Interpretation of Nominal Compounds: Combining 
Domain-Independent and Domain-Specific Information 
C6cile Fabre 
IRISA 
Campus de Beaulieu 
35200 Rennes 
\[~r&IICe 
cfabre~irisa.ft" 
Abstract 
A domain independent model is pro- 
posed for the automated interpretation 
of nominal compounds in English. This 
model is meant to account for productive 
rules of interpretation which are inferred 
from the morpho-syntactic and seman- 
tic characteristics of the nominal con- 
stituents. In particular, we make exten- 
sive use of Pustejovsky's principles con- 
cerning the predicative information asso- 
ciated with nominals. We argue that it is 
necessary to draw a line between gener- 
alizable semantic principles and domain- 
specific semantic information. We ex- 
plain this distinction and we show how 
this model may be applied to the in- 
terpretation of compounds in real texts, 
provided that complementary semantic 
information are retrieved. 
1 Motivation 
Interpreting nominal compounds consists in re- 
trieving the predicative relation between the con- 
stituents. In many cases, no surface information is 
available to deduce the relation, and in particular 
no morphological evidence of a link between the 
constituents and the underlying predicate. This 
problem has been tackled in several types of NLP 
systems, mainly: 
- domain-dependent systems. Such systems are 
very efficient but are limited to the domain they 
are built for: interpretation rules are inferred 
from the observation of specific semantic patterns 
(Marsh, 1984) or from a fine-grained conceptual 
representation (Ter Stal, 1996). 
domain-independent systems (Finin, 1980; 
Mac Donald, 1982), built to account for any kind 
of interpretation patterns, including rules that 
are not inferred from the properties of the con- 
stituents (what Finin calls productive rulcs, in op- 
position to structural rules). Frequency and prob- 
ability scores are added to the rules. Such numeric 
weighting of general semantic rules is hardly de- 
fensible in the absence of any reference to a do- 
main. 
Consequently, the questions that we propose 
to answer are: how far Call we go in design- 
ing a model of interpretation rules which account 
for productive patterns of interpretation, indepen- 
dently of any domain? Couversely, what domaim 
specific information must be available to enrich 
this general model? The aim of our research is to 
define as precisely as possible the border line be- 
tween what can be regularly described with gen- 
eral linguistic mechanisms, and what has to do 
with subregular or irregular phenomena which de- 
pend on corpus characteristics. This is a crucial 
issue when dealing with compound semantics be- 
cause regular semantic patterns (involving rela- 
tional properties of nominMs) and extralinguistic 
data are mingled. 
We have designed a model 1 that accounts for 
structural rules (in Finin's terminology) of in- 
terpretation of N N compounds 2, i.e. domain- 
independent rnles that are deduced from the 
rnorpho-syntactic and semantic characteristics of 
the nominal constituents. The interest of this gen- 
eral model is to base the interpretation of com- 
pounds exclusively on general principles regard- 
ing the association between nouns and predicative 
information. Besides, this non-specialized model 
of interpretation allows us to draw a comparison 
with nominal sequences across languages, and es- 
>Phis project is supported by the CNE'P (contract 
CNET-INRIA n°951B030). Our model of interpreta- 
tion of nominal compounds will be used to enrich im 
formation retrieval in a system that is open-domaln. 
2In this work, we only focus on non-recursive 
terms. The same interpretation mechanisms can be 
extended to compounds with three constituents or 
more, but furthermore these compounds raise the 
problem of ambiguous bracketing (Resnik, 1993). 
364 
pecially with l)¥ench sequences of the form "N de 
N" and "N g N", in which tile l)repositional link 
is semantically weak (l,'abre and Sdbillot, 1994). 
We first describe this model, showing how com- 
l)ound interl)retation must rely on an accurate de- 
scription of the predicative prol)erties of n()minal 
constituents. We then suggest how this general 
model may be apl)lied to the interpre.tation of 
compounds in texts, provided that it is made more 
specilic wil, h domain-dependent or text-specific in- 
formation. 
2 Domain-independent lnodel 
In this section, we briefly explain how the inter- 
pretation is carried out when conlpoullds contain 
explicit predicative information. We then focus 
on the interpretation of compounds in which the 
constituents are root nonainals. 
Ill what follows, semantic features are adapted 
front the WordNet lexical database :~ which pro- 
vides a rich but non-specialized semantic taxon- 
omy. We use a small part of this hierar(:hy in order 
to define, a set of semantic features that lat)el non> 
inal constituents. Sen:laaltic labels are also used I,o 
express seleetional restrictkms on arguments. 
2.1 Compounds with a devea'lml 
constituent 
Compounds including a deverbal constituent that 
subcategorizes the other (:onstitueut have been 
precisely descril)ed, in particular within the, gen- 
erative franmwork (Selkirk, 1982; IAeber, 11983). 
These results have been integrated in our model. 
The predicati(:e relation between the con- 
stituents is given by the verbal root of the (lever- 
bal noun. We differentiate two types of deverbals: 
a deverba\] may refer to the accomplishment or 
the result of the process denoted by the verb (e.g. 
parsing) or it may saturate the role assigned to 
one of the a rgunmnts of the verb and thus refer 
to one of the actors of the process (mainly agent 
or instrument, e.g. parser). In the former ease 
(action deverbals), the deverbal inherits the en- 
tire ~rgument structure of tile w~'rb; in the latter 
(subject devcrbals), it inherits the structure mi- 
mls the agent saturated by the sutfix. When the 
deverbal noun occupies the head position of the 
compound, the non-head may saturate one of the 
roles of the argument structure of the deverbal, 
either the theme role, a.s in sentence parsing --y 
parsc(theme: sentence4), or a semantic role (ill 
aWordNet is a trademark of lh'inceton University. 
4'l'he semant;ic interpre|;ation is ,'epresented in a 
tormula that exhit)its both t, he underlying pr(:dicate 
and the roles thaC each constitucnl, plays in th(" m'gu- 
the sense of Selkirk (1982)), referring to a cir- 
cmnstance of tile action (location, time, means, 
etc.}: hand parsing ~ parse(means: hand). When 
the deverbal noun is the non-head, it cannot sat- 
urate an internal argument within tile compound 
(Lieber, 1983); in this case, the head may only 
fill a semantic or an external argument: parsing 
program --~ pacsc(instrument: program). 
This first series of compounding patterns has 
often be considered as the only type of compound 
which can be described in semantic terms (Selkirk, 
1982). Our own position is to argue that the same 
predicate-argument pattern IIlay be used to deal 
with other types of compounds, provided that we 
rely on a richer semantic representation of nomi- 
na.ls, when no morpho-syntactic clues are available 
to constrain the semantic interpretation. 
2.2 Root conlponnds 
NominM compounds illustrate the distributional 
properties of nouns in the absenee of any ex- 
plicit verbal ln'edicate. They attest an rattier- 
lying event structure associated to nominal con- 
stituents, which makes it possible to derive a pred- 
icative relation from the mere collocation of two 
simple nouns. The idea that noun meaning in- 
volves ewmt-based description has been particu- 
larly emphasized by J. Pustejovsky (1991). We 
propose to apl)ly a crucial component of his gener- 
ative lea:icon, tile qualia st'ructurc, to tile semantic 
interpre~tation of conlpOullds. 
The key idea tllnt underlies the qualia sl,'uctu,v 
is that nouns are implicitly related to predicative 
information, and that a noun selects tbr the tyl)e 
of predicate, that can govern it. The four typ- 
ical nominal relations that constitute the qualia 
struetmv are tile telic role, that refers to the pur- 
pose and function of the referent, the agentive role, 
that concerns the factors involved in its origins, 
the constitutive role, that captures the relation be- 
tween an object and its constituent parts, and the 
Jormal role, that distinguishes the ol)ject within a 
larger domain. 
We illustrate the use of this theoretical flame- 
work R)r the interpretation of nolnitm.l contpounds. 
Telic role. The notion of telic role is directly 
applicable to the treatment of COml)ounds. It; re- 
calls Finin's notion of role nominals (Finin, 1980). 
A role nominal is typically linked to a verbal pred- 
ica.te that denotes its purpose; it, fills one of the 
roles included in the argument structm:e of the 
verb. For example, the noun pipeline typically 
refl'.rs to the external argument of tile verb trans- 
menl; structure o\[ that predicate: NI N2 -+ V(role_i: 
IN2, role_j: N1). The head constituent is underlined. 
365 
port (cf. WordNet textual gloss: '% long pipe used 
to transport liquids or gases"). Unlike subject de- 
verbals, role nominals are not provided with an 
argument structure that may be syntactically sat- 
isfied. Nevertheless, the argument structure of the 
underlying verb provides a clue for the distribu- 
tional properties of the noun within compounds. 
The verb tTunsport requires a subject and an ob- 
ject argument; since the noun pipeline refers to its 
first argument, the position which is left empty 
(the theme) may be occupied by the first con- 
stituent of a compound of the form N pipeline, as 
in oil pipeline -+ transport(instrument: pipeline, 
theme: oil). 
Agentive role. The agentive role is also se- 
lected by the compounding mechanism: the non- 
head may refer to the origin of the head noun, 
as in pancreas ptyalin -+ produce(agent: pan- 
creas, theme: ptyalin), in compiler message -+ 
emit(agent: compiler, theme: message), or in bul- 
let wound ~ cause(agent: bullet, theme: wound). 
We see that this relation covers different kinds of 
predicates which are instances of a more general 
relation of creation. 
Constitutive role. The constitutive role in- 
cludes various kinds of semantic associations, such 
as part-whole relations (outrigger canoe) or sub- 
stance relations (stone house). 
Formal role. The formal role involves a re- 
lation of characterization which concerns differ- 
ent aspects of an object (its size, shape, color, 
etc.). The nouns that denote such information are 
mostly elements of the ATTRIBUTE class, which is 
defined in WordNet as "an abstraction belonging 
to or characteristic of an entity". Each member of 
this class may appear at the head position of com- 
pounds in which the non-head denotes the entity 
that is characterized: desk height --+ character- 
ize(attribute: he_ight, entity: desk). These nouns 
are uni-relationM nouns that can appear as the 
head of "N1 of N2" groups, where N2 is a syn- 
tactic argument of N1 (e.g. height of the desk) 
(Isabelle, 1984). 
Consequently, Pustejovsky's notion of noun's 
qualia helps to characterize implicit predicative 
link in compounds. This semantic framework 
demonstrates that the association between nomi- 
nal constituents and underlying predicative rela- 
tion in root compounds is not arbitrary: it in- 
volves conceptual mechanisms that are triggered 
in other linguistic phenomena such as type coer- 
cion (Pustejovsky, 1991), anaphora (Fradin, 1984) 
or adjectival constructions (Bouillon and Viegas, 
1993). 
2.3 Implementation and results 
The implementation of these principles in our 
model is based on a conceptual framework in or- 
der to associate predicative information with nom- 
inal constituents. Two cases arise: when the link 
between a noun and a predicate is characteristic 
of a single noun, it is expressed in its lexical en- 
try. When it is shared by a whole class of nouns, 
it is seen as a characteristic feature of that class 
which accounts for a relational property that any 
member of the class inherits. For example, the 
telic role of the word pipeline, which involves the 
verb transport, cannot be generalized to a whole 
class of nouns. On the contrary, the predicate 
CONTAIN is a characteristic feature of the class 
CONTAINER. Consequently, several predicates and 
several roles are potentially associated with nom- 
inal constituents, either as instances of different 
attributes, or as a consequence of this inheritance 
mechanism. 
We have tested our model on a list of 100 com- 
pounds randomly picked up from a list of N N 
sequences in isolation 5. Our program generates 
any interpretation that can be calculated on ac- 
count of the mechanisms that we have described. 
Firstly, the list of predicates that are associated to 
the head constituent 6 is retrieved. Secondly, only 
the predicates that can provide a role to the other 
constituent are retained. 
It is difficult to assess the correction of the an- 
swers that are produced, since we are dealing with 
compounds in isolation. Other answers are some- 
times conceivable, if we apply less regular princi- 
ples of semantic associations (Downing 1977), so 
that we cannot compare our results with a closed 
set of correct answers. Moreover, we cannot set a 
clear-cut border line between probable and hardly 
conceivable interpretations. Having said this, we 
can estimate our results as follows: 71% of ~he 
compounds that we have examined receive accept- 
able answers. For example, our program generates 
two clearly acceptable solutions for the compound 
missile range: 
1) characterize(agent: range_7, theme: missile) 
2) shoot(locative: range_9, theme: missile) 
Contrary to Finin's and Mac Donald's models, 
5This list of 9000 binary nominals has been kindly 
put at our disposal by R. Sproat. The corpus is de- 
scribed in (Sproat, 1994). 
6 In most cases, the predicative information is asso- 
ciated with the head, except when the non-head is de- 
verbal, as in hunting lodge, or when the head refers to 
an under,pecified event structure, ,as in malaria pro- 
gram (fight) vs crop program (develop). Such com-- 
pounds illustrate the notion of co-compositionality 
(Pustejovsky 1991). 
366 
we are dealing with ambiguous constituents: nine 
meanings of the word range are listed, which cor- 
respond to the description given by WordNet for 
this noun. Only senses 7 ("scope", ATTmBUTE) 
and 9 ("a place for shooting projectiles", AaTE- 
FACT) are related to a predicative information 
that is compatible with the non-head, namely the 
formal role in the first case, and the relic role in 
the other. Some answers are more questionable: 
cardboard box -- 
1) constilutc(agent: cardboard, theme: box_/t, 
box_5, box_6, box_7) - objects made of cardboard 
(constitutive role) 
2) contain(locative: box_7, theme: cardboard) - 
box that contains cardboard (telic role) 
3) produce(agent: box_3, theme: cardboard) -- 
plant that produce cardboard (telic role) 
4) measure(agent: box_2, theme: cardboard)- a 
quantity of cardboard (formM role) 
Interpretations 2, 3 and 4 are surely mistaken 
in a standard context, if we refer to extralinguis- 
tic: knowledge (box_3 - a kind of shrub - does not 
prodnce cardlooard the way gum trees l)rodnce 
gum) or to lexicalization (the compound card- 
board box has only one usual meaning, namely 
constitute(agent: cardboard, theme: box_7, where 
box_7 refers to the container). Yet, each answer 
is conceivable because it corresponds to produc- 
tive semantic patterns and therefore to existing 
cognitive strategies. 
6% of the answers miss expected answers and 
23% give no answers at all. If we compare our 
results with those of Mac Donald (1982), we see 
that the part of silence is undoubtedly less im- 
portant in his system (no meaning is produced 
for 10 % of the compounds). Nevertheless, one 
crucial distinction must be emphasized: in Mac 
l)onMd's system, slots are defined in relation to 
nominals, and an interpretation is identified if one 
constituent can fill a slot of the other. These slots 
are supposed to represent any piece of real-world 
knowledge that is necessary to understand noun 
compomMs, but nothing precise is said about the 
information that needs to be stored. The solu- 
tion to improve this resnlt is unclear in such a 
system: missing interpretations correspond to ab- 
sent slots, but no indication is given regarding the 
slots that must be added. On the contrary, we 
have shown that a few general principles of pred- 
icative attachment to nominal constituents are in- 
volved in the interpretation of Compounds in our 
model; consequently, the analysis of incorrect an- 
swers allow us to determine in what cases domain- 
independent rnechanisms are unsulticient to per- 
form the interpretation and what kind of knowl- 
edge must be added to improve these results, ei- 
ther from domain-dependent or froln contextual 
information. One can classify the problems in two 
categories: 
Inappropriate selectional restrictions 
Only selectionM features can constrMn the in- 
terpretation when several predicates are possible, 
in order to distinguish between different roles (e.g. 
shoulder wound- the non-head affects a BODY 
PART VS bullet wound - the wound is caused by a 
WEAPON). Consequently, no interpretation is gen- 
erated when the semantics of the non-head does 
not match the constraints on the arguments of 
the predicate, and particnlarly in case of semantic 
shifts: stadium is a CONSTI~UCTION, but in sta- 
dium clo~sh, it is viewed as a LOCATION or 3.s a 
GI{OUP of people. This is a general issue in lex- 
ical semantics; yet, the problem is all the more 
difficult to handle in compounds as no syntactic 
clue (i.e. no prepositional link) is available to dis- 
tinguish between different (semantic or thematic) 
roles. It is also particnlarly problematic to solve 
ambignons role assignment when semantic roles 
are concerned (as in fear voters). 
Missing predicative llnk A general model 
cannot account for all possible compounding rela- 
tions. Not to mention contextual links (Downing, 
1977), some productive relations cannot be con- 
strained from the semantics of the constituents. 
Specific links such as ressemblance (carpet shark) 
or subclass relations (marathon tour) cannot be 
described with structural rules. Moreover, a pred- 
icative information may be missed when it entails 
fine extralinguistic knowledge (e.g. fl'uit fly: in- 
sect whose larvae feed on ft'nits). 
Generation of multiple interpretations and un- 
predicted patterns due to selectional violation 
or extralingnistic information are thus the two 
inherent limits of a domain-independent model 
of interpretation. Our aim is to give sugges- 
tions about the possibilities of refining this model 
when domMn-specific or contextual information 
are available. 
3 Domain-specific semantic 
inforlnation 
3.1 Detection of specific patterns 
Preferential patterns Statistical methods have 
been experimented by psycholinguists such as 
Pamela Downing (Downing, 1977) and Mary Ellen 
l{yder (Ryder, 1984): their purpose is to use sta- 
tistical knowledge to interpret new compounds. 
Ryder argues that a set of semantic rules is not 
sufficient to deal with the productivity of the 
compounding process, since the creation of new 
3 6'7 
compounds involves extralinguistie knowledge and 
cognitive strategies. According to her, "the pre- 
dictability is probabilistic", and she shows that 
the creation and interpretation of new compounds 
is based on knowledge about productive semantic 
patterns. For example, she lists highly frequent 
templates such as: 
N -t- PRODUCT : PRODUCT used on N (pet 
shampoo, laundry detergent) 
This pattern illustrates only one facet - the relic 
one - of the head noun (and is irrelevant for exam- 
ples such as egg shampoo or dishwasher detergent). 
This statistical result may differ considerably from 
one corpus to another. Consequently, fi'equency 
scores cannot be part of a domain-independent 
model. 
From our results, we see that two types of spe- 
cific information must be available to refine our 
domain-independent rules: firstly, we must spec- 
ify the relative frequence of each role to assess the 
best interpretation tbr a compound when several 
semantic relations apply. Secondly, we want to 
determine the semantic features that characterize 
the non-head for one given role; P.Resnik's aim is 
similar when he illustrates the use of selectional 
association in compounds (Resnik 1993), in order 
to find N N semantic patterns which help to per- 
form adequate bracketing of sequences with three 
constituents or more. Ite shows that it is diflicnlt 
to find clear-cut semantic groups in unrestricted 
texts. Yet, such techniques, that combine statis- 
tic measures and conceptual knowledge, are very 
promising to exhibit typical patterns of associa- 
tion in specific domains. 
Unpredieted patterns Exhibiting unpre- 
dieted patterns is a first step towards the determi- 
nation of specific interpretation schemes in a given 
domain. For example, let us consider a list of com- 
pounds matching the N pump pattern, such as: air 
pump, beer pump, breast pump, cattle pump, gear 
pump, piston pure,p, sand pump, stomach pump, 
drainage pump. In this list, we find compounds 
exhibiting: 
- the telic role of the noun: 
SUBSTANCE + pump --+ pump(instrmnent: 
pump, theme: SUnSTANCP 0 (sand, air) 
ACTION -t- pump ~ ACTION(instrument: pump 
(drainage) 
-. the constitutive role of the noun 
OBJECT + pump -+ constitute(theme: pump, 
agent: OBJECT) (gear, piston) 
These patterns are predicted and interpreted 
by our set of rules. Other types of associa- 
tions, too specific to be taken into account by our 
model, appear in the list: ANIMAL + pump (cat- 
tie pump) and ORGAN + pump (stomach pump, 
breast pump), in which the missing predicates are 
respectively feed - i.e. pump food Jor - and clean 
- i.e. pump the contents of. We see that the un- 
derlying relic relation is more complex, because 
it includes also an implicit argument (food, con- 
tents) of the predicate. These are typically the 
Specific patterns that cannot be taken into ac- 
count in a general model. Exhibiting semantic 
patterns in the texts is thus a way to autolnati- 
eally learn more specific patterns of associations in 
sublanguages. We are currently experimenting the 
way techniques of computer-aided acquisition for 
learning conceptual relations fi'om syntactic collo- 
cates (Velardi et al. 1991) can be applied to N N 
associations. 
3.2 Identification of the predicative link 
Our model associates a fixed verbal predicate with 
nouns or nominal classes to account for a given 
semantic facet. This predicate corresponds to the 
typical predicative information that occur ill the 
Wordnet textual gloss, when it is available. In 
fact, this predicate may vary fl'om one corpus to 
another, and we nmst take into account this vari- 
ation which corresponds to specific conceptual de- 
scriptions. Contextual information can contribute 
to identify the predicative relation by looking else- 
where in the text to see if the constituents of 
the compound are involved in another kind of lin- 
guistic construction, where their semantic relation 
would be explicit. Given a compound N1 N2, we 
may look for strings in which the couple (N1, N2) 
occurs in a different relation. In the following ex- 
amples, the context provides the missing verbal 
predicate: 
compiler warnings: (compiler,warning) = "it is 
reasonable for the compiler to emit a warning" 
In this example, which corresponds to the agen- 
tive role, we see that the two nouns are argmnents 
of the predicate that instantiates the underlying 
relation, which means that corpus-based methods 
can use a rich linguistic structure to identify the 
predicate. Pustejovsky et al. (1993) show how 
statistical techniques, such as mutual information 
measures can contribute to automatically acquire 
lexical information regarding the link between a 
noun and a predicate. Similar techniques are used 
by (Grefenstette and Teut~l 1.995) to determiue 
the support verb associated with deverbal nouns. 
Conclusion 
This paper describes a domain-independent model 
tbr the ,interpretation of nominal compounds; it 
shows how general knowledge and domain-specific 
368 
itiforinal;ion inay be combined for the interpreta- 
tion of nolnitlal colllpoulids. Otlr goal is to ac- 
count for l)roductive and actress-domain rules of 
interpretal,ion, l'\]xperimentation shows that the 
delinition of general rules, which inchide concep- 
tual description of the norninal constituents, im- 
plies the generation of multiple interpretations, es- 
pecially since we are dealing with arnbiguous nom- 
inal constituerits. 
We have \])reposed several ways of incoq)orat- 
ing specific 8elnantic inforination in our model, 
and we have suggested how corl)us observations 
can detect l)referential semantic relations and llll- 
predicted semantic patterns. Statistical observa.- 
tions can contribute to identify the most produc- 
tive compounding strategies for a given corl)us , 
and are especially very proniising a.s a way to (lea\[ 
with technical texts, in which the semantic vari- 
ety of cOinl)ounding relation is limited. '.l'his work 
is currently experinmnted in lPrenc\]l, where it el)- 
pears that tile saine eon(:el)tua\] franlcwork holds 
to account for the semantic role of prel)ositions (~ 
and de in binoininal sequences. 
References 
Pierrette l~ouillon and l,;velyl,e Viegas. 1993. 
Semantic I,exicons: the Cornerstone tbr l,exical 
Choice in Natural I,anguage (Jeneration, Prec. 
of tl~e sew'ulh I~d, ernational Workshop (4N,tu~'al 
Lang'ua.qc (;cnc'raLion. 
Pamela Downing. 1977. On the Creation and 
Use of English Compound Nouns. Language , 
53(4): 810-842. 
Cdcile Fabre and Pascalc Sdbillot. 1994. ln- 
terprdtation sdmantique des composds noniinaux 
anglais el; fran~a.is, l'roc, oJ the. Workshop on 
Compound Norms: Multilingual Aspects of Nomi- 
nal (7omposition, Genbve. 
'l'imothy Wilking 1,'inin. 1980. 'Phe Semantic 
Interpretation of Nominal Compounds, Prec. of 
the first co~@rcnce of AI. 
Bernard l,'radin. 1984. Anaphorisation et 
st&dotypes nominaux, Liugua, North-llolland, 
64: 3:x5-31i9. 
(h'egory (\]l eR:nstette and Siinone 'l'eut%l. 1995. 
Corl)us-based lnethod for auLoniatic idenLilication 
of ,nil)port verbs for noininalizations, l"roc, of 
EA CL, 1)ublin. 
Pierre lsabelle. 1984. Another Look at Nominal 
(\]otnpoullds, 13vet. of Colin q-.S4. 
I/,ochelle l,ieber. 71.983. Argument Linking and 
Compounds in I!',nglish, Linguistic lnqui'w, il d(2): 
2571-285. 
Da.vid 1t. Mac Donald. 1982, I/ndc,'s&mdi~lg 
(7o'mpounds Nouns, l'hD Thesis, (3arnegie Mellon 
/ Jniversity. 
Elaine Marsh. 11984. A Computational Analy- 
sis of Complex Noun Phrases in Navy Messages, 
P'~vc. of Colin q-84. 
James Pusteiovsky. 19911. The Generative I,ex- 
icon, Computational Linguistics, 1714): 408-441. 
James Pustejovsky, Peter Anick and Sabine 
Bergler. 11993. l,exical Semantic 'l%chniqnes 
for Corpus Analyses. Computational Lmg,i~tics, 
;I 9(2). 
Philil) Stuart Resnik. 1993. Selection alzd lr~- 
Jb'rmaliou: a Class-\]lascd Approach go l,emical l~c- 
l, lionships. PhD Thesis, University of Pennsylva- 
nia. 
Mary Ellen Ryder. 1994. O~de,wd Chaos: the 
interpretatiol~ of Euglish Noun-Noun Compotmds. 
University of California Press. 
l~',lisabeth Selkirk. 1982. 71~e S~\]ntaa: of Wo'rd,% 
M1T I'ress. 
Richard Sproat. 1!)94. l,;nglish Noun-Phrase 
Accent Prediction for Text-to-Speech, (\]ompulc.," 
,5'peech (rod Language, 8: 79-9/1. 
Wilco Ter Stal. 1996. Automated hd, e,cp.cctal, ion 
of Nominal Co'mpounds in a Technical Domain, 
Phi) Thesis. University of Twente, the Nether- 
lands. 
Pa.ola Velardi, Michela Fasolo and M aria 'l'cresa 
l'azienza. 1991. llow to l'hieode Semantic Knowl- 
edge: A Method \['or Meaning l{el)resentation and 
(Joml)uter-Aided Acquisition, Computatioual Li~t- 
9uisl, ics, 11712): 153-1170. 
369 
