, Budapest, 1988. 

Wilks, Yorick, "Preference Semantics," in Formal Seman- 
tics of Natural Language, Keenan, E. (ed.), Cambridge 
University Press, 1975. 
