GENERATION OF NOUN COMPOUNDS IN HEBREW: CAN 
SYNTACTIC KNOWLEDGE BE FULLY ENCAPSULATED? 
Yael Dahan Netzer and Michael Elhadad 
Ben Gurion University 
• Department of Mathematics and Computer Science, Beer Sheva, 84105, Israel 
(yaeln \[ elhadad) @cs. bgu. ac. il 
Abstract 
Hebrew •includes a very productive noun-compounding construction called smixut. Because 
smixut is marked morphologically and is restricted by many syntactic constraints, it has been 
the focus of many descriptive studies in Hebrew grammar. 
We present the treatment of smixut in HUGG, a FUF-based syntactic realization system 
capable of producing complex noun phrases in Hebrew. We contrast the treatment of smixut 
with noun-compounding in English and illustrate the potential for paraphrasing it introduces. 
We Specifically address the issue of determining when a smixut construction can be gener- 
ated as opposed to other semantically equivalent constructs. We investigate several competing 
hypotheses - smixut is lexically, semantically and/or pragmatically determined. For each hy: 
pothesis, we explain why the decision to produce a smixut construction cannot be reduced to a 
computation Over features produced by an outside module that Would not need to know about 
the smixut phenomenon. 
We conclude that smixut provides yet another theoretical example where the interface that 
a syntactic realization component presents to the other components of a generation architecture 
cannot be made as isolated as we would hope. While the syntactic constraints on smixut are 
encapsulated within HUGG, the input Specification language to HUGG must contain a feature 
that specifies that smixut is requested if possible. 
• However, because smixut accounts for close to half the cases of NP modifiers observed 
on a corpus of complex NPs, and it •appears to be the unmarked realization form for some 
frequent semantic relations, we empirically evaluate a default setting strategy for the feature 
use-smixut based on a simple semantic Classification of the relations head-modifier in the NP. 
This study provides a Solid ground for the definition of a small set of predicates in the input 
specification language to HUGG, that has applications beyond the selection of smixut -- for 
the determination of the order of modifiers in the NP and the use of stacking vs. conjunction 
--and for the definition of a bilingual input specification language. 
1 Introduction 
Over the past three years,• we have started developing HUGG, a syntactic realization component for 
Hebrew. One of our objectives is to investigate constraints on the design of the input specification 
language to a syntactic realization component through a contrastive analysis of the requirements 
of English and Hebrew. By design, we are attempting to keep the input to HUGG as similar 
as possible to the one we defined in the SURGE syntactic realization for English \[7\]. A detailed 
168 
I 
m 
U 
m 
I 
! 
! 
l 
l 
I 
! 
analysis of syntactic constructs specific to Hebrew becomes, therefore, critical to evaluate to which 
extent the input specification language can abstract away from knowledge of the syntax. 
We investigate in this paper one such construct: the Hebrew noun-compounding form known 
as srnixut. Because smixut is morphologically marked and remarkably productive in Hebrew, there 
exists a vast tradition of work in descriptive grammar of Hebrew providing functional analysis 
of the phenomenon \[11\] \[10\] \[13\]. This previous work has served as a fertile ground for our own 
generation-specific purposes. 
The specific issue we discuss in this paper is: what information in the input specification to 
the syntactic realization component can license the selection of a smixut construct. The classical 
objectives of modularity and knowledge encapsulation indicate that this decision should be a private 
decision of the syntactic realization component. Because there are so many syntactic constraints 
on the use of smixut, the objective of encapsulation is made even more desirable. 
After a thorough analysis of the different functions of the smixut construct and the constraints 
over its use, our conclusion, however, is that this reductionist strategy fails: we cannot explain the 
selection of a smixut construct without considering simultaneously lexical, semantic and pragmatic 
factors. 
Theoretically, in order to allow the syntactic realization component to select a smixut construct 
adequately, we are, therefore, left with two options: (1) either provide full, detailed access from the 
syntactic realization component to the complex semantic and pragmatic features that can impact on 
the decision; or else, (2) allow the other components to request the use of a smixut construct when 
they deem it adequate. In either case, modularity and encapsulation suffer. This analysis informs 
us in our design of a bilingual realization component: if a feature like use-smixut is required in 
the input to the syntactic component, this level of abstraction cannot be appropriate as a bilingual 
construction. It also informs us in the general ongoing debate over the design of reusable syntactic 
components and their place in the architecture of generators. 
From a more pragmatic perspective, however, we also provide a set of simple defaults for the 
generation of smixut based on a simple semantic classification of the relations head-modifier. We 
evaluate the validity of this classification by constructing input specifications for a corpus of more 
than 800 comp!ex, noun phrases and regenerating from them. The validation process includes two 
aspects: (1) we test that human coders agree on the semantic relations they use to label complex 
NPs; and (2) we verify that the generator's decision to produce a smixut construction corresponds 
to that observed in the corpus. Preliminary results are provided in Section 4.3. They encourage us 
to view in the set of semantic relations we propose a useful basis for the design of an interlingual 
input specification language. 
In the rest of the paper, we first briefly review the main approaches to the treatment of noun- 
compounds in English and in Hebrew. In Section 3 we provide descriptive data on the use of smixut 
in Hebrew. We then describe in Section 4 a first approach to the generation of smixut based on a 
simple semantic classification similar to that found in \[12\]. In Section 5, we identify the limitations 
of such an approach, illustrating that an explanation based on recoverable semantic relations cannot 
provide sufficient nor necessary conditions for the generation of smixut. However, the preliminary 
empirical evaluation we present in Section 4.3 demonstrates that the semantic relation approach 
provides a useful default that works "most of the time." 
169 
2 PreviOus Work 
2.1 Noun compounds in English 
Noun compounds in English axe partly "frozen" lexical constructions (e.g., computer science) and 
partly compositional constructions (e.g., computer equipment, farm equipment, city equipment...). 
The problematic aspect of this construction is that it seems to be very productive in English, but 
yet severely constrained (e.g., * science equipment). Compound constructions are also regularly 
ambiguous. 
The various approaches developed to explain the construction of noun compounds and their 
interpretation can be classified in three groups: semantic, pragmatic, and statistical/lexical. 
• Semantic theories explain the production of a noun compound N1 N2 as a derivation from a 
semantic relation N1 R NP where the relation • R is elided. The theory of recoverably deletable 
predicates (RDPs) Of \[!2\] proposes that only a small set of relations (canse, have, make...) can 
participate in this process. Because these relations were too general and sometimes vague, and 
because one can observe many cases of compounds that do not correspond to any of the proposed 
RDPs, others have proposed to define more precise domain specific models to explain the deletion 
of certain relations. 
Recognizing the importance of contextual factors, pragmatic theories predict the use of noun 
compounding when relations like naming or contrast play a role \[6\]. For example, when referring 
to two persons wearing a jacket and a Coat respectively, one can use compounds like the jacket 
man and the coat man even though, in neutral contexts, it would be difficult to interpret the same 
compound (i.e., the wear relation is not deletable). 
In \[5\], the explanation for compounding is provided in the form of lexical/syntactic knowledge. 
Generative devices inspired by \[14\] are found in the lexicon. In addition, statistical knowledge 
predicts which derivations are the most likely. 
From a generation perspective, the problem is less acute than for interpretation: we must decide 
whether to construct a compound as opposed to recover the missing relation between the head and 
the modifier. The problem has, therefore,• not received heavy attention for English generation. In 
the past, we used Levi's model in generation \[8\], but as part of the lexical chooser, and we did not 
include it within th e syntactic realization component. 
In Hebrew, however, the smixut Construction is extremely productive (in our corpus, smixut 
modification accounts for 40% of all modifiers, more than any other type of syntactic modification • 
in NPs). We, therefore, had to address the issue of when to generate smixut as a priority in the 
development of the NP grammar for HUGG. 
2.2 Noun compounds in Hebrew 
The structure of noun compounds in Hebrew - smixut, is marked and, therefore, it has been the 
focus of Hebrew language studies. The head (called nismax) is marked morphologically and it does 
not carry a mark of definiteness even when it is semantically definite. \[! 0\] \[2\] provide detailed 
studies of the syntactic constraints on the use of smixut. We provide an overview of the main 
constraints in Section 3. 
Although smixut is traditionally treated as a possessive construction, it can express many other 
relations between head and modifier. Levi \[11\] has extended her treatment of the noun-noun relation 
in English \[12\] and proposed that the same semantic relations can all be expressed by the Hebrew 
smixut construction. \[2\] and \[10\] (Chapter 6) also )rovide similar semantic classifications of the 
170 
elided relation in a smixut. We build on these studies in our implementation, but also investigate 
how a semantic account can be integrated with pragmatic and lexical constraints, 
3 Noun Compounds in Hebrew: Constraints 
We briefly present in this section the basic syntactic constraints over the use of smixut in Hebrew. 
The notion of "smixut" covers three main constructions: \[2\] \[10\]\[p24\]: 
compound separate construct double-genitive 
cadur ha-tinok ha-cadur Sel ha-tinok cadur-o Sel ha-tinok 
ball the-baby the ball of the baby ball-his of the-baby 
Smixut is identified by two main tests: first, when plural is used, only the head is marked 
morphologically with a special inflection: yeled (child - singular) vs. yelad-im (children - plural 
non-smixut) vs. yald~ei (children - plural head of smixut marking). 
Second, when definite is used, only the modifier is marked even though the head is understood as 
definite: aron 'mitbax (cabinet kitchen) (a kitchen cabinet) vs. aron ha-mitbax (cabinet the-kitchen) 
(the kitchen cabinet). 
Nominalizations are also built using a smixut construction with a gerund or a denominal as 
head: 
Bo ha-role Bo 'o Sel ha-role The arrival of the doctor: 
Arrive the-doctor Arrive=his of the-doctor 
We categorize the constraints on the use of smixut in four categories: syntactic, lexical, semantic 
and pragmatic. 
3.1 Syntactic Constraints 
One of the main constraints on the use• of smixut is that a head can have only one noun modifier 
(called somex in Hebrew). When several modifiers are attached to a head, this constraint forces 
other relations to be realized in other syntactic constructions (post-modifier adjective, prepositional 
phrase or relative clause). For example, when referring to a suit made of leather, the default 
realization (unmarked) is the smixut beged wor (suit-leather). The alternative realization beged 
me-wor (suit from-leather) with a qualifier PP is also possible, but less frequent. However, when 
referring to a bathing-suit in leather 1 the default realization is beged-yam me-wor (suit-sea from- 
leather). Because the somex (noun modifier) position of the head beged is occupied by the yam 
(sea) modifier, the second modifier •(leather) is relegated to another (not nominal) position. 
The head of a smixut must be a noun or a conjunction of nouns and it cannot be a compound 
itself. This means that smixut only allows right branching 2. This is in contrast with English, 
which allows right or left branching constructions: (computer communication) system vs. computer 
(communication system). 
Pronouns and proper nouns cannot head a smixut, and any pronoun in the modifier position 
'has the objective case and is agglutinated to the noun: ben-o son-him (his son). 
There are several restrictions on the combination •of smixut with different determiner types. 
Noun phrases in Hebrew are polydefinite -- that is, definiteness is marked on several of the con- 
stituents in the phrase. Any adjectival modifier is marked with agglutinated definite markers, the 
same as the head noun. Quantifiers and determiners can be also marked. In smixut, only the 
1Very frequent on Israeli beaches. 
2That is, if Hebrew is written left to right. 
171 
mod-n is marked as definite. Therefore, compounded nouns are understood as having the same 
definiteness value. As a consequence, if definiteness of the head and modifier differs, smixut cannot 
be reed: 
head-N + mod-N definite 
(The son of the king) 
ben ha-melex 
son the-king 
ben-o Sei ha-rnelex 
son-his of the-king 
head-N + mod-N indefinite ben me/ex ben Seg raelez 
(a son of a king) son king son of king 
head-N indefinite, mod-N definite benSeI ha-rnelez 
(a son of the king) son of the king 
head-N definite, mod-N indefinite ben-o Sel melex 
(The son of a king) son-his of king 
3.2 Lexical Constraints 
Not every noun can head a smixut construction: words which are lexical-compounds (cadur-sal - 
ball-basket - basketball), words of foreign origin, cannot be in nismaz form (i. e. the special inflection 
of compound nouns), and therefore any modifiers must be realized in another syntactic construction. 
Several criteria exist to distinguish frozen fromproductive smixut compounds: frozen com- 
pounds behave like regular smixut with-respect to plural marking (special morphology inflection). 
But depending on the level of cohesiveness of the frozen compound, definite marking may differ: 
beyt-sefer -house-book - (a school) may give ha-beyt-sefer (the school) instead of the predicted beyt 
ha-sefer for a productive smixut. 
In addition, for frozen constructs, many additional constraints exist: the head cannot be mod- 
ified (* beyt sefer kri'a - house book reading), cannot change its number (* beyt sfar-ym -house 
books), cannot be taken apart ( * bayt Sel sefer - house of book) or be•conjoined to another somex 
(noun modifier) (beyt sefer ve beyt Hol-im - house book and house patient-s - a school and hospital 
but * beyt sefer ve-Hol-im - house book and patient-s). Detailed references from linguistic and 
sociolinguistic aspects are found in \[4\] and \[3\] respectively. 
• 3.3 Semantic Constraints 
Smixut is often understood as a genitive type of construct, expressing dominantly a possessive 
relation between the head and the modifier. Very often, however, the relation expressed is not one 
of possession. 
The semantic relation realized by the smixut has an influence on the possible paraphrases the 
smixut can receive \[2\]: some semantic relations (including possessive) can be realized in a double 
genitive construction while others can only be realized by a simple smixut. The semantic relation 
also determines which types of modifiers can be accepted in the smixut construction. In general, 
when a double genitive construction is not possible, then pronouns cannot appear as modifiers, 
even in a simple smixut. 
In the case of gerunds, the only possible structures are compound and double-genitive while the 
separate construction is not possible. 
Levi \[11\] claims that smixut realizes in Hebrew a number of universal semantic processes which 
exist inother languages, thus extending her original analysis for English \[12\]. Her ':non predicative 
modifiers" theory claims that Noun-noun compounding is produced by two •syntactic processes: 
nominalization or deletion of the predicate : which corresponds to the observed uses of smixut for 
possessive and gerund constructions. 
Azar \[2\] classifies smixut into 15 semantic categories. This classification can be made parallel 
to Levi's RDPs. Glinert \[10\] also refers to such a classification in a similar manner. 
172 
,I 
I 
I 
1 
il 
I 
;I 
3.4 Pragmatic Constraints 
There are cases, however, when smixut can be constructed with no regard to the semantic set that 
was identified. Certain contexts license smixut constructions that would not be obtained otherwise 
- for example, contrast or naming \[6\]. 
In addition, smixut is associated with style and genre parameters. Seikevicz \[16\] analyzes 
transcripts of spoken Hebrew, and finds that smixut is used when using a 'Sel' preposition is not 
possible. 
Other pragmatic considerations for the use of smixut include the objective to generate a more 
compact text and to make of a compound an item available for further anaphorical reference. 
Finally, decision to compound a head with a plural or singular modifier is related to the gener- 
icity of the description and to the habituality of the relation,.as is the case in English \[15\] (p. 916): 
The table in the corner was laid for dinner The corner table 
The girl in the corner spoke to me ~ The corner girl 
4 When Can • Smixut Be Generated? 
Our main objective is to determine what features must be present in the input to the syntactic 
realization component to decide when to use a smixut construction. 
We observe that the •production of smixut is semantically constrained, and that the semantic 
relation holding between head and modifiers determines which syntactic paraphrases are possible 
(among smixut, double-genitive and separate construct). A set of semantic predicates similar to 
Levi's RDPs seems to play a role in the decision. On the other hand, being a member of that set 
is not a sufficient nor necessary condition to generate a smixut. 
In the SURGE grammar for English, we did not address this decision, and assume that the 
input includes a predefined syntactic construction (classifier-head). For Hebrew, we must find an 
alternative approach because: (1) smixut is extremely frequent (40~0 of the noun modifiers in our 
corpus); (2) smixut is the default realization for many relations but it cannot be used in many 
syntactic contexts. 
4.1 Exploiting a Semantic Classification 
Our strategy is to provide in the input to HUGG a reliable default indicating that smixut should 
be used when possible, but making it possible to fall back on an alternative realization (separate 
or double genitive, or qualifier modification) when smixut is not possible. 
For instance, lexical-compounds cannot be head a smixut, and, therefore, their modifier must 
be realized as a PP. The same semantic relation (e.g., material) will be realized in two different 
ways depending on the lexical property of the head: 
coat leather head I lex "mewyl" \] 
leather coat modifiers \[ material \[lax "wor" \] \] 
If the same input is provided, but the property of the head noun is different, a different Con- 
struction will be generated: 
bathing-suit •from-leather head lax "beged-yam" • cat noun-compound 
leather bathing-suit modifiers \[ material \[lax "wor" \] \] 
173 
A similar mechanism would determine that a smixut is not possible if the definiteness of the 
head and the modifier do.not match, as discussed above in the a son o.f the king example. 
The syntactic realizer also relies on the semantic classification of the relation head-modifier when 
several modifiers are attached to a single head. In this case, only a single modifier can be realized as 
a smixut. The others must be realized differently. In this case, the realizer must determine which 
relation takes priority to become the smixut, and it must also provide an appropriate paraphrase 
for the non-smixut modifier. 
For example, the English NP leather house shoe will be generated in one of the following ways: 3 
nawal bayit me-wor cat common 
house shoe ~om-leather lex ~nawal" 
shoe leather .for-the-house purpose lex "bayit" \] 
Beyond smixut-related decisions, determining a set of semantic relations is also useful to allow 
HUGG to determine appropriate defaults for prepositions in PP modifiers. For example, in the 
example above, HUGG can select the default preposition .for in shoe \]or the house because the 
relation of purpose is specified. 
The same classification is also useful to determine the order and the syntactic structure of a 
' multi-modifier sequence in complex NPs. 
In general, when several modifiers attach to a single head, a broken (conjuncted) sequence is 
created \[9\] (in contrast to English, where a stacking construction is generally used): .4 big white 
house vs. bayit gadol ve-lavan (house big and white). 
However, when adjectives realize a semantic relation that could have been reaiized by a smixut, 
they appear first in the sequence of modifiers and they do not require a conjunction \[1\]. 
cat common 
makdeHa HaSmalyt gadola lex "makdeHa" In 
a large electronic drill modifiers lex "HaSmaly" 
size lex "gadol" \] 
In this example, makdeHa HaSmalyt gadola is produced instead of * makdeha gadola ve- 
HaSmalyt because the electric modifier realizes the smixut-licensing relation of instrument. This 
phenomenon gives a further justification for the use of semantic relations in the input. 
4.2 Classification of Semantic Relations 
Since the syntactic realization component can make good use of a semantic classification in the 
input, we have designed the classification shown in Table 1, which synthesizes the lists provided by 
Levi, Glinert and Azar. 
In the table we present a basic list of relations with its occurrence percentage in our corpus. 
It can be viewed that some relations are much more productive than others - purpose, has-part. 
Our classification is finer than Levi's in distinguishing for example among different types of typical 
possessive relations (human-relator, has-part and ownership). This reflects slight differences in the 
default way of generation. Human-relator, for instance, is used as construct when the modifier is a 
pronoun, more often than with other ownership relations. 
3VVe discuss below the heuristics HUGG uses to decide between these 2 paraphrases. 
• 174 
I 
I 
I 
I 
I 
I 
i 
I 
I 
I 
I 
I 
I 
I 
I- 
Relation %corpus Example 
nominalization 
purpose 
has-part 
location 
content . 
human-relator 
12.00% 
11.24% 
11.24% 
6.50% 
6.21% 
6.21% 
hacHaSat meZi'ut / denial reality / reality denial 
Simlat Hatuna / dress wedding / wedding dress 
weyney ha-yeled / eyes the-boy / the boy's eyes 
PirHey midbar / flowers desert / desert flowers 
wugat tapuHym / cake apples / apple cake " 
Em habanym / mother the-sons / mother of the sons 
type 
owner 
producer 
matter 
material 
idioms 
relational 
name 
5.91% 
5.91% 
5.02% 
4.14% 
3.84% 
3.84% 
3:55% 
2.95% 
regeS Ahavah / feeling love / love feeling 
mytat horay / bed my-parents / my parents' bed 
reyaH bSamym / scent perfium / perfume scent 
miSpat reZaH / trial murder / murder trial 
cise weZ / chair wood / wooden chair 
cadur ha-AreZ / ball the-land / earth 
Zevaw ha-baZek / color the-batter / the color of the batter 
miSpaHat netanyahu / family netanyahu / The Netanyahu family 
experiencer 
config-units 
represented 
part-of 
time 
cause 
caused-by 
product 
instrument 
2.95% 
1.45% 
1.18% 
0.88% 
0.88% 
O.88% 
0.29% 
0.29% 
0.29% 
ce'ev-a \] pain-her-acc/her pain 
zer praHym / bouquet flowers \] a bouquet of flowers 
semel yokrah / symbol prestige / prestige symbol 
nawaley waker / shoe heel / high-heels shoes 
AruHat Zaharym / meal noon / dinner 
macat HaSmal / hit electricity / electric shock 
yetuS kadaHat / mosquitoe malaria/malaria mosquitoe 
mifwal keramika / factory ceramics / ceramics factory 
magheZ edim / iron steam / steam iron 
Table 1: Semantic relations that can produce a smixut 
4.3 Validation of the Classification 
To validate empirically the definition of our semantic classification, we gathered a corpus of 853 
complex NPs (NPs with more than one modifier) from written Hebrew sources (newspaper and 
novels). For each NP, we labeled the relations head-modifier in terms of the relations listed in 
Table 1. 
Our evaluation covers two aspects: we first verify that human coders agree on the labeling; we 
then verify that HUGG can generate from a labeled input a realization similar to that observed in 
the corpus. 
Preliminary evaluation of the agreement among human judges shows agreement of about 90% 
between three judges (we are currently extending the number of judges). The percentage agreement 
includes a category "undecided" which covers about 5% of the cases. This corresponds to cases 
where judges found the relation ambiguous or unclear. Judges agreed on the labeling of unclear 
relations. 
In our corpus, we observed the following distribution in terms of syntactic realization (this takes 
into account NPs with more than one modifier, explaining that the sum is > 100%): 
175 
139% Ismixut 131% \]Pp-qualifier I 
34% describer 8% relative clause 
When regenerating from the labeled input we have determined, I-IUGG's decision to generate 
a smixut corresponded to that observed in the corpus on more than 95% of the cases. 
5 Limitations of a Semantic Account 
While the semantic account described above provides good results, it cannot be the only mechanism 
licensing the production of smixut. We discuss in this section the type of interaction that must be 
allowed between discourse and pragmatic parameters and the syntactic realization component. 
In \[5\], the interaction between lexical semantics and pragmatics is explored, and two axioms 
are proposed to interface between the defaults of the lexical semantic and the arbitrary knowledge 
of pragmatics: (1) defaults survive and (2) discourse wins. A statistical method is then added in 
Order to resolve possible interpretations. It isassumed, then, that the grammar/lexicon delimits 
the range of compounds and indicates conventional interpretations, but that some compounds may 
only be resolved by pragmatics and that non-conventional contextual interpretations are always 
available. To provide interpretations , a general schema is encoded in the lexicon leaving undecidable 
cases to be resolved by pragmatics. Probabilities of possible interpretations are taken from corpus 
frequencies. Accordingly, a new rule is added: (3) Prefer Frequent Senses, which can still be 
overridden by contextual factors. 
From the generation perspective, the interaction between discourse licensed-relations and con- 
ventional readings must similarly be controlled by preference rules. 
For example, when referring to a city destroyed by the Barbarian, discourse readings cannot 
override the conventional reading in The Barbarian city: discourse cannot force the reading of a 
city destroyed by the Barbarians. 
This indicates that a non-monotonic form of reasoning, taking into account preference rules 
similar to that identified in \[5\] must be implemented at the pragmatic level. Clearly, this type 
of reasoning does not belong within the syntactic realization component. Therefore, we conclude 
that the feature use-smixut remains a necessary part of the input specification language to the 
syntactic realization component. 
6 Conclusion 
We have presented in this paper basic data on the Hebrew smixut construction. Our strategy to 
implement smixut in the HUGG syntactic realization is to provide a simple semantic classification 
in the input. We have demonstrated the many benefits this classification has within the realization 
process. 
Two main problems have been traditionally associated with such semantic accounts of noun- 
compounding: the relations are not well-defined enough and they are not necessary nor sufficient 
to explain all uses of compounding. We address these two problems in three ways: (1) we provide 
an empirical evaluation demonstrating high coder agreement when labeling complex NPs with the 
set of relations we identify; (2) we demonstrate empirically that the default strategy of "generating 
a smixut when a semantic relation licenses it" corresponds with the observed usage of smixut in 
more than 95% of the cases; and (3) we allow the pragmatic module to add a feature use-smixut 
in the input. 
176 
I 
I 
I 
! 
! 
! 
! 
I 
! 
! 
I 
! 
! 
! 
The same set of semantic relations is now being used in an extension to SURGE to allow similar 
paraphrasing decisions in English. 

References 
\[1\] Talila Atias. The order of adjectives in Hebrew. Master's thesis, Tel Aviv University, Israel, 
1981. (in Hebrew). 
\[2\] M. Azar. Classification of hebrew compounds. In R. Nit, editor, Academic Teaching of Contem- 
porary Hebrew. International Center for University Teaching of Jewish Civilization, Jerusalem, 
1985. (in Hebrew). 
\[3\] Ruth Berman and Dorit Ravid. On the lexical degree of the constructs in hebrew. Hebrew 
Computational Linguistics, (24):5-22, 1986. (In Hebrew). 
\[4\] Hagit Borer. On the morphological parallelism between compounds and constructs. Depart- 
ment of Cognitive Science, School of Social Sciences, University of California, Irvinne. 
\[5\] Ann Copestake and Alex Lascarides. Integrating symbolic and statistical representation: The 
lexicon pragmatics interface. In Proceedings of the 35th Annual Meeting of the Association for 
Computational Linguistics, pages 136-143. ACL, 1997. 
\[6\] P. Downinig. On the creation and use of english compound nouns. Language, 53(4)~810-842, 
1977. 
\[7\] Michael Elhadad. Using argumentation to control lexical choice: a unification-based imple- 
mentation. PhD thesis, Computer Science Department, Columbia University, 1992. 
\[8\] Michael Elhadad. Lexical choice for complex noun phrases: Structure, modifiers and deter- 
miners. Machine Translation, 11:159-184, 1996. 
David Gil. Stacked adjectives and conflgurationality. Linguistic Analysis, (12):141-158, 1983. 
L. Glinert. The Grammar of Modern Hebrew. Cambridge University, 1989. 
J.N. Levi. A semantic analysis of hebrew compound nominals. In Peter Cole, editor, Studies 
in Modern Hebrew syntax and semantics. North-Holland, Amsterdam, 1976. 
J.N. Levi. The Syntax and Semantics of Complex Nominals. Academic Press, New York, 1978. 
Uzi Ornan. The Nominal Phrase in Modern Hebrew. PhD thesis, Hebrew University, Jerusalem, 
1964: (in Hebrew). 
James D. Pustejovsky. The generative lexicon. Computational Linguistics, 17(4):409-441, 
December 1991. 
R. Quirk, S. Greenbaum, G. Leech, and J. Svartvik. A comprehensive grammar of the English 
language. Longman, 1985. 
Channa Seikevicz. The Possessive Construction in Modern Hebrew: A Sociolinguistic Ap- 
proach. PhD thesis, Georgetown University, Washington D.C., 1979. 
