The Syntax and Semantics of Prepositions in - IA de Investigación especializada en el MIT

The Syntax and Semantics of Prepositions in
the Task of Automatic Interpretation of
Nominal Phrases and Compounds:
A Cross-Linguistic Study

Roxana Girju∗
University of Illinois at
Urbana-Champaign

In this article we explore the syntactic and semantic properties of prepositions in the context
of the semantic interpretation of nominal phrases and compounds. We investigate the problem
based on cross-linguistic evidence from a set of six languages: Inglés, Español, italiano, Francés,
Portuguese, and Romanian. The focus on English and Romance languages is well motivated.
Most of the time, English nominal phrases and compounds translate into constructions of the
form N P N in Romance languages, where the P (preposition) may vary in ways that correlate
with the semantics. De este modo, we present empirical observations on the distribution of nominal
phrases and compounds and the distribution of their meanings on two different corpora, based
on two state-of-the-art classiﬁcation tag sets: Lauer’s set of eight prepositions and our list of 22
relaciones semánticas. A mapping between the two tag sets is also provided. Además, given a
training set of English nominal phrases and compounds along with their translations in the ﬁve
Romance languages, our algorithm automatically learns classiﬁcation rules and applies them
to unseen test instances for semantic interpretation. Experimental results are compared against
two state-of-the-art models reported in the literature.

1. Introducción

Prepositions are an important and frequently used category in both English and Ro-
mance languages. In a corpus study of one million English words, Fang (2000) muestra
that one in ten words is a preposition. Además, acerca de 10% del 175 most frequent
words in a corpus of 20 million Spanish words were found to be prepositions (Almela
et al. 2005). Studies on language acquisition (Romaine 1995; Celce-Murcia and Larsen-
Hombre libre 1999) have shown that the acquisition and understanding of prepositions in
languages such as English and Romance is a difﬁcult task for native speakers, y
even more difﬁcult for second language learners. Por ejemplo, together with articles,
prepositions represent the primary source of grammatical errors for learners of English
as a foreign language (Gocsik 2004).

∗ Linguistics and Computer Science Departments, University of Illinois at Urbana-Champaign, Urbana, IL

61801. Correo electrónico: girju@illinois.edu.

Envío recibido: 1 Agosto 2006; revised submission received: 20 Enero 2008; accepted for publication:
17 Marzo 2008.

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

Although the complexity of preposition usage has been argued for and documented
by various scholars in linguistics, psycholinguistics, and computational linguistics,
very few studies have been done on the function of prepositions in natural language
Procesando (NLP) applications. The reason is that prepositions are probably the most
polysemous category and thus, their linguistic realizations are difﬁcult to predict and
their cross-linguistic regularities difﬁcult to identify (Saint-Dizier 2005a).

In this article we investigate the role of prepositions in the task of automatic seman-
tic interpretation of English nominal phrases and compounds. The problem is simple to
deﬁne: Given a compositional noun phrase (the meaning of the phrase derives from the
meaning of the constituents) constructed out of a pair of nouns, N1 N2, one representing
the head and the other the modiﬁer, determine the semantic relationship between the
two nouns. Por ejemplo, the noun–noun compound family estate encodes a POSSESSION
relation, while the nominal phrase the faces of the children refers to PART-WHOLE. El
problema, although simple to state, is difﬁcult for automatic semantic interpretation.
The reason is that the meaning of these constructions is most of the time implicit (él
cannot be easily recovered from morphological analysis). Interpreting nominal phrases
and compounds correctly requires various types of information, from world knowledge
to lexico-syntactic and discourse information.

This article focuses on nominal phrases of the type N P N and noun compounds
(N N) and investigates the problem based on cross-linguistic evidence from a set of six
idiomas: Inglés, Español, italiano, Francés, Portuguese, and Romanian. The choice of
these constructions is empirically motivated. In a study of 6,200 (Europarl1) y 2,100
(CLUVI2) English token nominal phrase and compound instances randomly chosen
from two English–Romance parallel text collections of different genres, we show that
encima 80% of their Romance noun phrase translations are encoded by N P N and N N
constructions. Por ejemplo, beer glass, an English compound of the form N1 N2, trans-
lates into N2 P N1 instances in Romance: tarro de cerveza (‘glass of beer’) in Spanish,
bicchiere da birra (‘glass for beer’) in Italian, verre `a bi`ere (‘glass at/to beer’) in French, copo
de cerveja (‘glass of beer’) in Portuguese, and pahar de bere (‘glass of beer’) in Romanian.
In this article, in addition to the sense translation (in italics), when relevant we also
provide the word-by-word gloss (in ‘parentheses’). Además, we use N1, N2 to denote
the two lexical nouns that encode a semantic relation (where N1 is the syntactic modiﬁer
and N2 is the syntactic head), and Arg1, Arg2 to denote the semantic arguments of the
relation encoded by the two nouns. Por ejemplo, beer glass encodes a PURPOSE relation
where Arg1 (beer) is the purpose of Arg2 (‘glass’; thus ‘glass (usado) for beer’).

We argue here that the syntactic directionality given by the head-modiﬁer relation
(N1 N2 in noun compounds and N2 P N1 in nominal phrases) is not always the same
as the semantic directionality given by the semantic argument frame of the semantic
relation. Otherwise said, N1 does not always map to Arg1 and N2 to Arg2 for any given
relation.

Languages choose different nominal phrases and compounds to encode relation-
ships between nouns. Por ejemplo, English nominal phrases and compounds of the

1 http://www.isi.edu/koehn/europarl/.

This corpus contains over 20 million words in eleven ofﬁcial languages of the European Union covering
the proceedings of the European Parliament from 1996 a 2001.

2 CLUVI – Linguistic Corpus of the University of Vigo Parallel Corpus 2.1; http://sli.uvigo.es/CLUVI/.
CLUVI is an open text repository of parallel corpora of contemporary oral and written texts in some of
the Romance languages (such as Galician, Francés, Español, and Portuguese) and Basque parallel text
collections.

186

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

form N1 N2 (p.ej., wood stove) and N2 P1 N1 (p.ej., book on the table) usually translate
in Romance languages as N2 P2 N1 (p.ej., four `a bois in French – ‘stove at/to wood’,
and livre sur la table – ‘book on the table’). Romance languages have very few N N
compounds and they are of limited semantic categories, such as TYPE (p.ej., legge quadro
in Italian – ‘law framework’ – translates as framework law). Besides the unproductive
N N and the productive N P N phrases, Romanian also uses another productive con-
estructura: the genitive-marked noun–noun compounds (p.ej., frumuset¸ea fetei – beauty-
the girl-GEN – translated as the beauty of the girl). Whereas English N N compounds
are right-headed (p.ej., framework/Modiﬁer law/Head), Romance compounds are left-
headed (p.ej., legge/Head quadro/Modiﬁer). Además, the Romance preposition used in
the translations of English nominal phrase instances of the type N P N is one that comes
closest to having overlapping semantic range as intended in the English instance, pero
may not be the exact counterpart for the whole semantic range. Por ejemplo, Committee
on Culture translates as Comisi´on de la Cultura (Español) (‘Committee of the Culture’),
Commission de la Culture (Francés) (‘Committee of the Culture’), Commissione per la Cul-
tura (italiano) (‘Committee for the Culture’), Comiss˜ao para Cultura (Portuguese) (‘Com-
mittee for Culture’), and Comitet pentru Cultur˘a (Romanian) (‘Committee for Culture’).
Even those Romance prepositions that are spelled “de” are pronounced differently in
different Romance languages.

De este modo, the focus on nominal phrases and compounds in English and Romance lan-
guages is also motivated linguistically. The extension of this task to natural languages
other than English brings forth both new insights and new challenges. The Romance
prepositions used in the translations of English nominal phrases and compounds, may
vary in ways that correlate with the semantics. De este modo, Romance language prepositions
will give us another source of evidence for disambiguating the semantic relations in
English nominal phrases and compounds. We argue that, in languages with multiple
syntactic options such as English (N N and N P N) and Romanian (N N, genitive-
marked N N, and N P N), the choice between such constructions in context is governed
in part by semantic factors. Por ejemplo, the set of semantic relations that can be
encoded by pairs of nouns such as tea–cup and sailor–suit varies with the syntactic
construction used. In English, while the noun–noun compounds tea cup and sailor suit
encode only PURPOSE, the N P N constructions cup of tea and suit of the sailor encode
CONTENT-CONTAINER (a subtype of LOCATION) and MEASURE relations and POSSES-
SION, respectivamente. Similarmente, in Romanian both tea cup and cup of tea translate only as
the N P N instance cea¸sc˘a de ceai (‘cup of tea’), while sailor suit translates as costum de
marinar (‘suit of sailor’) and the suit of the sailor as the genitive-marked N N costumul
marinarului (‘suit-the sailor-GEN’). De este modo, we study the distribution of semantic relations
across different nominal phrases and compounds in one language and across all six
idiomas, and analyze the resulting similarities and differences. This distribution is
evaluated over the two different corpora based on two state-of-the-art classiﬁcation tag
conjuntos: Lauer’s set of eight prepositions (Lauer 1995) and our list of 22 relaciones semánticas.
A mapping between the two tag sets is also provided.

In order to test their contribution to the task of semantic interpretation, preposi-
tions and other linguistic clues are employed as features in a supervised, conocimiento-
intensive model. Además, given a training set of English nominal phrases and
compounds along with their translations in the ﬁve Romance languages, our algo-
rithm automatically learns classiﬁcation rules and applies them to unseen test instances
for semantic interpretation. As training and test data we used 3,124 Europarl and
2,023 CLUVI token instances. These instances were annotated with semantic relations
and analyzed for inter-annotator agreement. The results are compared against two

187

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

state-of-the-art approaches: a supervised machine learning model, semantic scattering
(Moldovan and Badulescu 2005), and a Web-based unsupervised model (Lapata and
Keller 2005). Además, we show that the Romanian linguistic features contribute more
substantially to the overall performance than the features obtained for the other Ro-
mance languages. This is explained by the fact that the choice of the linguistic construc-
ciones (either genitive-marked N N or N P N) in Romanian is highly correlated with their
significado.

The article is organized as follows. Sección 2 presents a summary of related work.
En la sección 3 we describe the general approach to the interpretation of nominal phrases
and compounds and list the syntactic and semantic interpretation categories used
along with observations regarding their distribution in the two different cross-linguistic
corpus. Secciones 4 y 5 present a learning model and experimental results. Sección 6
presents linguistic observations on the behavior of English and Romanian N N and
N P N constructions. Finalmente, en la sección 7 we provide an error analysis and in Section 8
we offer some discussion and conclusions.

2. Previous Work

2.1 Noun Phrase Semantic Interpretation

The semantic interpretation of nominal phrases and compounds in particular and noun
phrases (NPs) in general has been a long-term research topic in linguistics, computa-
tional linguistics,3 and artiﬁcial intelligence.

Noun–noun compounds in linguistics
Early studies in linguistics (Lees 1963) classiﬁed noun–noun compounds on purely
grammatical criteria using a transformational approach, criteria which failed to account
for the large variety of constraints needed to interpret these constructions. Later on, leví
(1978) attempted to give a tight account of noun–noun interpretation, distinguishing
two types of noun–noun compounds: (a) compounds interpreted as involving one of
nine predicates (CAUSE, HAVE, MAKE, USE, BE, EN, FOR, FROM, ABOUT) (p.ej., onion
tears encodes CAUSE) y (b) those involving nominalizations, a saber, compounds
whose heads are nouns derived from a verb, and whose modiﬁers are interpreted as
arguments of the related verb (p.ej., a music lover loves music). Levi’s theory was cast
in terms of the more general theory of Generative Semantics. In that theory it was
assumed that the interpretation of compounds was available because the examples were
derived from underlying relative clauses that had the same meanings. De este modo, honey bee,
expressing the relation MAKE, was taken to be derived from a headed relative a bee
that makes honey. Levi was committed to the view that a very limited set of predicates
constituted all of the relations that could hold between nouns in simple noun–noun
compounds. This reductionist approach has been criticized in studies of language use
by psycholinguists (Gleitman and Gleitman 1970; Downing 1977) who claim that noun–
noun compounds, which are frequent in languages like English, encode in principle an

3 In the past few years at many workshops, tutorials, and competitions this research topic has received
considerable interest from the computational linguistics community: the Workshops on Multiword
Expressions at ACL 2003, LCA 2004 and COLING/ACL 2006; the Computational Lexical Semantics
Workshop at ACL 2004; the Tutorial on Knowledge Discovery from Text at ACL 2003; the Shared Task on
Semantic Role Labeling at CONLL 2004 y 2005 and at SemEval 2007.

188

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

unbounded number of possible relations. One such example is apple juice seat—“a seat
in front of which an apple juice [es] placed” (Downing 1977, página 818)—which can only
be interpreted in the current discourse context.

In this article we tackle the problem using a uniﬁed framework. Although we
agree with Downing (1977) that pragmatics plays an important factor in noun–noun
interpretación, a large variety of noun–noun meanings can be captured with a well-
chosen set of semantic relations. Our proposed semantic classiﬁcation set differs from
that of Levi (1978) in the sense that it contains more homogenous categories. Levi’s
categories, en cambio, are more heterogeneous, including both prepositions and verbs,
some of which are too general (p.ej., the prepositions for, in and the verb to have), y
de este modo, too ambiguous. Además, in our approach to automatic semantic interpretation
we focus on both N N and N P N constructions and exploit a set of ﬁve Romance
idiomas.

Noun–noun compounds in computational linguistics
The automatic interpretation of nominal phrases and compounds is a difﬁcult task
for both unsupervised and supervised approaches. Actualmente, the best-performing
noun–noun interpretation methods in computational linguistics focus mostly on two
or three-word noun–noun compounds and rely either on ad hoc, domain-speciﬁc,
hand-coded semantic taxonomies, or statistical models on large collections of unlabeled
datos. Recent results have shown that symbolic noun–noun compound interpretation
systems using machine learning techniques coupled with a large lexical hierarchy
perform with very good accuracy, but they are most of the time tailored to a speciﬁc
domain (Rosario and Hearst 2001; Rosario, Hearst, and Fillmore 2002), or are general
purpose (Turney 2006) but rely on semantic similarity metrics on WordNet (Fellbaum
1998). Por otro lado, the majority of corpus statistics approaches to noun–noun
compound interpretation collect statistics on the occurrence frequency of the noun
constituents and use them in a probabilistic model (Lauer 1995). The problem is that
most noun–noun compounds are rare and thus, statistics on such infrequent instances
lead in general to unreliable estimates of probabilities. More recently, Lapata and Keller
(2005) showed that simple unsupervised models applied to the noun–noun compound
interpretation task perform signiﬁcantly better when the n-gram frequencies are
obtained from the Web (55.71% accuracy4), rather than from a large standard corpus.
Nakov and Hearst (2005) improve over Lapata and Keller’s method through the use of
surface features and paraphrases only for the task of noun–noun compound bracketing
(syntactic parsing of three-word noun compounds) without their interpretation.
Other researchers (Pantel and Ravichandran 2004; Pantel and Pennacchiotti 2006;
Pennacchiotti and Pantel 2006) use clustering techniques coupled with syntactic
dependency features to identify IS-A relations in large text collections. Kim and Baldwin
(2005) propose a general-purpose method that computes the lexical similarity of unseen
noun–noun compounds with those found in training. More recently Kim and Baldwin
(2006) developed an automatic method for interpreting noun–noun compounds based
on a set of 20 relaciones semánticas. The relations are detected based on a ﬁxed set of
constructions involving the constituent nouns and a set of seed verbs denoting the
semantic relation (p.ej., to own denotes POSSESSION). Then all noun–noun instances

4 These results were obtained on AltaVista on a general and abstract set of eight prepositions (Lauer 1995)

as semantic classiﬁcation categories: de, para, con, en, en, en, acerca de, and from.

189

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

in transitive sentential contexts (es decir., those sentences containing a transitive verb) son
mapped onto the selected set of constructions based on lexical similarity over the verbs.
Sin embargo, although the Web-based solution might overcome the data sparsity prob-
lem, current probabilistic models are limited because they do not take full advantage of
the structure and the meaning of language.

From a cross-linguistic perspective, there hasn’t been much work on the automatic
interpretation of nominal phrases and compounds. Busa and Johnston (1996), Johnston
and Busa (1996), and Calzolari et al. (2002), Por ejemplo, focus on the differences
between English and Italian noun–noun compounds. In their work they argue that a
computational approach to the cross-linguistic interpretation of these compounds has to
rely on a rich lexical representation model, such as those provided by FrameNet frames
(Panadero, Fillmore, and Lowe 1998) and qualia structure (Pustejovsky 1995). In the qualia
structure representation, Por ejemplo, the meaning of a lexical concept, como el
modiﬁer in a noun–noun compound, is deﬁned in terms of four elements representing
concept attributes along with their use and purpose. De este modo, qualia structure provides
a relational structure that enables the compositional interpretation of the modiﬁer in
relation to the head noun. Two implementations of such representations are provided
by the SIMPLE Project ontology (Lenci et al. 2000) and the OMB ontology (Pustejovsky
et al. 2006). The SIMPLE ontology, Por ejemplo, is developed for 12 European languages
and deﬁnes entry words that are mapped onto high-level concepts in EuroWordNet
(zorros 1998), a version of WordNet developed for European languages.

In this article, we use a supervised semantic interpretation model employing rich
linguistic features generated from corpus evidence coupled with word sense disam-
biguation and WordNet concept structure information. The results obtained are com-
pared against two state-of-the-art approaches: a supervised machine learning model,
semantic scattering (Moldovan and Badulescu 2005), and a Web-based unsupervised
modelo (Lapata and Keller 2005). In this research we do not consider extra cross-linguistic
información, such as semantic classes of Romance nouns (those provided by IS-A re-
laciones; p.ej., cat belongs to the class of animals) made available, Por ejemplo, by the
SIMPLE ontology. Sin embargo, such resources can be added at any time to further improve
the performance of noun–noun interpretation systems.

2.2 Semantics of Prepositions

Although prepositions have been studied intensively in linguistics (Herskovits 1987;
Zelinski-Wibbelt 1993; Linstromberg 1997; Tyler and Evans 2003; Evans and Chilton
2009, among others), they have only recently started to receive more attention in the
computational linguistics community.5 Moreover, the ﬁndings from these broad stud-
ies have not yet been fully integrated into NLP applications. Por ejemplo, a pesar de
information retrieval, and even question answering systems, would beneﬁt from the
incorporation of prepositions into their NLP techniques, they often discard them as stop
palabras.

5 The ﬁrst Workshop on the Syntax and Semantics of Prepositions, Tolosa, Francia, 2003; the second

ACL-SIGSEM Workshop on The Linguistic Dimensions of Prepositions and their Use in Computational
Linguistics Formalisms and Applications, Colchester, Reino Unido, 2005; the third ACL-SIGSEM Workshop on
Prepositions, trento, Italia, 2006.

190

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Prepositions in linguistics
Considerable effort has been allocated to the investigation of spatial prepositions mainly
based on a cognitive approach, not only in English (Herskovits 1987; Linstromberg 1997;
Tyler and Evans 2003; Evans and Chilton 2009), but also in many of the Indo-European
idiomas (Casadei 1991; Vandeloise 1993; Cadiot 1997; Melis 2002; Luraghi 2003). Estos
studies provide a detailed analysis of such prepositions trying to give a methodologi-
cal motivated account for the range of their polysemy. These works identify special
constraints on various prepositional patterns, such as semantic restrictions on the noun
phrases occurring as complements of the preposition. Por ejemplo, in prepositional
phrase constructions such as in NP, the head noun can be a container (in a cup), a
geometrical area (in a region), a geo-political area (in Paris), an atmospheric condition
(in the rain), etcétera. These selectional restrictions imposed by the preposition on the
noun phrases it combines with are presented in various formats from lists (Herskovits
1987; Linstromberg 1997) to semantic networks of cluster senses (Tyler and Evans 2003).
In this article we also focus on the polysemy of such prepositions, but we identify the se-
lectional restrictions automatically based on a specialization procedure on the WordNet
IS-A hierarchy. Sin embargo, unlike Herskovits, we do not consider pragmatic issues such
as relevance and tolerance. These account for the difference that pragmatic motivations
and context dependency make to how expressions are understood. Relevance has to do
with communicative goals and choice of means and is evident, Por ejemplo, in instances
such as cat on the mat which is still relevant even when only the paws and not the
whole cat are on the mat. Tolerance occurs in situations in which a book, Por ejemplo,
is described as on the table even though a set of ﬁles are placed between it and the
mesa.

The use of spatial prepositions can also trigger various inferences. Por ejemplo,
the man at his desk (cf. Herskovits 1987) implies, besides a LOCATION relation, that the
man is using the desk, thus an INSTRUMENT relation. Other inferences are more subtle,
involving spatial reasoning about the actions that can be performed on the arguments
of the preposition. One such instance is infant in a playpen (cf. Tyler and Evans 2003),
where the movement of the playpen involves the movement of the infant. In order to
identify such inferences the automatic interpretation system has to rely on pragmatic
conocimiento. In this research we do not deal with such inference issues, rather we identify
the meaning of N P N constructions based on the local context of the sentence.

Prepositions in computational linguistics
In order to incorporate prepositions into various resources and applications, it is neces-
sary to perform ﬁrst a systematic investigation of their syntax and semantics. Varios
investigadores (Dorr 1993; Litkowski and Hargraves 2005; Saint-Dizier 2005b; Lersundi
and Aggire 2006) have already provided inventories of preposition senses in English
and other languages. Others have focused on the analysis of verb particles (Baldwin
2006a, 2006b; Villavicencio 2006), the distributional similarity (Baldwin 2005) y el
semantics of prepositions (Kordoni 2005) in a multilingual context, and the meaning
of prepositions in applications such as prepositional phrase attachment (O’Hara and
Wiebe 2003; Kordoni 2006; Volk 2006).

Además, although there is a large amount of work in linguistics and computa-
tional linguistics relating to contrastive analysis of prepositions (Busa and Johnston
(1996); Johnston and Busa (1996); Jensen and Nilsson (2005); Kordoni (2005), inter alia),
a nuestro conocimiento, there have not been any attempts to provide an investigation of the
prepositions’ role in the task of automatic noun phrase interpretation in a large cross-
linguistic English–Romance framework.

191

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

3. Linguistic Considerations of Nominal Phrases and Compounds

The meaning of nominal phrases and compounds can be compositional (p.ej., spoon
handle—PART–WHOLE, kiss in the morning—TEMPORAL), or idiosyncratic, cuando el
meaning is a matter of convention (p.ej., soap opera, sea lion). These constructions can
also encode metaphorical names (p.ej., ladyﬁnger), proper names (p.ej., John Doe), y
dvandva compounds6 in which neither noun is the head (p.ej., player–coach).

Además, they can also be classiﬁed into synthetic (verbal, p.ej., truck driver) y
raíz (non-verbal, p.ej., tea cup) constructions.7 It is widely held (leví 1978; Selkirk 1982b)
that the modiﬁed noun of a synthetic noun–noun compound, Por ejemplo, may be
associated with a theta-role of the compound’s head noun, which is derived from a
verb. Por ejemplo, in truck driver, the noun truck satisﬁes the THEME relation associated
with the direct object in the corresponding argument structure of the verb to drive.

In this article we address English–Romance compositional nominal phrases and
compounds of the type N N (noun–noun compounds which can be either genitive-
marked or not genitive-marked) and N P N, and disregard metaphorical names, adecuado
names, and dvandva structures. In the following we present two state-of-the-art se-
mantic classiﬁcation sets used in automatic noun–noun interpretation and analyze their
distribution in two different corpora.

3.1 Lists of Semantic Classiﬁcation Relations

Although researchers (Jespersen 1954; Downing 1977) argued that noun–noun com-
pounds, and noun phrases in general, encode an inﬁnite set of semantic relations,
many agree (leví 1978; Finin 1980) there is a limited number of relations that occur
with high frequency in these constructions. Sin embargo, the number and the level of
abstraction of these frequently used semantic categories are not agreed upon. They can
vary from a few prepositions (Lauer 1995) to hundreds and even thousands of more
speciﬁc semantic relations (Finin 1980). The more abstract the category, the more noun
phrases are covered, but also the larger the variation as to which category a phrase
should be assigned. Lauer, Por ejemplo, classiﬁes the relation between the head and the
modiﬁer nouns in a noun–noun compound by making use of a set of eight frequently
used prepositions: de, para, con, en, en, en, acerca de, and from. Sin embargo, according to this
classiﬁcation, the noun–noun compound love story, por ejemplo, can be classiﬁed both
as story of love and story about love. The main problem with these abstract categories
is that much of the meaning of individual compounds is lost, and sometimes there is
no way to decide whether a form is derived from one category or another. En el otro
mano, lists of very speciﬁc semantic relations are difﬁcult to build as they usually contain
a very large number of predicates, such as the list of all possible verbs that can link the
noun constituents. Finin, Por ejemplo, uses semantic categories such as dissolved in to
build interpretations of compounds such as salt water and sugar water.

In this article we experiment with two sets of semantic classiﬁcation categories
deﬁned at different levels of abstraction. The ﬁrst is a core set of 22 relaciones semánticas
(SRs), a set which was identiﬁed by us from the linguistics literature and from various
experiments after many iterations over a period of time (Moldovan and Girju 2003).

6 The term dvandva comes from Sanskrit, translates literally as ‘two-and-two’ and means ‘pair’.
7 In the linguistic literature the words “synthetic” and “root” have been coined for noun–noun compounds.
Because these terms apply also to nominal phrases, we use them in relation to these constructions as well.

192

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Moldovan and Girju proved empirically that this set is encoded by noun–noun pairs in
noun phrases; the set is a subset of their larger list of 35 semantic relations used in a large
set of semantics tasks. This list, presented in Table 1 along with examples and semantic
argument frames, is general enough to cover a large majority of text semantics while
keeping the semantic relations to a manageable number. A semantic argument frame is
deﬁned for each semantic relation and indicates the position of each semantic argument
in the underlying relation. Por ejemplo, “Arg2 is part of (entero) Arg1” identiﬁes the part
(Arg2) and the whole (Arg1) entities in this relation. This representation is important
because it allows us to distinguish between different arrangements of the arguments
for given relation instances. Por ejemplo, most of the time, in N N compounds Arg1
precedes Arg2, whereas in N P N constructions the position is reversed (Arg2 P Arg1).
Sin embargo, this is not always the case as shown by N N instances such as ham/Arg2
sandwich/Arg1 and spoon/Arg1 handle/Arg2, both encoding PART–WHOLE. More details
on subtypes of PART–WHOLE relations are presented in Section 6.2. A special relation
here is KINSHIP, which is encoded only by N P N constructions and whose argument
order is irrelevant. De este modo, the labeling of the semantic arguments for each relation as
Arg1 and Arg2 is just a matter of convention and they were introduced to provide a
consistent guide to the annotators to easily test the goodness-of-ﬁt of the relations. El
examples in column 4 are presented with their WordNet senses identiﬁed in context
from the CLUVI and Europarl text collections, where the speciﬁc sense is represented
as the sense number preceded by a “#” sign.

The second set is Lauer’s list of eight prepositions (exempliﬁed in Table 2) and can
be applied only to noun–noun compounds, because in N P N instances the preposition
is explicit. We selected these two state-of-the-art sets as they are of different size and
contain semantic classiﬁcation categories at different levels of abstraction. Lauer’s list is
more abstract and thus capable of encoding a large number of noun–noun compound
instances found in a corpus (p.ej., many N1 N2 instances can be paraphrased as N2 of

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

Mesa 1
The set of 22 semantic relations along with examples interpreted in context and the semantic
argument frame.

No.

Semántico
relaciones

Default argument frame

Examples

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

POSSESSION
KINSHIP
PROPERTY
AGENT
TEMPORAL
DEPICTION-DEPICTED
PART-WHOLE
HYPERNYMY (IS-A)
CAUSE
MAKE/PRODUCE
INSTRUMENT
LOCATION
PURPOSE
FUENTE
TOPIC
MANNER
MEANS
EXPERIENCER
MEASURE
TYPE
THEME
BENEFICIARY

Arg1 POSSESSES Arg2
Arg1 IS IN KINSHIP REL. WITH Arg2
Arg2 IS PROPERTY OF Arg1
Arg1 IS AGENT OF Arg2
Arg1 IS TEMPORAL LOCATION OF Arg2
Arg2 DEPICTS Arg1
Arg2 IS PART OF (entero) Arg1
Arg1 IS A Arg2
Arg1 CAUSES Arg2
Arg1 PRODUCES Arg2
Arg1 IS INSTRUMENT OF Arg2
Arg2 IS LOCATED IN Arg1
Arg1 IS PURPOSE OF Arg2
Arg1 IS SOURCE OF Arg2
Arg1 IS TOPIC OF Arg2
Arg1 IS MANNER OF Arg2
Arg1 IS MEANS OF Arg2
Arg1 IS EXPERIENCER OF Arg2
Arg2 IS MEASURE OF Arg1
Arg2 IS A TYPE OF Arg1
Arg1 IS THEME OF Arg2
Arg1 IS BENEFICIARY OF Arg2
OTHERS

family#2/Arg1 estate#2/Arg2
the sister#1/Arg2 of the boy#1/Arg1
lubricant#1/Arg1 viscosity#1/Arg2
investigation#2/Arg2 of the police#1/Arg1
morning#1/Arg1 news#3/Arg2
a picture#1Arg2 of my nice#1/Arg1
faces#1/Arg2 of children#1/Arg1
daisy#1/Arg1 ﬂower#1/Arg2
scream#1/Arg2 of pain#1/Arg1
chocolate#2/Arg2 factory#1/Arg1
laser#1/Arg1 treatment#1/Arg2
castle#1/Arg2 in the desert#1/Arg1
cough#1/Arg1 syrup#1/Arg2
grapefruit#2/Arg1 oil#3/Arg2
weather#1/Arg1 report#2/Arg2
performance#3/Arg2 with passion#1/Arg1
bus#1/Arg1 service#1/Arg2
the fear#1/Arg2 of the girl#1/Arg1
inches#1/Arg2 of snow#2/Arg1
framework#1/Arg1 law#2/Arg2
acquisition#1/Arg2 of stock#1/Arg1
reward#1/Arg2 for the ﬁnder#1/Arg1
cry of death

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

193

Ligüística computacional

Volumen 35, Número 2

Mesa 2
Lauer’s set of prepositions along with examples interpreted in context.

No.

Preposition

Examples

1
2
3
4
5
6
7
8

de
para
con
en
en
en
acerca de
de

sea bottom (bottom of the sea)
leisure boat (boat for leisure)
spoon feeding (feeding with a spoon)
London house (house in London)
Saturday snowstorm (snowstorm on Saturday)
night ﬂight (ﬂight at night)
war story (story about war)
almond butter (butter from almonds)

N1), whereas our list contains ﬁner grained semantic categories (p.ej., only some N1 N2
instances encode a CAUSE relation).

In the next section, we present the coverage of these semantic lists on two different
corpus, how well they solve the interpretation problem of noun phrases, y el
mapping from one list to another.

3.2 Corpus Analysis

For a better understanding of the semantic relations encoded by N N and N P N
instancias, we analyzed the semantic behavior of these constructions on two large
cross-linguistic corpora of examples. Our intention is to answer questions like:

(1) What syntactic constructions are used to translate the English instances to the target

Romance languages and vice versa? (cross-linguistic syntactic mapping)

(2) What semantic relations do these constructions encode? (cross-linguistic semantic

mapping)

(3) What is the corpus distribution of the semantic relations per each syntactic construction?

(4) What is the role of English and Romance prepositions in the semantic interpretation of

nominal phrases and compounds?

For questions (1) y (2), we expand the work of Selkirk (1982b), Grimshaw
(1990), Giorgi and Longobardi (1991), and Alexiadou, Haegeman, and Stavrou (2007)
on the syntax of noun phrases in English and Romance languages by providing
cross-linguistic empirical evidence for in-context instances on two different corpora
based on the set of 22 semantic tags. Following a conﬁgurational approach, Giorgi and
Lombardos, Por ejemplo, focus only on synthetic nominal phrases, such as the capture
of the soldier (THEME), where the noun capture is derived through nominalization
from the verb to capture. Besides synthetic constructions, we also consider root nominal
phrases and compounds, such as family estate (POSSESSION).

Los datos
In order to perform empirical investigations of the semantics of nominal phrases and
compounds, and to train and test a learning model for the interpretation of noun–noun

194

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

instances encoded by these constructions, we collected data from two text collections
with different distributions and of different genres, Europarl and CLUVI.

The Europarl data were assembled by combining four of the bilingual sentence-
aligned corpora made public as part of the freely available Europarl corpus. Specif-
icamente, the Spanish–English, Italian–English, French–English and Portuguese–English
corpora were automatically aligned based on exact matches of English translations.8
Entonces, only those English sentences which appeared verbatim in all four language pairs
were considered. The resulting English corpus contained 10,000 sentences which were
syntactically parsed using Charniak’s parser (Charniak 2000). From these we extracted
6,200 token instances of N N (49.62%) and N P N (50.38%) constructions.

CLUVI (Linguistic Corpus of the University of Vigo) is an open text repository of
parallel corpora of contemporary oral and written languages, a resource that besides
Galician also contains literary text collections in other Romance languages. Because the
collection provides translations into only two of the Romance languages considered
aquí, Spanish and Portuguese, we focused only on the English–Portuguese and English–
Spanish literary parallel texts from the works of Agatha Christie, James Joyce, and H. GRAMO.
Wells, among others. Using the CLUVI search interfaces we created a sentence-aligned
parallel corpus of 4,800 unique English–Portuguese–Spanish sentences. The English
version was syntactically parsed using Charniak’s parser (Charniak 2000) después de lo cual
each N N and N P N instance was manually mapped to the corresponding translations.
The resulting corpus contains 2,310 English token instances with a distribution of
25.97% N N and 74.03% N P N.

Corpus annotation and inter-annotator agreement
For each corpus, each nominal phrase and compound instance was presented separately
to two experienced annotators9 in a Web interface in context along with the English
sentence and its translations. Because the corpora do not cover some of the languages
(Romanian in Europarl, and Romanian, italiano, and French in CLUVI), three other
native speakers of these languages who were ﬂuent in English provided the translations,
which were added to the list. The two computational semantics annotators had to tag
each English constituent noun with its corresponding WordNet sense.10 If the word was
not found in WordNet the instance was not considered. The annotators were also asked
to identify the translation phrases, tag each instance with the corresponding semantic
relation, and identify the semantic arguments Arg1 and Arg2 in the semantic argument
frame of the corresponding relation. Whenever the annotators found an example encod-
ing a semantic relation or a preposition paraphrase other than those provided, or if they
did not know what interpretation to give, they had to tag it as OTHER-SR (p.ej., melody
of the pearl: here the context of the sentence did not indicate the association between the
two nouns; cry of death: the cry announcing death), and OTHER-PP (p.ej., box by the wall,
searches after knowledge) respectivamente.

Tagging each noun constituent with the corresponding WordNet sense in context is
important not only as a feature employed in the training models, but also as guidance
for the annotators to select the right semantic relation. Por ejemplo, in the follow-
ing sentences, daisy ﬂower expresses a PART–WHOLE relation in Example (1) and an

8 This version of the Europarl text collection does not include Romanian.
9 The annotators have extensive expertise in computational semantics and are ﬂuent in at least three of the

Romance languages considered for this task.

10 We used version 2.1 of WordNet.

195

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

IS-A relation in Example (2) depending on the sense of the noun ﬂower (cf. WordNet
2.1: ﬂower#2 is a “reproductive organ of angiosperm plants especially one having
showy or colorful parts,” whereas ﬂower#1 is “a plant cultivated for its blooms or
blossoms”).

(1)

(2)

Usually, more than one daisy#1 ﬂower#2 grows on top of a single stem.

Try them with orange or yellow ﬂowers of red-hot poker, solidago, or other late
daisy#1 ﬂowers#1, such as rudbeckias and heliopsis.

In cases where noun senses were not enough for relation selection, the annotators

had to rely on a larger context provided by the sentence and its translations.

Además, because the order of the semantic arguments in a nominal phrase or
noun–noun compound is not ﬁxed (Girju et al. 2005), the annotators were presented
with the semantic argument frame for each of the 22 semantic relations and were
asked to tag the instances accordingly. Por ejemplo, in PART–WHOLE instances such
as chair/Arg1 arm/Arg2 the part arm follows the whole chair, whereas in spoon/Arg1
handle/Arg2 the order is reversed. In the annotation process the translators also used
the ﬁve corresponding translations as additional information in selecting the semantic
relation. Por ejemplo, the context provided by the Europarl English sentence in Exam-
por ejemplo (3) does not give enough information for the disambiguation of the English nominal
phrase judgment of the presidency, where the modiﬁer noun presidency can be either
AGENT or THEME in relation to the nominalized noun head judgment. The annotators
had to rely on the Romance translations in order to identify the correct meaning in
contexto (THEME): valoraci´on sobre la Presidencia (Sp. – Spanish), avis sur la pr´esidence
(Fr. – French), giudizio sulla Presidenza (Él. – Italian), veredicto sobre a Presidˆencia (Port. –
Portuguese), evaluarea Pre¸sendint¸iei (Ro. – Romanian).

Most of the time, one instance was tagged with one semantic relation, y uno
preposition paraphrase (in case of noun–noun compounds), but there were also situa-
tions in which an example could belong to more than one category in the same context.
Por ejemplo, Texas city is tagged as PART–WHOLE, but also as a LOCATION relation using
the 22-SR classiﬁcation set, and as of, de, in based on the 8-PP set (p.ej., city of Texas,
city from Texas, and city in Texas). En general, 8.2% CLUVI and 4.8% Europarl instances
were tagged with more than one semantic relation, and almost half of the noun–noun
compound instances were tagged with more than one preposition.

(3)

En.:

Sp.:

Fr.:

It.:

Port.:

Ro.:

If you do , the ﬁnal judgment of the Spanish presidency will be even more
positive than it has been so far.
Si se hace, la valoraci ´on sobre la Presidencia espa ˜nola del Consejo
ser´a a ´un mucho m´as positiva de lo que es hasta ahora.
Si cela arrive, notre avis sur la pr´esidence espagnole du Conseil sera
encore beaucoup plus positif que ce n’est d´ej`a le cas.
Se ci riuscir`a, il nostro giudizio sulla Presidenza spagnola sar`a ancora
pi `u positivo di quanto non sia stato ﬁnora.
Se isso acontecer, o nosso veredicto sobre a Presidˆencia espanhola ser´a
ainda muito mais positivo do que o actual.
Dac˘a are loc, evaluarea Pres¸edint¸iei spaniole va ﬁ ˆınc˘a mai positiva
decˆat pˆan˘a acum.

196

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

De este modo, the corpus instances used in the corpus analysis phase have the following
format: (cid:2)NPEn; NPEs; NPIt; NPFr; NPPort; NPRo; objetivo(cid:3). The word target is one of the
23 (22 + OTHER-SR) semantic relations and one of the eight prepositions considered for
noun compound instances, and one of the 23 semantic relations for N P N instances. Para
ejemplo, (cid:2)development cooperation; cooperaci´on para el desarrollo; cooperazione allo sviluppo;
coop´eration au d´eveloppement; coopera¸c˜ao para o desenvolvimento; cooperare de dezvoltare;
PURPOSE / FOR(cid:3).

Inter-annotator agreement was measured using kappa, one of the most frequently
used measures of inter-annotator agreement for classiﬁcation tasks: K = Pr(A)−Pr(mi)
,
1−Pr(mi)
where Pr(A) is the proportion of times the annotators agree and Pr(mi) is the probability
of agreement by chance. The K coefﬁcient is 1 if there is a total agreement among
the annotators, y 0 if there is no agreement other than that expected to occur by
chance.

The kappa values along with percentage agreements obtained on each corpus are
mostrado en la tabla 3. We also computed the number of instances that were tagged with
OTHER by both annotators for each semantic relation and preposition paraphrase, encima
the number of examples classiﬁed in that category by at least one of the judges. Para
the instances that encoded more than one classiﬁcation category, the agreement was
measured on the ﬁrst relation on which the annotators agreed.

The agreement obtained for the Europarl corpus is higher than that for CLUVI
on both classiﬁcation sets. En general, the K coefﬁcient shows a fair to good level of
agreement for the corpus data on the set of 22 relaciones, with a higher agreement for the
preposition paraphrases. Sin embargo, according to Artstein (2007), kappa values can drop
signiﬁcantly if the frequency distribution of the annotation categories in the text corpus
is skewed. This is the case here, as will be shown in the next section. De este modo, for a better
understanding of the annotation results we also computed the percentage agreement,
which is indicated for each classiﬁcation set in parentheses in Table 3.

7.8% of Europarl and 5.7% of CLUVI instances that could not be tagged with
Lauer’s prepositions were included in the OTHER-PP category. A partir de estos, 2.1% y
2.3%, respectivamente, could be paraphrased with prepositions other than those considered
by Lauer (p.ej., bus service: service by bus), y 5.7% y 3.4%, respectivamente, could not be
paraphrased with prepositions (p.ej., daisy ﬂower).

In the next section we discuss the distribution of the syntactic and semantic inter-

pretation categories on the two different cross-linguistic corpora.

Mesa 3
The inter-annotator agreement on the annotation of the nominal phrases and compounds in the
two corpora. For the instances that encoded more than one classiﬁcation category, the agreement
was measured on the ﬁrst relation on which the annotators agreed. N/A = not applicable.

Kappa Agreement
(% agreement)

Cuerpo

Classiﬁcation tag sets

N N

N P N

OTHER

Europarl

CLUVI

8 PPs
22 SRs

0.80 (85.4%)
0.61 (76.1%)

N/A
0.67 (80.8%)

0.77 (84.7%)
0.56 (73.8%)

N/A
0.58 (75.1%)

91%
78%

86%
69%

197

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

3.3 Distribution of Syntactic Constructions and Semantic Relations

A. Cross-linguistic distribution and mapping of nominal phrases and compounds

Mesa 4 shows the distribution of various syntactic constructions used for the
translation of the 6,200 (3,076 N N and 3,124 N P N) Europarl and 2,310 (600 N N
y 1,710 N P N) CLUVI English token instances in each of the ﬁve target languages
consideró. The data show that N N and N P N constructions cover over 83% del
translation patterns for both text corpora. Sin embargo, whereas the distribution of both
constructions is balanced in the Europarl corpus (acerca de 45%, with the exception of
Romanian for which N P N constructions are less frequent), in CLUVI the N P N
constructions occur in more than 85% of the cases (de nuevo, with the exception of Ro-
manian where they represent about 56% of the data). The high percentage obtained for
N P N instances in CLUVI is explained by the fact that Romance languages have very
few N N compounds which are of limited semantic types, such as TYPE. Además, él
is interesting to note here that some of the English instances are translated into both
noun–noun (N N) and noun–adjective (N A) compounds in the target languages. Para
ejemplo, love affair translates into either the N A construction enredo amoroso (Español),
aventure amoureuse (Francés), relazione amorosa (italiano), rela¸cao amorosa (Portuguese),
and aventur˘a amoroasˇa (Romanian), or using the more common N de N pattern aventura
de amor (Español), aventure d’amour (Francés), storia d’amore (italiano), estoria de amor
(Portuguese), and aventur˘a de dragoste (Romanian). There are also instances which
translate as one word in the target language, mostrado en la tabla 4, columna 6. Por ejemplo,

Mesa 4
The distribution of syntactic constructions used in the translation of 6,200 Europarl and 2,310
English NN and N P N instances. N A = noun–adjective; pph = other syntactic paraphrase.

Cuerpo

Idioma

N N

N P N

N A

palabra

pph

Total

Syntactic distribution

2,747
(44.31%)
2,896
(46.71%)
2,896
(46.71%)
2,858
(46.1%)
4,010
(64.68%)

32
(1.39%)
25
(1.08%)
25
(1.08%)
25
(1.08%)
758
(32.81%)

2,896
(46.71%)
2,413
(38.92%)
2,487
(40.12%)
2,301
(37.11%)
1,596
(25.74%)

1,967
(85.15%)
2,046
(88.57%)
1,959
(84.81%)
1,990
(86.15%)
1,295
(56.06%)

372
(5.99%)
520
(8.38%)
483
(7.79%)
594
(9.58%)
297
(4.79%)

94
(4.07%)
75
(3.25%)
107
(4.63%)
163
(7.05%)
88
(3.81%)

37
(0.6%)
111
(1.8%)
36
(0.58%)
75
(1.21%)
74
(1.19%)

154
(6.66%)
113
(4.89%)
163
(7.06%)
88
(3.81%)
125
(5.41%)

148
(2.39%)
260
(4.19%)
298
(4.80%)
372
(6%)
223
(3.6%)

63
(2.73%)
51
(2.21%)
56
(2.42%)
44
(1.91%)
44
(1.91%)

6,200

2,310

Francés

italiano

Europarl

Español

Portuguese

Romanian

Francés

italiano

CLUVI

Español

Portuguese

Romanian

198

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

ankle boot is translated into bottine in French and stivaletto in Italian. The rest of the data
is encoded by other syntactic paraphrases, as shown in Table 4, columna 7. Por ejemplo,
bomb site is translated into Italian as luogo dove `e esplosa la bomba (‘the place where
the bomb has exploded’). Además, Mesa 5 shows the distribution of the prepositions
present in the N P N translations.

Mesa 5
The distribution of N P N constructions used in the translation of the English noun phrase
instances on both text corpora. The preposition a is used to denote a, anuncio, and de to denote simple
and articulated prepositions (de, di, du, de la, della, degli, d’, etc.).

Cuerpo

Idioma

N P N distribution

Inglés

Francés

italiano

Europarl

Español

Portuguese

Romanian

Inglés

Francés

italiano

CLUVI

Español

Portuguese

Romanian

de (81.15%); para (3.27%); en (4.61%); en (2.43%);
en (1.22%); de (0.67%); con (2.85%);
por (1.5%); against (0.42%); a través de (0.29%);
bajo (0.42%); después (0.38%); antes (0.85%)
de (75.69%); `a (2.93%); pour (6.42%); par (1.42%);
en (1.62%); avec (1.6%) ; devant (1.6%);
apr`es (1.21%); dans (2.11%); sur (2.6%);
contre (0.4%); avant (0.4%)
de (71.78%); a (7%); su (1.29%); a (3.11%);
da (6.59%); por (6.22%); a través de (0.79%); en (0.79%);
estafa (1.41%); contra (0.62%); davanti (0.2%);
dopo (0.2%)
de (83.39%); a (1.81%); en (1.41%); para (3.5%);
por (2.61%); estafa (3.18%); sobre (3.3%);
contra (0.4%); en materia de (0.4%)
de (78.4%); a (0.8%); em (0.8%); para (3.5%);
por (1.6%); com (0.8%); sobre (1.3%);
antes de (0.4%)
de (82.2%); ˆınainte de (1.82%); cu (1.82%); pentru (4.51%);
despre (1.63%); la (0.38%); datorit˘a (0.38%);
pe (6.08%); pe calea (0.37%); ˆın (0.81%)

de (83.80%); para (1.17%); en (5.90%); en (2.40%);
en (0.76%); con (1.99%); against (1.17%);
a través de (0.41%); encima (0.41%); arriba (0.41%);
beside (0.41%); acerca de (0.41%); behind (0.76%)
de (82.33%); `a (6.2%); pour (1.42%); en (1.8%);
sur (7.02%); contre (0.41%); pr`es de (0.41%);
`a cot´e de (0.41%)
de (75.42%); a (8.07%); su (1.32%);
da (6.6%); por (6.21%); en (0.78%); estafa (0.4%);
contra (0.4%); sopra (0.2%); accanto a (0.2%);
dietro de (0.2%); a través de (0.2%)
de (85.96%); a (2.81%); en (3.89%); para (0.71%);
por (1.74%); estafa (2.1%); sobre (1.38%);
contra (0.36%); detr´as (0.71%); encima (0.36%)
de (78.4%); a (0.8%); em (0.82%); para (3.5%);
por (1.6%); com (0.8%); sobre (1.3%);
acima de (0.4%)
de (85.21%); cu (1.82%); pentru (4.5%);
la (0.4%); datorit˘a (0.4%); pe (5.08%);
despre (1.58%); ˆın (0.79%); lˆanga ( 0.2%)

Total

3,124

2,896

2,413

2,487

2,301

1,596

1,710

1,967

2,046

1,959

1,990

1,295

199

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

For the purposes of this research, desde el 6,200 Europarl and 2,310 CLUVI
instancias, we selected those which had all the translations encoded only by N N and
N P N constructions. columnas 3 y 4 en mesa 4 show the number of N N and N P N
translation instances in each Romance language. Out of these, we considered only
3,124 Europarl and 2,023 CLUVI token instances representing the examples encoded
by N N and N P N in all languages considered, after inter-annotator agreement.

B. Cross-linguistic distribution of semantic relations and their mapping to nominal
phrases and compounds

A closer look at the N N and N P N translation instances in Table 4 shows that
their syntactic distribution is inﬂuenced by the text genre and the semantics of the
instancias. Por ejemplo, in Europarl most of the N N instances were naming noun–
noun compounds referring to entities such as member states and framework law which
were repeated in many sentences. Many of them encoded TYPE relations (p.ej., member
estado, framework law) cual, most of the time, are encoded by N N patterns in the target
idiomas (stato membro and legge quadro in Italian, respectivamente). In the CLUVI corpus,
por otro lado, the N N Romance translations represented only 1% of the data. A
notable exception here is Romanian (64.68% of Europarl and 32.8% of CLUVI). Esto es
explained by the fact that, in Romanian, many noun phrases are represented as genitive-
marked noun compounds (N1 N2). In Romanian the genitive case is realized either as
a sufﬁx attached to the modiﬁer noun N2 or as one of the genitival articles a/al/ale. Si
the modiﬁer noun N2 is determined by an indeﬁnite article then the genitive mark is
applied to the article, not to the noun, for example o fat˘a – unei fete (‘a girl – of/to a
girl’) and un b˘aiat – unui b˘aiat (‘a boy – of/to a boy’). Similarmente, if the modiﬁer noun is
determined by the deﬁnite article (which is enclitic in Romanian), the genitive mark is
added at the end of the noun together with the article. Por ejemplo, fata–fetei (the girl –
girl-GEN), cartea–c˘art¸ii (the book – book-GEN). De este modo, the noun phrase the beauty of the
girl, por ejemplo, is translated as frumuset¸ea fetei (‘beauty-the girl-GEN’), and the beauty of
a girl as frumuset¸ea unei fete (‘beauty-the of/to a girl’).

En general, in Romanian the choice between the N de N and the genitive-marked
N N constructions depends on the speciﬁcity of the instance. Some noun–noun instances
refer to a speciﬁc entity (existential interpretation), in which case the construction
preferred is the genitive-marked N N, or they can refer in general to the category of
those entities (generic interpretation),11 thus using N de N. Por ejemplo, the instance
the bite of the scorpion (AGENT) translates into mu¸sc˘atura scorpionului (‘bite-the scorpion-
GEN’), whereas a scorpion bite (AGENT) translates into mu¸sc˘atur˘a de scorpion (‘bite of
scorpion’).

Many semantic relations that allow both the generic and the existential interpre-
tations can be encoded by both N P N and genitive-marked N N constructions as
shown by the example above. Sin embargo, there are situations when the generic and
the existential interpretations change the meaning of the noun–noun pair. One such
example is the suit of the sailor (POSSESSION) translated as costumul marinarului (‘suit-
the sailor-GEN’), and sailor suit (PURPOSE) translated as costum de marinar (‘suit of
sailor’).

11 The words existential and generic are borrowed here from the vast linguistic literature on deﬁnite and
indeﬁnite descriptions. Aquí, nouns such as ﬁremen can have different readings in various contexts:
Firemen are available (existential reading), vs. Firemen are altruistic (generic reading).

200

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

At the other extreme there are relations which prefer either the generic or the
existential interpretation. Por ejemplo, some POSSESSION-encoding instances such as
the budget of the University translate as ‘bugetul Universit˘at¸ii’ (budget-the University-GEN)
and not as ‘bugetul de Universitate’ (budget-the of University). Other relations such as
PURPOSE and SOURCE identify generic instances. Por ejemplo, (a) olive oil (FUENTE)
translates as ‘ulei de m˘asline’ (oil of olive), and not as ‘uleiul m˘aslinei’ (oil-the olive-
GEN), y (b) the milk glass (PURPOSE) translates as ‘paharul de lapte’ (glass-the of milk)
and not as ‘paharul laptelui’ (glass-the milk-GEN). Other examples include CAUSE and
TOPIC. This observation is very valuable for the interpretation of nominal phrases and
compounds and is used in the learning model to discriminate among the possible
interpretaciones.

Tables 6 y 7 show the semantic distribution of the instances on both text corpora.
This distribution is represented both in number of tokens (the total number of instances
per relation) and types (the unique number of instances per relation). In Europarl,
the most frequently occurring relations are TYPE and THEME that together represent
acerca de 50% of the data with an equal distribution. The next most frequent relations
are TOPIC, PURPOSE, AGENT, and PROPERTY with an average coverage of about 8%.
Además, eight relations of the 22-SR set (KINSHIP, DEPICTION, CAUSE, INSTRUMENT,
FUENTE, MANNER, MEASURE, and BENEFICIARY) did not occur in this corpus. El
9.61% of the OTHER-SR relation represents the ratio of those instances that did not
encode any of the 22 relaciones semánticas. It is interesting to note here the large difference
between the number of types versus tokens for the TYPE relation in Europarl. Esto es
accounted for by various N N instances such as member states that repeat across the
cuerpo.

This semantic distribution contrasts with the one in CLUVI. Aquí, the most fre-
quent relation by far is PART–WHOLE (40.53%), followed by LOCATION (8.95%), AGENT
(6.23%), and IS-A (5.93%). The missing relations are KINSHIP, MANNER and BENEFI-
CIARY. A larger percentage of OTHER-SR instances (12.95%) did not encode any of the
22 relaciones semánticas. Además, in CLUVI 256 instances were tagged with more than
one semantic relation with the following distribution: 46.8% MEASURE/PART–WHOLE
(p.ej., a couple of cigarettes), 28.2% PART–WHOLE/LOCATION (p.ej., bottom of the sea), 10.9%
MEASURE/LOCATION (p.ej., cup of chocolate), 8.2% PURPOSE/LOCATION (p.ej., waste gar-
den), y 5.9% THEME/MAKE-PRODUCE (p.ej., makers of songs). In Europarl, en el otro
mano, there were only 97 such cases: 81.4% THEME/MAKE-PRODUCE (p.ej., bus manufac-
turers) y 18.6% MEASURE/PART–WHOLE (p.ej., number of states).

One way to study the contribution of both the English and Romance prepositions
to the interpretation task is to look at their distribution over the set of semantic relations
on two reasonably large text corpora of different genres. Por supuesto, this approach does
not provide an analysis that generates an exhaustive generalization over the properties
of the language. Sin embargo, as Tables 6 y 7 espectáculo, there are dependencies between the
structure of the Romance language translations and the semantic relations encoded by
the nominal phrases and compounds, although the most frequently occurring preposi-
tions are de and its English equivalent of. Here we use the preposition de to represent a
set of translation equivalents in Romance languages (p.ej., the Italian counterpart is di).
These prepositions are semantically underspeciﬁed, encoding a large set of semantic
relaciones. The many-to-many mappings of the prepositions to the semantic classes
adds to the complexity of the interpretation task. Por ejemplo, in the Europarl corpus
LOCATION is encoded in French by de, sur, devant, and `a pr`es de, while TOPIC is encoded
in English by of, para, en, about and noun compounds, and in Spanish by de, sobre, en
materia de.

201

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

Mesa 6
Mapping between the set of 22 semantic classiﬁcation categories and the set of English and
Romance syntactic constructions on the Europarl corpus. The preposition de is used here to
denote simple and articulated prepositions (de, di, du, de la, della, degli, d’, etc.). También, the dash “–”
refers to noun–noun compounds where there is no connecting preposition. The mapping was
obtained on the 3,124 Europarl instance corpus. En. = English; Sp. = Spanish; Él. = Italian;
Fr. = French; Port. = Portuguese; Ro. = Romanian.

Nr.
1

SRs

POSSESSION

En.

de, –

Sp.

de, –

Él.
de, –

Fr.

Port.

Total

Token
[%]
2.85

0
6.05

Tipo
[%]
2.4

0
6.05

Ro.

de, –

7.47

7.08

de,
avant

de
acima de

de,
ˆınainte de

0.04

de, para,
en, –

de, para,
en, por, –

de, en,
en, en, –

de, estafa
a

de,–

2
3

6
7

9
10

11
12

14
15

16
17

19
20

22
23

KINSHIP
PROPERTY

AGENT

TEMPORAL

DEPICTION
PART–WHOLE

CAUSE
MAKE/
PRODUCE

INSTRUMENT
LOCATION

PURPOSE

FUENTE
TOPIC

MANNER
MEANS

EXPERIENCER

MEASURE
TYPE

THEME

BENEFICIARY
OTHER–SR

de, en,
con, –

de, estafa

de, a,
a, –

de, `a

IS–A
(HYPERNYMY)

de, –
con

de, –

de, para, –
en, de

de, en,
en, –
en
de, –
para

de, para,
en, –
acerca de

por
de, –

de,–
en
en

–

de, para,
en, –

de, en,
sobre

de, por,
para,
contra

de, sobre,
en materia
de

por, en,

–

de, su,
a, en

de, da,
por, a,
–

de, a,
su

de, sur,
`a, pr`es de,
devant
contre, `a,
de, –
pour

por, en,
a, a través de

en, `a,
par

–

de, a

de,

–

de, a

de,
sobre

por

–

0
3.20

0
2.75

0.8

0
1.43

0
2.14

0
1.43

0
2.14

7.48

7.23

0
11.03

0
0.07

de
estafa

–

de, pe,
la, ˆın

de,
pentru

de,
despre

pe, cu,
pe calea

de, –

0.04

–

0
24.47

0
1.7

23.13

19.2

0
9.61

0
8.13

3,124

2,190

de, por

a, de

de, `a

de, a,
com

de,
pentru

Total no. of examples

Ejemplo

Union resources
‘resursele uniunii’ (Ro.)
(resource-the union-GEN)

trafﬁc density
‘densit`a del trafﬁco’ (It.)
(density of trafﬁc)
request of a member
‘richiesta di uno membro’ (It.)
(request of a member)
year before the constitution
‘a ˜no anterior a la constituci´on’ (Sp.)
(the year previous of the)

constitution

Union citizen
‘citoyen de l’ Union’ (Fr.)
(citizen of the Union)
process of decay
‘proces de descompunere’ (Ro.)
(process of decay)

paper plant
‘f´abrica de papel’ (Sp.)
(plant of paper)

place of the meeting
‘lieu de la r´eunion’ (Fr.)
(place of the meeting)
building stone
‘pedras de constru¸c˜ao’ (Port.)
(stones of building)

policy on asylum
‘pol´ıtica en materia de asilo’ (Sp.)
(policy in regard to asylum)

travel by train
‘calatorie cu tenul’ (Ro.)
(travel with train-the)
suffering of the people
‘sofrimento das pessoas’ (Port.)
(suffering of the people)

framework law
‘legge quadro’ (It.)
(law framework)
conﬂict prevention
‘prevenire de conﬂict’ (Ro.)
(prevention of conﬂict)

tobacco addiction
‘adicci´on a tabaco’ (Sp.)
(addiction to tobacco)

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Además, in the Europarl corpus, 31.64% of the instances are synthetic phrases en-
coding AGENT, MEANS, LOCATION, THEME, and EXPERIENCER. Out of these instances,
98.7% use the preposition of and its Romance equivalent de. In the CLUVI corpus,
14.1% of the examples were verbal, from which the preposition of/de has a coverage
de 77.66%.

Based on the literature on prepositions (Lyons 1986; Pregonero 1998;

Ionin,
Matushansky, and Ruys 2006) and our own observations, the preposition of/de in both
root and synthetic nominal phrases may have a functional or a semantic role, acting
as a linking device with no apparent semantic content, or with a meaning of its own.
De este modo, for the interpretation of these constructions a system must rely on the meaning of
preposition and the meaning of the two constituent nouns in particular, and on context

202

Girju

The Syntax and Semantics of Prepositions

Mesa 7
Mapping between the set of 22 semantic classiﬁcation categories and the set of English and
Romance syntactic constructions on the CLUVI corpus. The preposition de is used here to denote
simple and articulated prepositions (de, di, du, de la, della, degli, d’, etc.). También, the dash “–” refers
to noun–noun compounds where there is no connecting preposition. The mapping was obtained
sobre el 2,023 CLUVI instance corpus. En. = English; Sp. = Spanish; Él. = Italian; Fr. = French;
Port. = Portuguese; Ro. = Romanian.

Nr.
1

SRs

POSSESSION

En.

de, –

Sp.
de, –

Él.

de, –

Fr.

Port.

Ro.

de, –

Total

Token
[%]
1.35

0
2.97

Tipo
[%]
1.21

0
2.76

de, –

6.23

5.78

de, –

2.97

0.3

40.53

34.35

5.93

5.4

de,
datorit˘a

2.72

0.29

de, cu

0.29

de, `a

de, –

de, `a

KINSHIP
PROPERTY

AGENT

TEMPORAL

DEPICTION–
DEPICTED

PART–WHOLE

de, para,
en, –

de, para,
en, por, –

de, en,
en, en, –

de, –

de, estafa

de, en,
con, –

IS–A
(HYPERNYMY)

de, –
con

CAUSE

de, –

MAKE/
PRODUCE

de, para,
en, de, –

de, estafa

de, –

INSTRUMENT

para, con

de, –

LOCATION

de, en,
en, en, –

de, en,
sobre,

PURPOSE

de, –
para

FUENTE

de, de

de, por,
para,
contra
de

de, a,
–

de, –

de, da

de, a,
estafa

de, su,
a, en,
dietro de,
accanto a,
sopra
de, da,
por, a, –
contra
de

de, sur,
`a, pr`es de,
`a cot´e de

de, em
acima de

de, pe, la,
ˆın, lˆanga

8.65

8.01

contre, a,
de, –
pour
de

de,
pentru

4.45

0.94

0.15

TOPIC

MANNER
MEANS

de, para, en,
acerca de, –

de,
sobre

de, a,
su

de, por

por

a través de

EXPERIENCER

de, en, –

MEASURE

por

TYPE
THEME

BENEFICIARY
OTHER–SR

para, –
de, en

de, por

de, a

Total no. of examples

de,
sobre

de,
despre

0.79

por

0
0.15

de, –

0.64

de,
pentru

de, a, –

3.81

2.72

0
4.05

0
12.95

0
3.94

0
8.81

2,023

1,734

de, a,

–

de
com

–

2
3

16
17

20
21

22
23

Ejemplo

police car
‘coche de polizia’ (Sp.)
(car of police)

beauty of the buildings
‘belleza de los ediﬁcios’ (Sp.)
(beauty of the buildings)
return of the family
‘regresso da fam´ılia’ (Port.)
(return of the family)
spring rain
‘pluie de printemps’ (Fr.)
(rain of spring)
picture of a girl
‘retrato de uma rapariga’ (Port.)
(picture of a girl)
ruins of granite
‘ruinas de granito’ (Sp.)
(ruins of granite)
sensation of fear
‘sensa¸c´ao de medo’ (Port.)
(sensation of fear)
cries of delight
‘cri de joie’ (Fr.)
(cries of delight)
noise of the machinery
‘ruido de la maquinaria’ (Sp.)
(noise of the machinery)
a ﬁnger scratch
‘o zgˆarietur˘a de unghie’ (Ro.)
(a scratch of ﬁnger)
book on the table
‘livre sur la table’ (Fr.)
(book on the table)

nail brush
‘spazzolino per le unghie’ (It.)
(brush for the nails)
oil of cloves
‘´oleo de cravinho’ (Port.)
(oil of cloves)
love story
‘histoire d’amour’ (Fr.)
(story of love)

travel by car
‘c˘al˘atorie cu ma¸sina’ (Ro.)
(travel by car)
the agony of the prisoners
‘l’agonia dei prigionieri’ (It.)
(the agony of.the prisoners)
a cup of sugar
‘o cea¸sc´a de zah˘ar’ (Ro.)
(a cup of sugar)

lack of intelligence
‘manque d’intelligence’ (Fr.)
(lack of intelligence)

cry of death
‘cri de mort’ (Fr.)
(cry of death)

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

en general. Because the two corpora used in this paper contain both root and synthetic
instancias, we employed two semantic resources for this task: WordNet noun semantic
classes and a collection of verb classes in English that correspond to special types of
nominalizations. These resources are deﬁned in Section 4.2. Además, en la sección 6
we present a detailed linguistic analysis of the prepositions of in English and de in
Romance languages, and show how their selection correlates with the meaning of the
construction.

203

Ligüística computacional

Volumen 35, Número 2

4. Modelo

4.1 Mathematical Formulation

Given the syntactic constructions considered, the goal is to develop a procedure for
the automatic annotation of the semantic relations they encode. The semantic relations
derive from various lexical and semantic features of each instance.

The semantic classiﬁcation of instances of nominal phrases and compounds can be
formulated as a learning problem, and thus beneﬁts from the theoretical foundation
and experience gained with various learning paradigms. The task is a multi-class clas-
siﬁcation problem since the output can be one of the semantic relations in the set. Nosotros
cast this as a supervised learning problem where input/output pairs are available as
training data.

An important ﬁrst step is to map the characteristics of each instance (es decir., list of
properties that describe the instance, usually not numerical) into feature vectors. Let us
deﬁne xi as the feature vector of an instance i and let X be the space of all instances; eso
es, xi ∈ X.

The multi-class classiﬁcation is performed by a function that maps the feature space
X into a semantic space S, F : X → S, where S is the set of semantic relations from Table 1,
a saber, rj ∈ S, where rj is a semantic relation.

Let T be the training set of examples or instances T = (x1r1 .. xlrl) ⊆ (X x S)l where
l is the number of examples x each accompanied by its semantic relation label r. El
problem is to decide which semantic relation to assign to a new, unseen example xl+1.
In order to classify a given set of examples (members of X), one needs some kind of
measure of the similarity (or the difference) between any two given members of X.

De este modo, the system receives as input an English nominal phrase and compound
instances along with their translations in the Romance languages, plus a set of extra-
linguistic features. The output is a set of learning rules that classify the data based on
the set of 22 semantic target categories. The learning procedure is supervised and takes
into consideration the cross-linguistic lexico-syntactic information gathered for each
instancia.

4.2 Feature Space

The set of features allows a supervised machine learning algorithm to induce a function
that can be applied to accurately classify unseen instances. Based on the study of the
instances and their semantic distribution presented in Section 3, we have identiﬁed and
experimented with the following features presented subsequently for each language in-
volved. Features F1–F5 have been employed by us in our previous research (Moldovan
et al. 2004; Girju et al. 2005; Girju, Badulescu, and Moldovan 2006). All the other features
are novel.

A. English features

F1 and F2. Semantic class of noun speciﬁes the WordNet sense of the head noun (F1), y
the modiﬁer noun (F2) and implicitly points to all its hypernyms. The semantics of the
instances of nominal phrases and compounds is heavily inﬂuenced by the meaning of
the noun constituents. One such example is family#2 car#1, which encodes a POSSESSION
relation. The hypernyms of the head noun car#1 are: {motor vehicle}, {self-propelled

204

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

vehicle} … {entidad} (cf. WordNet 2.1). These features will help generalize over the se-
mantic classes of the two nouns in the instance corpus.

F3 and F4. WordNet derivationally related form speciﬁes if the head noun (F3), y el
modiﬁer noun (F4) are related to a corresponding verb in WordNet. WordNet contains
information about nouns derived from verbs (p.ej., statement derived from to state; cry
from to cry; death from to die).

F5. Prepositional cues link the two nouns in a nominal phrase. These can be either simple
or complex prepositions such as of or according to. In case of N N instances (p.ej., member
estado), this feature is “–”.

F6 and F7. Type of nominalized noun indicates the speciﬁc class of nouns the head (F6) o
modiﬁer (F7) belongs to depending on the verb from which it derives. Primero, we check if
the noun is a nominalization or not. For English we used the NomLex-Plus dictionary of
nominalizations (Meyers at al. 2004) to map nouns to corresponding verbs.12 One such
example is the destruction of the city, where destruction is a nominalization. F6 and F7 may
overlap with features F3 and F4 which are used in case the noun to be checked has no
entry in the NomLex-Plus dictionary.

These features are of particular importance because they impose some constraints
on the possible set of relations the instance can encode. They take the following values:
a) active form nouns, b) unaccusative nouns, C) unergative nouns, and d) inherently
passive nouns. We present them in more detail subsequently.

a. Active form nouns are derived through nominalization from psych verbs and rep-
resent states of emotion, such as love, miedo, desire, Etcétera. They have an intrinsic
active voice predicate–argument structure and, de este modo, resist passivisation. Por ejemplo,
we can say the desire of Anna, but not the desire by Anna. This is also explained by
the fact that in English the AGENT or EXPERIENCER relations are mostly expressed
by the clitic genitive ‘s (p.ej., Anna’s desire) and less or never by N P N constructions.
Citing Anderson (1983), Giorgi and Longobardi (1991) mention that with such nouns
that resist passivisation, the preposition introducing the internal argument, even if
it is of, has always a semantic content, and is not a bare case-marker realizing the
genitive case. Además, they argue that the meaning of these nouns might pattern
differently in different languages. Consider for example the Italian sentences (4) y (5)
below and their English equivalents (see Giorgi and Longobardi 1991, pages 121–
122). In English the instance Anna’s desire identiﬁes the subject of desire (y por lo tanto
encodes an EXPERIENCER relation), whereas in Italian it can identify either the subject
(EXPERIENCER) as in Example (4), or the object of desire (THEME) as in Example (5),
the disambiguation being done at the discourse level. In Example (6) the prenominal
construction il suo desiderio encodes only EXPERIENCER.

(4)

(5)

Il desiderio di Anna fu esaudito. (EXPERIENCER)
(The desire of Anna was fulﬁlled.)
‘Anna’s desire was fulﬁlled.’

Il desiderio di Anna lo porter`a alla rovina. (THEME)
(The desire of Anna him will ruin.)
‘The desire for Anna will ruin him.’

12 NomLex-Plus is a hand-coded database of 5,000 verb nominalizations, de-adjectival, and de-adverbial
nouns including the corresponding subcategorization frames (verb-argument structure information).

205

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

(6)

Il suo desiderio fu esaudito. (EXPERIENCER)
(The her desire was fulﬁlled.)
‘Her desire was fulﬁlled.’

Sin embargo, our observations on the Romanian training instances in Europarl and CLUVI
(captured by features F12 and F13 below) indicate that the choice of syntactic construc-
tions can help in the disambiguation of instances that include such active nouns. De este modo,
whereas genitive-marked N N compounds identify only the subject (thus encoding
EXPERIENCER), the N de/pentru N constructions identify only the object (thus encoding
THEME). Such examples are dorint¸a Anei (‘desire-the Anna-GEN’ – Anna’s desire) (EX-
PERIENCER) and dorint¸a de/pentru Ana (‘desire-the of/for Anna’ – the desire for Anna)
(THEME).

Another example is the love of children and not the love by the children, where children
are the recipients of love, not its experiencers. In Italian the instance translates as l’amore
per i bambini (‘the love for the children’), whereas in Romanian it translates as dragostea
pentru copii (‘love-the for children’). These nouns mark their internal argument through
of in English and most of the time require prepositions such as for in Romance languages
y viceversa.

b. Unaccusative nouns are derived from ergative verbs that take only internal ar-
guments (p.ej., those that indicate an object and not a subject grammatical role). Para
ejemplo, the transitive verb to disband allows the subject to be deleted as in the following
oraciones:

(7)

(8)

The lead singer disbanded the group in 1991.

The group disbanded.

De este modo, the corresponding unaccusative nominalization of to disband, the disbandment of
el grupo, encodes THEME and not AGENT.

C. Unergative nouns are derived from intransitive verbs. They can take only AGENT
relaciones semánticas. One such case is exempliﬁed in the instance l’arrivo della cavalleria in
Italian which translates in English as the arrival of the cavalry and in Romanian as sorirea
cavaleriei (‘arrival-the cavalry-GEN’).

d. Inherently passive nouns. These nouns, like the verbs they are derived from, assume
an implicit AGENT relation and, being transitive, associate to their internal argument
the THEME relation. One such example is the capture of the soldier which translates in
Italian as la cattura del soldato (‘the capture of the soldier’), la capture du soldat in French
(‘the capture of soldier’), and la captura de soldado in Spanish and Portuguese (‘the
capture of soldier’), where the nominalization capture (cattura, captura, captura in Italian,
Francés, and Spanish and Portuguese respectively) is derived from the verb to capture.
Aquí, whereas English and Italian, Español, Portuguese, and French use the N of/de N
construction (as shown in Examples (9) y (10) for English and Italian), Romanian
uses genitive-marked noun compounds. In Romanian, sin embargo, nominalizations are
formed through sufﬁxation, where a sufﬁx is added to the root of the verb it comes from.
Different sufﬁxes attached to the same verb may lead, sin embargo, to more than one nom-
inalization, producing different meanings. The verb to capture (a captura in Romanian),
Por ejemplo, can result through sufﬁxation in two nominalizations: capturare (con el

206

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

inﬁnitive sufﬁx -are and encoding an implicit AGENT relation) and captur˘a (a través de
zero derivation and encoding an implicit THEME relation) (Cornilescu 2001). De este modo, el
noun phrase capturarea soldatului (‘capture-the soldier-GEN’) encodes a THEME relation,
while captura soldatului (‘capture-the soldier-GEN’) encodes an AGENT relation. In all the
Romance languages with the exception of Romanian, this construction is ambiguous,
unless the AGENT is explicitly stated or inferred as shown in Example (9) for Italian.
The same ambiguity might occur sometimes in English, with the difference that besides
the of-genitive, English also uses the s-genitive: the soldier’s capture (AGENT is preferred
if the context doesn’t mention otherwise), the soldier’s capture by the enemy (THEME), el
capture of the soldier (THEME is preferred if the context doesn’t mention otherwise), el
capture of the soldier by the enemy (THEME).

(9)

La cattura del soldato (da parte del nemigo) `e cominciata come un atto terroristico.
(THEME)
‘The capture of the soldier (by the enemy) has started as a terrorist act.’

(10)

La sua cattura `e cominciata come un atto terroristico. (THEME)
‘His capture has started as a terrorist act.’

These nouns have a different behavior than that of active form nouns. As shown
previamente, the object of inherently passive nouns can move to the subject position as
in the soldier’s capture by the enemy, whereas it cannot do so for active form nouns (p.ej.,
*Anna’s desire by John). Similarmente, in Italian, although active form nouns allow only the
subject reading in prenominal constructions (p.ej., il suo desiderio – ‘her desire’), inher-
ently passive nouns allow only the object reading (p.ej., la sua cattura – ‘his capture’).

For Romanian, the nominalization sufﬁxes were identiﬁed based on the morpho-

logical patterns presented in Cornilescu (2001).

We assembled a list of about 3,000 nouns that belong to classes a–d using the infor-
mation on subcategorization frames and thematic roles of the verbs in VerbNet (Kipper,
Dang, and Palmer 2000). VerbNet is a database which encodes rich lexical information
for a large number of English verbs in the form of subcategorization information,
selectional restrictions, thematic roles for each argument of the verb, and alternations
(the syntactic constructions in which the verb participates).

B. Romance features

F8, F9, F10, F11, and F12. Prepositional cues that link the two nouns are extracted from
each translation of the English instance: F8 (Sp.), F9 (Fr.), F10 (It.), F11 (Port.), and F12
(Ro.). These can be either simple or complex prepositions (p.ej., de, in materia de [Sp.]) en
all ﬁve Romance languages, or the Romanian genitival article a/al/ale. For N N instances,
this feature is “–”.

F13. Noun inﬂection is deﬁned only for Romanian and shows if the modiﬁer noun in N N
instances is not inﬂected or is inﬂected and modiﬁes the head noun which is or is not
a nominalization. This feature is used to help differentiate between instances encoded
by genitive-marked N N constructions and noun–noun compounds, when the choice
of syntactic construction reﬂects different semantic content. Two such examples are the
noun–noun compound lege cadru (law framework) (TYPE) which translates as framework

207

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

law and the genitive-marked N N instance frumuset¸ea fetei (‘beauty-the girl-GEN’) (PROP-
ERTY) meaning the beauty of the girl. It also covers examples such as capturarea soldatului
(‘capture-the soldier-GEN’), where the modiﬁer soldatului is inﬂected and the head noun
capturarea is a nominalization derived through inﬁnitive sufﬁxation.

In the following Example we present the feature vector for the instance the capture

of the soldiers.

(11)

The instance the capture of the soldiers has the following Romance translations:

(cid:2)capture#4/Arg2 of soldiers#1/Arg1; captura de soldados; capture du soldats; cattura dei soldati;
captura dos soldados; capturarea soldat¸ilor; THEME(cid:3).
Its corresponding feature vector is:

(cid:2)entity#1/Arg2; entity#1/Arg1; captura; –; de; inherently passive noun; –; de; de; de; de; –;

mod-inﬂected-inf-nom; THEME(cid:3),

where mod-inﬂected-inf-nom indicates that the noun modiﬁer soldat¸ilor in the Ro-
manian translation capturarea soldat¸ilor (‘capture-the soldiers-GEN’) is inﬂected and that
the head noun capturarea is an inﬁnitive nominalization.

4.3 Learning Models

Several learning models can be used to provide the discriminating function f. Nosotros
have experimented with the support vector machines model and compared the results
against two state-of-the-art models: semantic scattering, a supervised model described
in Moldovan et al. (2004), Girju et al. (2005), and Moldovan and Badulescu (2005), y
Lapata and Keller’s Web-based unsupervised model (Lapata and Keller 2005).

Each model was trained and tested on the Europarl and CLUVI corpora using a
7:3 training–testing ratio. All the test nouns were tagged with the corresponding sense
in context using a state-of-the-art WSD tool (Mihalcea and Faruque 2004). The default
semantic argument frame for each relation was used in the automatic identiﬁcation of
the argument positions.

A. Support vector machines
Support vector machines (SVMs) are a set of related supervised learning methods used
for creating a learning function from a set of labeled training instances. La función
can be either a classiﬁcation function, where the output is binary (is the instance of
category X?), or it can be a general regression function. For classiﬁcation, SVMs operate
by ﬁnding a hypersurface in the space of possible inputs. This hypersurface will attempt
to split the positive examples from the negative examples. The split will be chosen
to have the largest distance from the hypersurface to the nearest of the positive and
negative examples. Intuitivamente, this makes the classiﬁcation correct for testing data that
is similar but not identical to the training data.

In order to achieve classiﬁcation in n semantic classes, n > 2, we built a binary
classiﬁer for each pair of classes (a total of C2
n classiﬁers), and then we used a voting
procedure to establish the class of a new example. For the experiments with semantic
relaciones, the simplest voting scheme has been chosen; each binary classiﬁer has one
vote, which is assigned to the class it chooses when it is run. Then the class with the
largest number of votes is considered to be the answer. The software used in these
experiments is the package LIBSVM (http://www.csie.ntu.edu.tw/∼cjlin/libsvm/)
which implements an SVM model. We tested with the radial-based kernel.

208

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

After the initial instances in the training and testing corpora were expanded with
the corresponding features, we had to prepare them for the SVM model. The set-up
procedure is now described.

Corpus set-up for the SVM model:
The processing method consists of a set of iterative procedures of specialization of
the examples on the WordNet IS-A hierarchy. De este modo, after a set of necessary specializa-
tion iterations, the method produces specialized examples which through supervised
machine learning are transformed into sets of semantic rules for the semantic interpre-
tation of nominal phrases and compounds. The specialization procedure is described
subsequently.

Initially, the training corpus consists of examples that follow the format exempliﬁed
at the end of Section 4.2 (Ejemplo [11]). Note that for the English instances, each noun
constituent was expanded with the corresponding WordNet top semantic class. At this
punto, the generalized training corpus contains two types of examples: unambiguous
and ambiguous. The second situation occurs when the training corpus classiﬁes the
same noun–noun pair into more than one semantic category. Por ejemplo, both rela-
tionships chocolate#2 cake#3 (PART–WHOLE) and chocolate#2 article#1 (TOPIC) are mapped
into the more general type (cid:2)entity#1, entity#1, PART–WHOLE/TOPIC(cid:3).13 We recursively
specialize these examples to eliminate the ambiguity. By specialization, the semantic
class is replaced with the corresponding hyponym for that particular sense, eso es,
the concept immediately below in the hierarchy. These steps are repeated until there
are no more ambiguous examples. Para este ejemplo, the specialization stops at the
ﬁrst hyponym of entity: physical entity (for cake) and abstract entity (for article). Para el
unambiguous examples in the generalized training corpus (those that are classiﬁed
with a single semantic relation), constraints are determined using cross-validation on
the SVM model.

B. Semantic scattering
The semantic scattering (SS) model was initially tested on the classiﬁcation of genitive
constructions, but it is also applicable to nominal phrases and compounds (Moldovan
et al. 2004). SS is a supervised model which, like the SVM model described previously,
relies on WordNet’s IS-A semantic hierarchy to learn a function which separates positive
and negative examples. Esencialmente, it consists of using a training data set to establish a
boundary G∗ on WordNet noun hierarchies such that each feature pair of noun–noun
senses fij on this boundary maps uniquely into one of a predeﬁned list of semantic
relaciones. The algorithm starts with the most general boundary corresponding to the
entity WordNet noun hierarchy and then specializes it based on the training data until
a good approximation is reached.14 Any feature pair above the boundary maps into
more than one semantic relation. Due to the specialization property on noun hierarchies,
feature pairs below the boundary also map into only one semantic relation. For any new
pair of noun–noun senses, the model ﬁnds the closest boundary pair which maps to one
semantic relation.

The authors deﬁne with SCm = { f m
i

} the sets of semantic class features
for modiﬁer noun and, respectivamente, head noun. A pair of nouns maps

} and SCh = { f h
j

13 The specialization procedure applies only to features 1 y 2.
14 Moldovan et al. (2004) used a list of 35 semantic relations – actually only 22 of them proved to be encoded

by nominal phrases and compounds.

209

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

(cid:3) (henceforth fij). The probability of a
uniquely into a semantic class feature pair (cid:2) f m
i
norte(r,fij )
semantic relation r given the feature pair fij, PAG(r| fij ) =
norte( fij ) is deﬁned as the ratio between
the number of occurrences of a relation r in the presence of the feature pair fij over the
number of occurrences of the feature pair fij in the corpus. The most probable semantic
relation ˆr is

, f h
j

ˆr = arg max
r∈R

PAG(r| fij) = arg máx
r∈R

PAG( fij|r)PAG(r)

(1)

From the training corpus, one can measure the quantities n(r, fij) y N( fij). Depend-

ing on the level of abstraction of fij two cases are possible:

Case 1. The feature pair fij is speciﬁc enough such that there is only one semantic

relation r for which P(r| fij) = 1 y 0 for all the other semantic relations.

Case 2. The feature pair fij is general enough such that there are at least two semantic
relations for which P(r| fij) (cid:11)= 0. In this case Equation (1) is used to ﬁnd the most appro-
priate ˆr.

Definición
A boundary G∗ in the WordNet noun hierarchies is a set of synset pairs such that:

a) for any feature pair on the boundary, denoted f G∗
ij
only one relation r, y

∈ G∗, f G∗

ij maps uniquely into

b) for any f u
ij

(cid:12) f G∗
ij

, f u

ij maps into more than one relation r, y

≺ f G∗
ij

, f l

C) for any f l
ij maps uniquely into a semantic relation r.
ij
Here relations (cid:12) and ≺ mean ‘semantically more general’ and ‘semantically more
speciﬁc’, respectivamente.
As proven by observation, there are more concept pairs under the boundary G∗
than above it, eso es, | {f l
ij

} | (cid:14) | {f u
ij

} |.

Boundary Detection Algorithm

Step 1. Create an initial boundary.
The initial boundary denoted G1 is formed from combinations of the entity#1 –
entity#1 noun class pairs. For each training example a corresponding feature fij is
ﬁrst determined, after which it is replaced with the most general corresponding
feature consisting of top WordNet hierarchy concepts. Por ejemplo, both instances
family#2 estate#2 (POSSESSION) and the sister#1 of the boy#1 (KINSHIP) are mapped into
entity#1 – entity#1. At this level, the noun–noun feature encodes a number of semantic
relaciones. For each feature, one can determine the most probable relation using Equa-
ción (1). Por ejemplo, the feature entity#1 – entity#1 can be encoded by any of the 23
relaciones.

The next step is to construct a lower boundary by specializing the semantic classes
of the ambiguous features. A feature fij is ambiguous if it corresponds to more than
one relation and its most relevant relation has a conditional probability less than 0.9.

210

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

To eliminate irrelevant specializations, the algorithm specializes only the ambiguous
classes that occur in more than 1% of the training examples.

The specialization procedure consists of ﬁrst identifying the features fij to which
correspond more than one semantic relation, then replacing these features with their
hyponym synsets. Thus one feature breaks into several new specialized features.
Por ejemplo, the feature entity#1 – entity#1 generated through generalization for
the examples family#2 estate#2 and the sister#1 of the boy#1 is specialized now as
kin group#1 – real property#1 and female sibling#1 – male person#1 corresponding to the
direct hyponyms of the nouns in these instances. The net effect is that the semantic
relations that were attached to fij will be ‘scattered’ across the new specialized features
which form the second boundary. The probability of the semantic relations that are
encoded by these specialized features is recalculated again using Equation (1). El
number of relations encoded by each of this boundary’s features is less than the one for
the features deﬁning the previous boundary. This process continues until each feature
has only one semantic relation attached. Each iteration creates a new boundary.

Step 2. Test the new boundary.
The new boundary is more speciﬁc than the previous boundary and it is closer to the
ideal boundary. One does not know how well it behaves on unseen examples, pero el
goal is to ﬁnd a boundary that classiﬁes these instances with high accuracy. De este modo,
the boundary is ﬁrst tested on only 10% of the annotated examples (different from
el 10% of the examples used for testing). If the accuracy is larger than the previous
boundary’s accuracy, the algorithm is converging toward the best approximation of the
boundary and thus it repeats Step 2 for the new boundary. If the accuracy is lower than
the previous boundary’s accuracy, the new boundary is too speciﬁc and the previous
boundary is a better approximation of the ideal boundary.

C. Lapata and Keller’s Web-based unsupervised model
Lauer (1995) was the ﬁrst to devise and test an unsupervised probabilistic model for
noun–noun compound interpretation on Grolier’s Encyclopedia, an eight million word
cuerpo, based on a set of eight preposition paraphrases. His probabilistic model com-
putes the probability of a preposition p given a noun–noun pair n1 − n2 and ﬁnds
the most likely preposition paraphrase p∗ = argmaxpP(pag|n1, n2). Sin embargo, as Lauer
noticed, this model requires a very large training corpus to estimate these proba-
bilities. More recently, Lapata and Keller (2005) replicated the model using the Web
as training corpus and showed that the best performance was obtained with the trigram
model f (n1, pag, n2). In their approach, they used as the count for a given trigram the num-
ber of pages returned by using the trigram as a query. These co-occurrence frequencies
were estimated using inﬂected queries which are obtained by expanding a noun–noun
compound into all its morphological forms; then searching for N P N instances, for each
of the eight prepositions P in Lauer’s list. All queries are performed as exact matches
using quotation marks. Por ejemplo, for the test noun–noun compound instance war sto-
ries, all possible combinations of deﬁnite/indeﬁnite articles and singular/plural noun
forms are tried resulting in the queries story about war, a/the story about war, story about
a/the war, stories about war, stories about the wars, story about wars, story about the wars,
etcétera. These forms are then submitted as literal queries, and the resulting hits are
summed up. The query, and thus the preposition, with the largest number of hits is
selected as the correct semantic interpretation category.

211

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

For the Europarl and CLUVI test sets, we replicated Lapata and Keller’s (2005) ex-
periments using Google.15 We formed inﬂected queries with the patterns they proposed
and searched the Web.

5. Experimental Results

We performed various experiments on both the Europarl and CLUVI testing corpora
using seven sets of supervised models. Mesa 8 shows the results obtained against SS
and Lapata and Keller’s model on both corpora and the contribution of the features
exempliﬁed in seven versions of the SVM model. Supervised models 1 y 2 are deﬁned
only for the English features. Aquí, features F1 and F2 measure the contribution of the
WordNet IS-A lexical hierarchy specialization. Sin embargo, supervised model 1, cual es
also the baseline, does not differentiate between unambiguous and ambiguous training
examples and thus does not specialize those that are ambiguous. These models show
the difference between SS and SVM and the contribution of the other English features,
such as preposition and nominalization (F1–F7).

The table shows that overall the performance is better for the Europarl corpus
than for CLUVI. For the supervised models 1 y 2, SS [F1 + F2] gives better re-
sults than SVM [F1 + F2]. The inclusion of the other English features (SVM [F1–F7])
adds more than 10% exactitud (with a higher increase in Europarl) for the supervised
modelo 1.

The results obtained are presented using the standard measure of accuracy (el

number of correctly labeled instances over the number of instances in the test set).

5.1 The Contribution of Romance Linguistic Features

Our intuition is that the more information we use from other languages for the interpre-
tation of an English instance, the better the results. De este modo, we wanted to see the impact
of each Romance language on the overall performance. Supervised model 3 shows the
results obtained for English and the Romance language that contributed the least to the
actuación (English and Spanish for the entire English feature subset F1–F8). Here we
computed the performance on all ﬁve English–Romance language combinations and
chose the Romance language that provided the best result. De este modo, supervised models 3
a través de 7 add Spanish, Francés, Portuguese, italiano, and Romanian in this order and
show the contribution of each Romance preposition and all English features.

The language ranking in Table 8 shows that Romance languages considered here
have a different contribution to the overall performance. Whereas the addition of
Portuguese in CLUVI decreases the performance, in Europarl it increases it, if only
by a few points. Sin embargo, a closer analysis of the data shows that this is mostly due
to the distribution of the corpus instances. Por ejemplo, Francés, italiano, Español, y
Portuguese are consistent in the choice of preposition (p.ej., if the preposition de [de ] es
used in French, then the corresponding preposition is used in the other four language
translations). A notable exception here is Romanian which provides two possible con-
structions with almost equal distribution: the N P N and the genitive-marked N N. El
table shows (in the increase in performance between supervised models 6 y 7) eso

15 As Google limits the number of queries to 1,000 per day per computer, we repeated the experiment using
10 computers for a number of days. Although Keller and Lapata used AltaVista for the interpretation of
two noun–noun compounds, they showed that there is almost no difference between the correlations
achieved using Google and AltaVista counts.

212

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Mesa 8
The performance obtained by ﬁve versions of the cross-linguistic SVM model compared against
the baseline, an English SVM model, and the SS model. The results obtained are presented
using the standard measure of accuracy (number of correctly labeled instances over the number
of instances in the test set).

Learning models

8-PÁGINAS

22-SR 8-PP

22-SR

Resultados [%]

CLUVI

Europarl

Supervised model 1: Base
(English nominal features only)
(no WordNet specialization)
Supervised model 2
(English features)

Supervised model 3
(English and Spanish features)
Supervised model 4
(Inglés, Español, y francés

características)

Supervised model 5
(Inglés, Español, Francés,
and Portuguese features)

Supervised model 6
(Inglés, Español, Francés,

Portuguese, and Italian features)

Supervised model 7 (SVM)
(English and all Romance features: F1–F13)
Lapata and Keller’s Web-based
unsupervised model (Inglés)

SS (F1+F2)
SVM (F1+F2)
SVM (F1-F7)
SS (F1+F2)
SVM (F1+F2)
SVM (F1-F7)
SVM (F1-F7+F8)

42.01
34.17
–
55.20
41.8
–
–

–

SVM
(F1-F7+F8+F9)

SVM
(F1-F7+F8+F9)
(+F11)
SVM
(F1-F7+F8+F9+
F10+F11)
–

46.03
38.11
50.1
61.02
46.18
61.04
63.11

65.81

64.31

66.05

35.8
30.02
–
54.12
41.03
–
–

–

36.2
33.01
43.33
57.01
41.3
67.63
68.04

69.58

69.92

71.25

72.82

–

76.34

41.10

–

42.12

–

this choice is not random, but inﬂuenced by the meaning of the instances (features F12,
F13). This observation is also supported by the contribution of each feature to the overall
actuación. Por ejemplo, in Europarl, the WordNet verb and nominalization features
of the head noun (F3, F6) have a contribution of 5.12%, whereas for the modiﬁer nouns
they decrease by about 2.7%. The English preposition (F5) contributes 6.11% (Europarl)
y 4.82% (CLUVI) to the overall performance.

The most frequently occurring preposition in both corpora is the underspeciﬁed
preposition de (de ), encoding almost all of the 22 relaciones semánticas. The many-to-
many mappings of the preposition to the semantic classes adds to the complexity
of the interpretation task. A closer look at the Europarl and CLUVI data shows that
Lauer’s set of eight prepositions represents 88.2% (Europarl) y 91.8% (CLUVI) del
N P N instances. A partir de estos, the most frequent preposition is of with a coverage of 79%
(Europarl) y 88% (CLUVI). Because the polysemy of this preposition is very high, nosotros
wanted to analyze its behavior on the set of most representative semantic relations in
both corpora. Además, we wanted to see what prepositions were used to translate the
English nominal phrase and compound instances in the target Romance languages, y
thus to capture the semantic (ir)regularities among these languages in the two corpora
and their contribution to the semantic interpretation task.

For most of the N P N instances, we noticed consistent behavior of the target
Romance languages in terms of the prepositions used. This behavior can be classiﬁed

213

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

roughly in four categories exempliﬁed subsequently: Ejemplo (12) shows a combination
of the preposition of/de and more speciﬁc prepositions; Ejemplo (13) shows different
prepositions than the one corresponding to the English equivalent in the instance; y
Examples (14) y (15) show corresponding translations of the equivalent preposi-
tion in English in all Romance languages with variations in Romanian (p.ej., de for of,
para/pour/par/pentru for for).

(12)

(13)

(14)

(15)

Committee on Culture (En.) – Comisi ´on de la Cultura (Sp.) – commission de la
cultura (Fr.) – commissione per la cultura (It.) – Comiss˜ao para Cultura (Port.) –
comitet pentru cultur˘a (Ro.) (PURPOSE)

the supervision of the administration (En.) – control sobre la administraci ´on
(Sp.) – contr ˆole sur l’administration (Fr.) – controllo sull’amministrazione (It.) –
controlo sobre administrac¸ ˜ao (Port.) – controlul asupra administrat¸iei (Ro.)
(THEME)

lack of protection (En.) – falta de protecci ´on (Sp.) – manque de protection (Fr.) –
mancanza di tutela (It.) – falta de protecc¸ ˜ao (Port.) – lips˘a de protect¸ie (Ro.)
(THEME)

the cry of a man (En.) – el llanto de un hombre (Sp.) – un cri d’homme (Fr.) –
l’urlo di un uomo (It.) – o choro de um bˆebado (Port.) – strig˘atul unui om (Ro.)
(AGENT)

Because the last three categories are the most frequent in both corpora, we analyzed
their instances. Most of the time Spanish, Francés, italiano, and Portuguese make use of
speciﬁc prepositions such as those in Examples (12) y (13) to encode some semantic
relations such as PURPOSE and LOCATION, but rely on N de N constructions for almost
all the other relations. English and Romanian, sin embargo, can choose between N N and
N P N constructions. In the next section we present in more detail an analysis of the
semantic correlations between English and Romanian nominal phrases and compounds
and their role in the semantic interpretation task.

6. Linguistic Observations

In this section we present some linguistic observations derived from the analysis of
the system’s performance on the CLUVI and Europarl corpora. More speciﬁcally, nosotros
present different types of ambiguity that can occur in the interpretation of nominal
phrases and compounds when using more abstract interpretation categories such as
Lauer’s eight prepositions. We also show that the choice of syntactic constructions
in English and Romanian can help in the identiﬁcation of the correct position of the
semantic arguments in test instances.

6.1 Observations on Lapata and Keller’s Unsupervised Model

In this section we show some of the limitations of the unsupervised probabilistic ap-
proaches that rely on more abstract interpretation categories, such as Lauer’s set of
eight prepositions. Para esto, we used Lapata and Keller’s approach, a state-of-the-art
knowledge-poor Web-based unsupervised probabilistic model which provided a per-
formance of 42.12% on Europarl and 41.10% on CLUVI. We manually checked the ﬁrst

214

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Mesa 9
Experimental results with Lapata and Keller’s Web-based unsupervised interpretation model on
different types of test sets from the Europarl corpus.

Noun–noun compound
test set

Ambiguity of noun constituents

Accuracy
[%]

Set#1
Set#2
Set#3
Set#4

one part of speech, one WordNet sense
multiple parts of speech, one WordNet sense
one part of speech, multiple WordNet senses
multiple parts of speech, multiple WordNet senses

35.28%
31.22%
50.63%
43.25%

ﬁve entries of the pages returned by Google for each most frequent N P N paraphrase
para 100 CLUVI and Europarl instances and noticed that about 35% of them were wrong
due to syntactic (p.ej., part of speech) and/or semantic ambiguities. Por ejemplo, baby cry
generated instances such as “it will make moms cry with the baby,” where cry is a verb,
not a noun. This shows that many of the NP instances selected by Google as matching
the N P N query are incorrect, and thus the number of hits returned for the query is over-
estimated. De este modo, because we wanted to measure the impact of various types of noun–
noun compound ambiguities on the interpretation performance, we further tested the
probabilistic Web-based model on four distinct test sets selected from Europarl, cada
containing 30 noun–noun compounds encoding different types of ambiguity: In Set#1
the noun constituents had only one part of speech and one WordNet sense; in Set#2 the
nouns had at least two possible parts of speech and were semantically unambiguous; en
Set#3 the nouns were ambiguous only semantically; and in Set#4 they were ambiguous
both syntactically and semantically. Mesa 9 shows that for Set#1, the model obtained an
accuracy of 35.28%, while for more semantically ambiguous compounds it obtained an
average accuracy of about 48% (50.63% [Set#3] y 43.25% [Set#4]. This shows that for
more syntactically ambiguous instances, the Web-based probabilistic model introduces
a signiﬁcant number of false positives, thus decreasing the accuracy (cf. conjuntos #1 vs. #2
y #3 vs. #4).

Además, further analyses of the results obtained with Lapata and Keller’s model
showed that about 30% of the noun–noun compounds in sets #3 y #4 were ambiguous
with at least two possible readings. Por ejemplo, paper bag can be interpreted out-
of-context both as bag of paper (bag made of paper—STUFF–OBJECT, a subtype of
PART–WHOLE) and as bag for papers (bag used for storing papers—PURPOSE). simí-
mucho, gingerbread bowl can be correctly paraphrased both as bowl of/with gingerbread
(CONTENT–CONTAINER) and as bowl of gingerbread (bowl made of gingerbread—STUFF–
OBJECT). The following two examples show the two readings of the noun–noun com-
pound gingerbread bowl as found on Google:

(16)

Stir a bowl of gingerbread,
Smooth and spicy and brown,
Roll it with a rolling pin,
Up and up and down,
…16

16 An excerpt from the “Gingerbread Man” song.

215

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

(17)

The gingerbread will take the shape of the glass bowl. Let it cool for a few
minutes and then carefully loosen the foil and remove the gingerbread from
the glass. And voil`a: your bowl of gingerbread.

These ambiguities partially explain why the accuracy values obtained for sets #3
y #4 are higher then the ones obtained for the other two sets. The semantic ambiguity
also explains why the accuracy obtained for set #2 is higher than that for set #4. For these
sets of examples the syntactic ambiguity affected the accuracy much less than the se-
mantic ambiguity (eso es, more N P N combinations were possible due to various noun
senses). This shows one more time that a large number of noun–noun compounds are
covered by more abstract categories, such as prepositions. Además, these categories
also allow for a large variation as to which category a compound should be assigned.

6.2 Observations on the Symmetry of Semantic Relations: A Study on English

and Romanian

Nominal phrases and compounds in English, nominal phrases in the Romance lan-
guages considered here, and genitive-marked noun–noun compounds in Romanian
have an inherent directionality imposed by their ﬁxed syntactic structure. Por ejemplo,
in English noun–noun compounds the syntactic head always follows the syntactic
modiﬁer, whereas in English and Romance nominal phrases the order is reversed. Two
such examples are tea/Modiﬁer cup/Head and glass/Head of wine/Modiﬁer.

The directionality of semantic relations (es decir., the order of the semantic arguments)
sin embargo, is not ﬁxed and thus it is not always the same as the inherent direc-
tionality imposed by the syntactic structure. Two such examples are ham/Modiﬁer/
Arg2 sandwich/Head/Arg1 and spoon/Modiﬁer/Arg1 handle/Head/Arg2. A pesar de
both instances encode a PART–WHOLE relation (Arg1 is the semantic argument iden-
tifying the whole and Arg2 is the semantic argument identifying the part), their se-
mantic arguments are not listed in the same order (Arg1 Arg2 for spoon handle and
Arg2 Arg1 for ham sandwich). For a better understanding of this phenomenon, nosotros
performed a more thorough analysis of the training instances in both CLUVI and
Europarl. Because the choice of syntactic constructions in context is governed in part
by semantic factors, we focused on English and Romanian because they are the only
languages from the set considered here with two productive syntactic options: N N
and N P N (Inglés) and genitive-marked N N and N P N (Romanian). De este modo, nosotros
grouped the English–Romanian parallel instances per each semantic relation and each
syntactic construction and checked if the relation was symmetric or not, according to
the following deﬁnition.

Definición
We say that a semantic relation is symmetric relative to a particular syntactic
construction if there is at least one relation instance whose arguments are in a different
order than the order indicated by the relation’s default argument frame for that
construction.

Por ejemplo, PART–WHOLE is symmetric with regard to nominal phrases because
the semantic arguments of the instance the building/Arg1 with parapets/Arg2 are in a
different order than the one imposed by the relation’s default argument frame (Arg2
P Arg1) for nominal phrases (cf. Mesa 1).

216

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Because the relation distribution is skewed in both corpora, we focused only on
those relations encoded by at least 50 instances in both Europarl and CLUVI. Para
ejemplo, in English the POSSESSION relation is symmetric when encoded by N P N and
noun–noun compounds. Por ejemplo, we can say the girl with three dogs and the resources
of the Union, but also family estate and land proprietor. The ﬁndings are summarized and
presented in Table 10 along with examples. Some relations such as IS-A, PURPOSE, y
MEASURE cannot be encoded by genitive-marked noun–noun compounds in Romanian
(indicated by “–” in the table). A checkmark symbol indicates if the relation is symmetric
('(cid:1)') or not (‘x’) for a particular syntactic construction. It is interesting to note that not
all the relations are symmetric and this behavior varies from one syntactic construction
to another and from one language to another. Although some relations such as AGENT
and THEME are not symmetric, others such as TEMPORAL, PART–WHOLE, and LOCATION
are symmetric irrespective of the syntactic construction used.

Symmetric relations pose important challenges to the automatic interpretation of
nominal phrases and compounds because the system has to know which of the nouns
is the semantic modiﬁer and which is the semantic head. In this research, the order
of the semantic arguments has been manually identiﬁed and marked in the training
corpus. Sin embargo, this information is not provided for unseen test instances. Hasta ahora,
in our experiments with the test data the system used the order indicated by the
default argument frames. Another solution is to build argument frames for clusters of
prepositions which impose a particular order of the arguments in N P N constructions.
Por ejemplo, in the N2 P N1 phrases the books on the table (LOCATION) and relaxation
during the summer (TEMPORAL), the semantic content of the prepositions on and during
identiﬁes the position of the physical and temporal location (p.ej., that N1 is the time
or location). This approach works most of the time for relations such as LOCATION
and TEMPORAL because in both English and Romance languages they rely mostly on
prepositions indicating location and time and less on underspeciﬁed prepositions such
as of or de. Sin embargo, a closer look at these relations shows that some of the noun–noun
pairs that encode them are not symmetric and this is true for both English and Romance.
Por ejemplo, cut on the chin and house in the city cannot be reversed as chin P cut or
city P house. One notable exception here is indicated by examples such as box of/with
matches – matches in/inside the box and vessels of/with blood – blood in vessels17 encoding
CONTENT–CONTAINER. Another special case is when P1 and P2 are location antonyms
(p.ej., the book under the folder and the folder on the book). Sin embargo, even here symmetry
is not always possible, being inﬂuenced by pragmatic factors (Herskovits 1987) (p.ej.,
we can say the vase on the table, but not the table under the vase—this has to do with the
difference in size of the objects indicated by the head and modiﬁer nouns. De este modo, un mayor
object cannot be said to be placed under a smaller one).

It is important to stress here the fact that our deﬁnition of symmetry of semantic
relations does not focus in particular on the symmetry of an instance noun–noun pair
that encodes the relation, although it doesn’t exclude such a case. We call this lexical
symmetry and deﬁne it here.

Definición
We say that a noun–noun pair (N1N1N1 – N2N2N2) is symmetric relative to a particular syntactic
construction and the semantic relation it encodes in that construction if the order of the
nouns in the construction can be changed provided the semantic relation is preserved.

17 Here the noun vessels refers to a type of container.

217

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

Mesa 10
A summary of the symmetry properties of a set of the 12 most frequent semantic relations in
CLUVI and Europarl. “–” means the semantic relation is not encoded by the syntactic
construction, “(cid:1)” and “x” symbols indicate whether the relation is or is not symmetric.

Symmetry

Inglés

Romanian

Semántico
relaciones

POSSESSION

No.

N N
(cid:1)

N P N
(cid:1)

genitive-marked
N N
(cid:1)

N P N

PROPERTY

AGENT

TEMPORAL

(cid:1)

PART–WHOLE (cid:1)

(cid:1)

HYPERNYMY
(is-a)

LOCATION

(cid:1)

PURPOSE

TOPIC

MEASURE

TYPE

THEME

–

(cid:1)

–

(cid:1)

–

(cid:1)

218

Examples

En.: family#2/Arg1 estate#2/Arg2 vs.
land#1/Arg2 proprietor#1/Arg1
Ro.: terenul/Arg2 proprietarului/Arg1

(land-the owner-GEN)
(‘the owner’s land’)
proprietarul/Arg1 magazinului/Arg2
(owner-the store-GEN)
(‘the owner of the store’)

En.: calm#1/Arg2 of evening#1/Arg1 vs.

spots#4/Arg1 of color#1/Arg2

Ro.: pete/Arg1 de culoare/Arg2

(‘spots of color’)
miros/Arg2 de camfor/Arg1
(‘odour of camphor’)

En.: the investigation#2/Arg2 of the police#1/Arg1
Ro.: investigat¸ia/Arg2 polit¸iei/Arg1
(investigation-the police-GEN)

En.: news#3/Arg2 in the morning#1/Arg1 vs.
the evening#1/Arg1 of her arrival#2/Arg2

Ro.: placinte/Arg2 de dimineat¸˘a/Arg1

(cakes of morning)
(‘morning cakes’) vs.
ani/Arg1 de subjugare/Arg2
(‘years of subjugation’)

En: faces#1/Arg2 of children#1/Arg1 vs.
the shell#5Arg2 of the egg#2/Arg1

Ro: fet¸ele/Arg2 copiilor/Arg1
(faces-the children-GEN)
(‘the faces of the children’) vs.
coaj˘a/Arg1 de ou/Arg2
(shell of egg)
(‘egg shell’)

En.: daisy#1/Arg1 ﬂower#1/Arg2
Ro.: meci/Arg2 de fotbal/Arg1

(match of football)
(‘football match’)

En.: castle#2/Arg2 in the desert#1/Arg1 vs.

point#2/Arg1 of arrival#1/Arg2

Ro.: castel/Arg2 in de¸sert/Arg1

(castle in desert)
(‘castle in the desert’) vs.
punct/Arg1 de sosire/Arg2
(‘point of arrival’)

En.: war#1/Arg1 canoe#1/Arg2
Ro.: pirog˘a/Arg2 de r˘azboi/Arg1

(canoe of war)

En.: war#1/Arg1 movie#1/Arg2
Ro.: ﬁlm/Arg2 despre r˘azboi/Arg1

(‘movie about war’)

En.: inches#1/Arg2 of snow#2/Arg1
Ro.: inci/Arg2 de zapad˘a/Arg1 (inches of snow)
En.: framework#1/Arg1 law#2/Arg2
Ro.: lege/Arg2 cadru/Arg1 (law framework)
En.: examination#1/Arg2 of machinery#1/Arg1
Ro.: veriﬁcarea/Arg2 ma¸sinii/Arg1

(examination-the machinery-GEN)
(‘the examination of the machinery’)

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Por ejemplo, the pair building–parapets in the nominal phrases the building/Arg1 with
parapets/Arg2 and the parapets/Arg2 of the building/Arg1 encodes a PART–WHOLE relation.
Aquí, both the noun–noun pair and the semantic relation are symmetric relative to
N P N. Sin embargo, the situation is different for instances such as the book/Arg2 under
the folder/Arg1 and the folder/Arg2 on the book/Arg1, both encoding LOCATION. Aquí, el
book–folder pair is symmetric in N P N constructions (in the ﬁrst instance the book is the
syntactic head and the folder is the modiﬁer, whereas in the second instance the order
is reversed). Sin embargo, the LOCATION relation they encode is not symmetric (in both
instancias, the order of the semantic arguments matches the default argument frame for
LOCATION). It is interesting to notice here that these two location instances are actually
paraphrases of one another. This can be explained by the fact that both the book and the
folder can act as a location with respect to the other, and that the prepositions under and
on are location antonyms. En comparación, the building with parapets is not a paraphrase
of the parapets of the building. Aquí, the nouns building and parapets cannot act as a
whole/part with respect to each other (p.ej., the only possible whole here is the noun
building, and the only possible part here is the noun parapets). This is because parts and
wholes have an inherent semantic directionality imposed by the inclusion operation on
the set of things representing parts and wholes, respectivamente.

In this research we consider the identiﬁcation and extraction of semantic relations
in nominal phrases and compounds, but we do not focus in particular on the acquisition
of paraphrases in these constructions. Our goal is to build an accurate semantic parser
which will automatically annotate instances of nominal phrases and compounds with
semantic relations in context. This approach promises to be very useful in applications
that require semantic inference, such as textual entailment (Tatu and Moldovan 2005).
Sin embargo, a thorough analysis of the semantics of nominal phrases and compounds
should focus on both semantic relations and paraphrases. We leave this topic for future
investigación.

Because we wanted to study in more detail the directionality of semantic relations,
we focused on PART–WHOLE. These relations, and most of the semantic relations con-
sidered here, are encoded mostly by N of/de N constructions, genitive-marked N N
(Romanian), and noun–noun compounds (Inglés) y por lo tanto, the task of argument order
identiﬁcation becomes more challenging. For the purpose of this research we decided
to take a closer look at the PART–WHOLE relation in both CLUVI and Europarl where
together it accounted for 920 token and 636 type instances. We show subsequently a
detailed analysis of the symmetry property on a classiﬁcation of PART–WHOLE relations
starting with a set of ﬁve PART–WHOLE subtypes identiﬁed by Winston, Chafﬁn, y
Hermann (1987):18 (1) Component–Integral object, (2) Member–Collection, (3) Portion–Mass,
(4) Stuff–Object, y (5) Place–Area.

(1) Component–Integral object

This is a relation between components and the objects to which they belong. Integral
objects have a structure with their components being separable and having a functional
relation with their wholes. This type of PART–WHOLE relation can be encoded by N of N
and less often by N N constructions. Además, here the existential interpretation is
preferred over the generic one. Such examples are the leg of the table and the table
leg which translate in Romanian as piciorul mesei (‘leg-the table-GEN’). In Romanian a

18 Winston, Chafﬁn, and Hermann (1987) identiﬁed six subtypes of PART–WHOLE relations, one of which,

(Feature–Activity), is not presented here because it is not frequently encoded by N N and N P N
constructions.

219

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

generic interpretation is also possible, but with change of construction and most of the
time of semantic relation (p.ej., picior de mas˘a – ‘leg of table’ encoding PURPOSE19).

This relation subtype is symmetric in English for both N N and N P N con-
structions. In Romanian, sin embargo, it is symmetric only when encoded by N P N.
Además, it is interesting to note that Modiﬁer/Arg1 Head/Arg2 noun–noun com-
pound instances translate as genitive noun–noun compounds in Romanian, mientras
Modiﬁer/Arg2 Head/Arg1 instances translate as N P N, with P different from of. Para
ejemplo, chair/Arg1 arm/Arg2 and ham/Arg2 sandwich/Arg1 translate in Romanian as
Head/Arg2 Modiﬁer/Arg1 – brat¸ul scaunului (‘arm-the chair-GEN’) and Head/Arg1 P
Modiﬁer/Arg2 – sandwich cu ¸sunc˘a (‘sandwich with ham’).

For N P N instances in Romanian and English both Arg1 P1 Arg2 and Arg2 P2 Arg1
argument orderings are possible, but with a different choice of preposition (with P1
different from of/de). Por ejemplo, one can say the parapets/Arg2 of the building/Arg1,
but also the building/Arg1 with parapets/Arg2. A closer look at such instances shows that
symmetry is possible when the modiﬁer (in this case the part) is not a mandatory part
of the whole, but an optional part with special features (p.ej., color, forma). Por ejemplo,
the car with the door is less preferred than the car with the red door which differentiates the
car from other types of cars.

(2) Stuff–Object

This category encodes the relations between an object and the material of which it
is partly or entirely made. The parts are not similar to the wholes which they compose,
cannot be separated from the whole, and have no functional role. The relation can
be encoded by both N of N and N N English and Romanian patterns and the choice
between existential and generic interpretations correlates with the relation symmetry.
For N N constructions this relation subtype is not symmetric, while for N P N it
is symmetric only in English. Such examples are brush/Arg2 hut/Arg1 in English, y
metalul/Arg2 scaunului/Arg1 (‘metal-the seat-GEN’ – the metal of the seat) and scaun de metal
(‘chair of metal’ – metal chair) in Romanian.

N P N instances can only be encoded by of in English or de/din (of/from) in Ro-
manian. If the position of the arguments is Arg1 of Arg2 and Arg2 is an indeﬁnite noun
indicating the part then the instance interpretation is generic. Por ejemplo, seat of metal
translates as scaun de/din metal (‘chair of/from metal’) in Romanian. It is important to
note here the possible choice of the preposition from in Romanian, a preposition which
is rarely used in English for this type of relation.

When the position of the arguments changes (p.ej., Arg2 of Arg1), the same preposi-
tion of is used and the semantic relation is still STUFF–OBJECT, but the instance is more
speciﬁc having an existential interpretation. Por ejemplo, the metal of the seat translates in
Romanian as metalul scaunului (‘metal-the seat-GEN’) and not as metalul de scaun (‘metal-
the of seat’).

(3) Portion–Mass

According to Selkirk (1982a), Ionin, Matushansky, and Ruys (2006), and our own
observations on the CLUVI and Europarl data, this type of PART–WHOLE relation can
be further classiﬁed into mass, measure, and fraction partitives. Here the parts are
separable and similar to each other and to the whole they are part of. An example of a
mass partitive is half/Arg2 of the cake/Arg1 which translates in Romanian as jum˘atate/Arg2

19 This reading is possible if the leg is separated from the table.

220

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

de/din prajitur˘a/Arg1 (‘half of/from cake’). Note that here the noun cake is indeﬁnite in
Romanian, and thus the instance is generic. An existential interpretation is possible
when the noun is modiﬁed by a possessive (p.ej., half of your cake).

Measure partitives are also called vague PART–WHOLE relations (Selkirk 1982b)
because they can express both PART–WHOLE and MEASURE depending on the context.
They are encoded by N1 of N2 constructions, where N2 is indeﬁnite, and can indi-
cate both existential and generic interpretations. Two such examples are bottles/Arg1 of
wine/Arg2 and cup/Arg1 of sugar/Arg2. In Romanian, the preposition used is either de
(de ), or cu (con). Por ejemplo, sticle/Arg1 de/cu vin/Arg2 (‘bottles of/with wine’) y
cea¸sc˘a/Arg1 de/cu zah˘ar/Arg2 (‘cup of/with sugar’).

Fraction partitives indicate fractions of wholes, such as three quarters/Arg2 of a
pie/Arg1 (trei p˘atrimi/Arg2 de pl˘acint˘a/Arg1 [Romanian]–[‘three quarters of pie’]) y uno
third/Arg2 of the nation/Arg1 (o treime/Arg2 din populat¸ia/Arg1 [Romanian]–[‘one third from
population-the’] and not o treime de populat¸ia – [‘a third of population-the’]). Here again,
we notice the choice of the Romanian preposition din and not de when the second noun
is deﬁnite. The preposition from indicates the idea of separation of the part from the
entero, an idea which characterizes PART–WHOLE relations.

Portion–Mass relations cannot be encoded by N N structures in either English or

Romanian and they are not symmetric in N P N constructions.

(4) Member–Collection

This subtype represents membership in a collection. Members are parts, pero puede
not play any functional role with respect to their whole. Eso es, comparado con
Component–Integral instances such as the knob of the door, where the knob is a round
handle one turns in order to open a door, in an example like bunch of cats, the cats don’t
play any functional role to the whole bunch.

This subtype can be further classiﬁed into a basic subtype (p.ej., the member of the
equipo), count partitives (p.ej., two of these people), fraction count partitives (p.ej., two
out of three workers), and vague measure partitives (p.ej., a number/lot/bunch of cats).
Although the basic Member–Collection partitives are symmetric for N N (Romanian
solo) and N P N (English and Romanian), the other subtypes can be encoded only by
N P N constructions and are not symmetric in English or in Romanian. Por ejemplo,
the children/Arg2 of the group/Arg1 and children/Arg2 group/Arg1 translate as copiii/Arg2
din grup/Arg1 (‘children-the from group’) and as grup/Arg1 de copii/Arg2 (‘group of
children’).

The second and the third subtypes translate in Romanian as doi/Arg2 din ace¸sti
oameni/Arg1 (‘two from these people’) and doi/Arg2 din trei lucr˘atori/Arg1 (‘two from three
workers’), by always using the preposition din ( de) instead of de (de ). En el otro
mano, vague measure partitives translate as un num˘ar/Arg1 de pisici/Arg2 (‘a number of
cats’) and not as un num˘ar din pisici (‘a number from cats’). Although all these subtypes
need to have a plural modiﬁer noun and are not symmetric, count partitives always
have an existential interpretation, whereas fraction count and vague measure partitives
have a generic meaning.

(5) Location–Area

This subtype captures the relation between areas and special places and locations
within them. The parts are similar to their wholes, but they are not separable from
a ellos. De este modo, this relation overlaps with the LOCATION relation. One such example is the
surface/Arg2 of the water/Arg1. Both nouns can be either deﬁnite or indeﬁnite and the rela-
tion is not symmetric when the part is a relational noun (p.ej., surface, end). In Romanian,

221

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

both N de N and genitive-marked N N constructions are possible: suprafat¸a/Arg2
apei/Arg1 (‘surface-the water-GEN’) and suprafat¸˘a/Arg2 de ap˘a/Arg1 (‘surface of water’).
The relation is symmetric only for N P N in both English and Romanian.

Mesa 11 summarizes the symmetry properties of all ﬁve PART–WHOLE subtypes

accompanied by examples.

De este modo, features such as the semantic classes of the two nouns (F1, F2), and the syntac-
tic constructions in English and Romanian—more speciﬁcally, the preposition features
para ingles (F5) and Romanian (F12) and the inﬂection feature for Romanian (F13)—
can be used to train a classiﬁer for the identiﬁcation of the argument order in nominal
phrases and compounds encoding different subtypes of PART–WHOLE relations. Para
ejemplo, the argument order for Portion–Mass instances can be easily identiﬁed if it is
determined that they are encoded by N2 of/de N1 in English and Romanian and the head
noun N2 is identiﬁed as a fraction in the WordNet IS-A hierarchy, thus representing Arg2
(the part). It is interesting to note here that all the other Member–Collection subtypes with
the exception of the basic one are also encoded only by N of/de N, but here the order
is reversed in both English and Romanian (N1 of/de N2), where the head noun N1, si
identiﬁed as a collection concept in WordNet, represents the whole concept (Arg1).

This approach can also be applied to other symmetric relations by classifying them
into more speciﬁc subtypes for argument order identiﬁcation. De este modo, local classiﬁers
can be trained for each subtype on features such as those mentioned herein and tested
on unseen instances. Sin embargo, for training this procedure requires a sufﬁciently large
number of examples for each subtype of the semantic relation considered.

Mesa 11
A summary of the symmetry properties of the ﬁve subtypes of PART–WHOLE semantic relation in
CLUVI and Europarl. “–” means the semantic relation is not encoded by the syntactic
construction, “(cid:1)” and “x” symbols indicate whether the relation is or is not symmetric.

No.
1

Semántico
relaciones
Component –
Integral obj.

Inglés

Romanian

Symmetry

N N
(cid:1)

N P N
(cid:1)

genitive-marked
N N
X
Arg2 Arg1

N P N
(cid:1)

Stuff –
Object

X
Arg2 Arg1

(cid:1)

X
Arg2 Arg1

X
Arg1 de Arg2

Examples

En.: chair#1/Arg1 arm#5/Arg2 vs.
ham#1/Arg2 sandwich#1/Arg1
Ro: ‘brat¸ul/Arg2 scaunului/Arg1’
(arm-the chair-GEN) vs.
‘sandwhich/Arg1 cu ¸sunc˘a/Arg2 ’
(sandwich with ham)

En.: dress#1/Arg1 of silk#1/Arg2 vs.

the silk#1/Arg2 of the dress#1/Arg1

Ro.: ‘rochie/Arg1 de m˘atase/Arg2’

(dress of silk) vs.
‘m˘atasea/Arg2 rochiei/Arg1’
(silk-the dress-GEN)

–

X
Arg2 of Arg1

X
Arg1 of Arg2

–

X
Arg2 de Arg1

X
Arg1 de Arg2

En.: half#1/Arg2 of the cake#3/Arg1 vs.
Ro: ‘jum˘atate/Arg2 de/din prajitur˘a/Arg1’

(half of/from cake)

En.: a bunch#1/Arg1 of cats#1/Arg2
Ro.: ‘o gr˘amad˘a/Arg1 de pisici/Arg2’

(a bunch of cats)

X
Arg1 Arg2

(cid:1)

Location –
Area

X
Arg1 Arg2

(cid:1)

X
Arg2 Arg1

(cid:1)

En.: president#4/Arg2 of the committee#1/Arg1 vs.

committee#1/Arg1 of idiots#1/Arg2

Ro.: copiii/Arg2 din grup/Arg1
(children-the from group)
(‘the children from the group’)
grup/Arg1 de copii/Arg2
(‘group of children’)

En.: the swamps#1/Arg2 of the land#7/Arg1 vs.
the land#7/Arg1 with swamps#1/Arg2

Ro.: oaz˘a ˆın de¸sert
(oasis in desert) vs.

de¸sert cu oaz˘a ˆın (desert with oasis)

Portion –
Mass

Member –
Collection
(count,
fraction count,
and vague measure
partitives)
Member –
Collection
(basic
partitive)

222

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

This analysis shows that the choice of lexico-syntactic structures in both English and
Romanian correlates with the meaning of the instances encoded by such structures. En
the next section we present a list of errors and situations that, currently, our system fails
to recognize, and suggest possible improvements.

7. Error Analysis and Suggested Improvements

A major part of the difﬁculty of interpreting nominal phrases and compounds stems
from multiple sources of ambiguity. These factors range from syntactic analysis, a
semantic, pragmatic, and contextual information and translation issues. En esta sección
we show various sources of error we found in our experiments and present some
possible improvements.

A. Error analysis
Two basic factors are wrong part-of-speech and word sense disambiguation tags. De este modo,
if the syntactic tagger and WSD system fail to annotate the nouns with the correct senses,
the system can generate wrong semantic classes which will lead to wrong conclusions.
Además, there were also instances for which the nouns or the corresponding senses of
these nouns were not found in WordNet. Había 42.21% WSD and 6.7% Etiquetado de punto de venta
errors in Europarl and 54.8% y 7.32% in CLUVI. Además, 6.9% (Europarl) y
4.6% (CLUVI) instances had missing senses.

There are also cases when local contextual information such as word sense disam-
biguation is not enough for relation detection and when access to a larger discourse
context is needed. Various researchers (Sp¨arck Jones 1983; Lascarides and Copestake
1998; Lapata 2002) have shown that the interpretation of noun–noun compounds, para
ejemplo, may be inﬂuenced by discourse and pragmatic knowledge. This context may
be identiﬁed at the level of local nominal phrases and compounds or sentences or at the
document and even collection level. Por ejemplo, a noun–noun compound modiﬁed
by a relative clause might be disambiguated in the context of another argument of the
same verb in the clause, which can limit the number of possible semantic relations. Para
instancia, the interpretation of the instance museum book in the subject position in the
following examples is inﬂuenced by another argument of the verbs bought in Exam-
por ejemplo (18), and informed in Example (19):

(18)

(19)

el [museum book]TOPIC John bought in the bookshop at the museum

el [museum book]LOCATION that informed John about the ancient art

Prepositions such as spatial ones are also amenable to visual interpretations due
to their usage in various visual contexts. Por ejemplo, the instance nails in the box (cf.
Herskovits 1987) indicates two possible arrangements of the nails: either held by the
box, or hammered into it. We cannot capture these subtleties with the current procedure
even if they are mentioned in the context of the sentence or discourse.

B. Suggested improvements
In this article we investigated the contribution of English and Romance prepositions to
the task of interpreting nominal phrases and compounds, both as features employed
in a learning model and as classiﬁcation categories. An interesting extension of this
approach would be to look into more detail at the functional–semantic aspect of these
prepositions and to deﬁne various tests that would classify them as pure functional
components with no semantic content or semantic devices with their own meaning.

223

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

Además, our experiments focused on the detection of semantic relations encoded
by N N and N P N patterns. A more general approach would extend the investigation
to adjective–noun constructions in English and Romance languages as well.

Another direction for future work is the study of the semantic (ir)regularities among
English and Romance nominal phrases and compounds in both directions. Tal
analysis might be also useful for machine translation, especially when translating into a
language with multiple choices of syntactic constructions. One such example is tarro
de cerveza (‘glass of beer’) in Spanish which can be translated as either glass of beer
(MEASURE) or beer glass (PURPOSE) en Inglés. The current machine translation language
models do not differentiate between such options, choosing the most frequent instance
in a large training corpus.

The drawback of the approach presented in this article, as for other very precise
learning methods, is the need for a large number of training examples. If a certain
class of negative or positive examples is not seen in the training data (and therefore
it is not captured by the classiﬁcation rules), the system cannot classify its instances.
De este modo, the larger and more diverse the training data, the better the classiﬁcation rules.
Además, each cross-linguistic study requires translated data, which is not easy to
obtain in electronic form, especially for most of the world’s languages. Sin embargo, más
and more parallel corpora in various languages are expected to be forthcoming.

8. Discussion and Conclusions

In this article we investigated the contribution of English and Romance prepositions to
the interpretation of N N and N P N instances and presented a supervised, conocimiento-
intensive interpretation model.

Our approach to the interpretation of nominal phrases and compounds is novel
in several ways. We investigated the problem based on cross-linguistic evidence from
a set of six languages: Inglés, Español, italiano, Francés, Portuguese, and Romanian.
De este modo, we presented empirical observations on the distribution of nominal phrases and
compounds and the distribution of their meanings on two different corpora, based
on two state-of-the-art classiﬁcation tag sets: Lauer’s set of eight prepositions (Lauer
1995) and our list of 22 relaciones semánticas. A mapping between the two tag sets was
also provided. A supervised learning model employing various linguistic features was
successfully compared against two state-of-the-art models reported in the literature.

It is also important to mention here the linguistic implications of this work. Nosotros
hope that the corpus investigations presented in this article provide new insight for the
machine translation and multilingual question answering communities. The translation
of nominal phrase and compound instances from one language to another is highly
correlated with the structure of each language, or set of languages. In this article we
measured the contribution of a set of ﬁve Romance languages to the task of semantic in-
terpretation of English nominal phrases and compounds. More speciﬁcally, we showed
that the Romanian linguistic features contribute more substantially to the overall per-
formance than the features obtained for the other Romance languages. The choice of
the Romanian linguistic constructions (either N N or N P N) is highly correlated with
their meaning. This distinct behavior of Romanian constructions is also explained by
the Slavic and Balkanic inﬂuences. An interesting future research direction would be to
consider other Indo- and non Indo-European languages and measure their contribution
to the task of interpreting nominal phrases and compounds in particular, and noun
phrases in general.

224

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Expresiones de gratitud
We would like to thank all our annotators
without whom this research would not have
sido posible: Silvia Kunitz (italiano) y
Florence Mathieu-Conner (Francés). Nosotros también
thank Richard Sproat, Tania Ionin, and Brian
Drexler for their suggestions on various
versions of the article. And last but not least,
we also would like to thank the reviewers for
their very useful comments.

Referencias
Alexiadou, Artemis, Liliane Haegeman, y
Melita Stavrou. 2007. Noun Phrases in the
Generative Perspective. Mouton de Gruyter,
Berlina.

Almela, Ram ´on, Pascual Cantos, Aquilino
S´anchez, Ram ´on Sarmiento, and Mois´es
Almela. 2005. Frequencias del Espa ˜nol.
Dicctionario de estudios l´exicos y morfol´ogicos.
Ed. Universitas, Madrid.

anderson, Mona. 1983. Prenominal genitive

NPs. The Linguistic Review, 3:1–24.

Artstein, Ron. 2007. Quality Control of Corpus
Annotation Through Reliability Measures.
Asociación de Lingüística Computacional
Conferencia (LCA), Prague, checo
República.

Panadero, Collin, Charles Fillmore, y juan
Lowe. 1998. The Berkeley FrameNet
Proyecto. In Proceedings of the 36th Annual
Meeting of the Association for Computational
Linguistics and 17th International Conference
on Computational Linguistics (COLING-ACL
1998), pages 86–90, Montréal.

Baldwin, Timothy. 2005. Distributional

similarity and collocational prepositional
phrases. In Patrick Saint-Dizier, editor,
Syntax and Semantics of Prepositions,
pages 197–210, Desorden, Dordrecht.
Baldwin, Timothy. 2006a. Automatic

identiﬁcation of English verb particle
constructions using linguistic features. En
Third ACL-SIGSEM Workshop on
Prepositions, pages 65–72, trento, Italia.

Baldwin, Timothy. 2006b. Representing

and Modeling the Lexical Semantics of
English Verb Particle Constructions.
In The European Association for
Ligüística computacional (EACL), el
ACL-SIGSEM Workshop on Prepositions,
trento.

nominals in the generative lexicon. In AISB
Workshop on Multilinguality in the Lexicon,
sussex.

Cadiot, Piere. 1997. Les pr´epositions abstraites

en fran¸cais. Armand Colin, París.

Calzolari, Nicoletta, Charles J. Fillmore,

Ralph Grishman, Nancy Ide, Alessandro
Lenci, Catherine MacLeod, and Antonio
Zampolli. 2002. Towards best practice for
multiword expressions in computational
lexicons. In The International Conference on
Language Resources and Evaluation LREC,
pages 1934–1940, Las Palmas.

Casadei, Federica. 1991. Le locuzioni

preposizionali. Struttura lessicale e gradi
di lessicalizzazione. Lingua e Stile, XXXVI:
43–80.

Celce-Murcia, Marianne and Diane

Larsen-Freeman. 1999. The grammar book,
2nd edition. Heinle and Heinle, Bostón,
MAMÁ.

Charniak, Eugene. 2000. A

maximum-entropy-inspired parser. In The
1st Conference of the North American Chapter
of the Association for Computational
Lingüística (NAACL), pages 132–139,
seattle, Washington.

Cornilescu, alejandra. 2001. Romanian
nominalizations: Case and aspectual
estructura. Journal of Linguistics, 37:467–501.

Dorr, bonnie. 1993. Máquina traductora: A

View from the Lexicon. CON prensa,
Cambridge, MAMÁ.

Downing, Pamela. 1977. On the creation and
use of English compound nouns. Idioma,
53:810–842.

evans, Vyvyan and Paul Chilton, editores.
2009. Idioma, Cognition and Space: El
State of the Art and New Directions.
Advances in Cognitive Linguistics.
Equinox Publishing Company, Londres.
Fang, Alex C. 2000. A lexicalist approach

towards the automatic determination for
the syntactic functions of prepositional
phrases. Natural Language Engineering,
6:183–201.

Fellbaum, Christiane. 1998. WordNet—An
Electronic Lexical Database. CON prensa,
Cambridge, MAMÁ.

Finin, Timothy W. 1980. The Semantic

Interpretation of Compound Nominals.
Doctor. tesis, University of Illinois at
Urbana-Champaign, Urbana-Champaign,
IL.

Pregonero, cris. 1998. Partitives, double

Giorgi, Alessandra and Giuseppe

genitives and anti-uniqueness.
Natural Language and Linguistic Theory,
16:679–717.

Lombardos. 1991. The syntax of noun
phrases. Prensa de la Universidad de Cambridge,
Londres.

Busa, Federica and Michael Johnston. 1996.
Cross-linguistic semantics for complex

Girju, Roxana, Alexandra Badulescu, y

Dan Moldovan. 2006. Automatic discovery

225

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

of part-whole relations. computacional
Lingüística, 32(1):83–135.

Conferencia sobre Inteligencia Artificial (AAAI),
pages 691–696, austin, Texas.

Girju, Roxana, Dan Moldovan, Marta Tatu,

Kordoni, Valia. 2005. Prepositional

and Daniel Antohe. 2005. Sobre el
Semantics of Noun Compounds. Computadora
Speech and Language, Special Issue on
Multiword Expressions, 19(4):479–496.

Gleitman, Lila R. and Henry Gleitman. 1970.
Phrase and Paraphrase: Some Innovative Uses
of Language. norton, Nueva York.

Gocsik, karen. 2004. English as a Second
Idioma. Dartmouth College Press,
Hanover, NH.

Grimshaw, Jane. 1990. Argument Structure.

CON prensa, Cambridge, MAMÁ.

Herskovits, Annette. 1987. Language and
Spatial Cognition: An Interdisciplinary
Study of the Prepositions in English.
Prensa de la Universidad de Cambridge, Cambridge,
MAMÁ.

Ionin, Tania, Ora Matushansky, and Eddy
Ruys. 2006. Parts of speech: Toward a
uniﬁed semantics for partitives. En
Conference of the North East Linguistic
Sociedad (NELS), pages 357–370,
Amherst, MAMÁ.

Jensen, Per Anker and J˝orgen F. Nilsson.
2005. Ontology-based semantics for
prepositions. In Patrick Saint-Dizier,
editor, Syntax and Semantics of Prepositions,
volumen 29 of Text, Speech and Language
Tecnología. Saltador, Dordrecht.

Jespersen, Otón. 1954. A Modern English

Grammar on Historical Principles. Jorge
allen & Unwin Ltd., Heidelberg and
Londres.

Johnston, Michael and Federica Busa. 1996.
Qualia structure and the compositional
interpretation of compounds. In Evelyne
Viegas, editor, Breadth and Depth of
Semantics Lexicon, pages 77–88, Desorden
Académico, Dordrecht.

kim, Su Nam and Timothy Baldwin. 2005.

Automatic interpretation of noun
compounds using WordNet similarity.
In The International Joint Conference on
Natural Language Processing (IJCNLP),
pages 945–956, Jeju.

kim, Su Nam and Timothy Baldwin. 2006.
Interpreting semantic relations in noun
compounds via verb semantics. In The
International Conference on Computational
Lingüística / la Asociación para
Ligüística computacional (COLING/ACL) –
Main Conference Poster Sessions,
pages 491–498, Sídney.

Kipper, Karin, Hoa Trang Dang, and Martha
Palmer. 2000. Class-based construction
of a verb lexicon. In The National

226

arguments in a multilingual context.
In Patrick Saint-Dizier, editor,
Syntax and Semantics of Prepositions,
volumen 29 of Text, Speech and Language
Tecnología. Saltador, Dordrecht,
pages 307–330.

Kordoni, Valia. 2006. PPs as verbal

argumentos: From a computational
semantics perspective. In The European
Asociación de Lingüística Computacional
(EACL), the ACL-SIGSEM Workshop on
Prepositions, trento, Italia.

Lapata, Mirella. 2002. The Disambiguation of
nominalisations. Ligüística computacional,
28(3):357–388.

Lapata, Mirella and Frank Keller. 2005.

Web-based models for natural language
Procesando. ACM Transactions on Speech and
Procesamiento del lenguaje, 2:1–31.

Lascarides, Alex and Ann Copestake. 1998.
Pragmatics and word meaning. Diario de
Lingüística, 34:387–414.

Lauer, Marca. 1995. Corpus statistics meet the
noun compound: Some empirical results.
In The Association for Computational
Linguistics Conference (LCA), pages 47–54,
Cambridge, MAMÁ.

Lees, Robert B. 1963. The Grammar of English
Nominalisations. Moutón, The Hague.

Lenci, Alessandro, Nuria Bel, Federica Busa,

Nicoletta Calzolari, Elisabetta Gola,
Monica Monachini, Antoine Ogonowski,
Ivonne Peters, Wim Peters, Nilda Ruimy,
Marta Villegas, and Antonio Zampolli.
2000. SIMPLE: A general framework for
the development of multilingual lexicons.
International Journal of Lexicography,
13:249–263.

Lersundi, Mikel and Eneko Aggire. 2006.

Multilingual inventory of interpretations
for postpositions and prepositions. En
Patrick Saint-Dizier, editor, Syntax and
Semantics of Prepositions, volumen 29 of Text,
Speech and Language Technology. Saltador,
Dordrecht, pages 69–82.

leví, Judith. 1978. The Syntax and Semantics
of Complex Nominals. Prensa académica,
Nueva York.

Linstromberg, Seth. 1997. English Prepositions
Explained. John Benjamins Publishing Co.,
Amsterdam/Philadelphia.

Litkowski, Kenneth C. and Orin Hargraves.

2005. The Preposition Project. In The
ACL-SIGSEM Workshop on the Linguistic
Dimensions of Prepositions and their Use in
Computational Linguistics Formalisms and

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Girju

The Syntax and Semantics of Prepositions

Aplicaciones, pages 171–179, Colchester,
Reino Unido.

Luraghi, Silvia. 2003. Prepositions in Greek and
Indo-European. Benjamín, Ámsterdam.

Lyons, Christopher. 1986. The syntax of

English genitive constructions. Diario de
Lingüística, 22:123–143.

Melis, Ludo. 2002. Les pr´epositions du fran¸cais.
L’essentiel franc¸ais. Ophrys, Paris/Gap.
Meyers, A., R. Reeves, Catherine Maclead,
Rachel Szekely, Veronika Zielinsk, brian
Joven, y r. Grishman. 2004. El
cross-breeding of dictionaries. En
Proceedings of the 5th International
Conference on Language Resources and
Evaluation (LREC-2004), pages 1095–1098,
Lisbon.

Mihalcea, Rada and Ehsanul Faruque. 2004.

SenseLearner: Minimally supervised word
sense disambiguation for all words in
open text. In Senseval-3: Third International
Workshop on the Evaluation of Systems for the
Semantic Analysis of Text, pages 155–158,
Barcelona.

Moldovan, Dan and Adriana Badulescu.

2005. A semantic scattering model for the
automatic interpretation of genitives. En
Proceedings of Human Language Technology
Conference and Conference on Empirical
Métodos en el procesamiento del lenguaje natural,
pages 891–898, vancouver.

Moldovan, Dan, Adriana Badulescu, Marta
Tatu, Daniel Antohe, and Roxana Girju.
2004. Models for the semantic
classiﬁcation of noun phrases. In The
Human Language Technology Conference /
North American Association for
Computational Linguistics Conference
(HLT/NAACL), Workshop on Computational
Lexical Semantics, pages 60–67,
Bostón, MAMÁ.

Moldovan, Dan and Roxana Girju. 2003.
Proceedings of the Tutorial on Knowledge
Discovery from Text. Asociación para
Ligüística computacional, Sapporo,
Japón.

Nakov, Preslav and Marti Hearst. 2005.
Search engine statistics beyond the
n-gram: Application to noun compound
bracketing. In The 9th Conference on
Computational Natural Language Learning,
pages 835–842, vancouver.

O’Hara, Tom and Janyce Wiebe. 2003.

Preposition semantic classiﬁcation via
Treebank and FrameNet. In Conference on
Computational Natural Language Learning
(CONLL), pages 79–86, Edmonton.
Pantel, Patrick and Marco Pennacchiotti.
2006. Espresso: Leveraging generic

patterns for automatically harvesting
relaciones semánticas. In The International
Computational Linguistics Conference /
Asociación de Lingüística Computacional
(COLING/ACL), pages 113–120, Sídney.
Pantel, Patrick and Deepak Ravichandran.
2004. Automatically labeling semantic
classes. In The Human Language Technology
Conferencia del Capítulo Norteamericano
of the Association for Computational
Lingüística (HLT/NAACL), pages 321–328,
Bostón, MAMÁ.

Pennacchiotti, Marco and Patrick Pantel.

2006. Ontologizing semantic relations. En
The International Computational Linguistics
Conferencia / Asociación de Computación
Lingüística (COLING/ACL), pages 793–800,
Sídney.

Pustejovsky, James. 1995. The Generative
Lexicon. CON prensa, Cambridge, MAMÁ.

Pustejovsky, James, Catherine Havasi, Roser

Sauri, Patrick Hanks, Jessica Littman,
Anna Rumshisky, Jose Castano, and Marc
Verhagen. 2006. Towards a generative
lexical resource: The brandeis semantic
ontology. In The International Conference on
Language Resources and Evaluation (LREC),
pages 385–388, Genoa.

Romaine, Suzanne. 1995. Bilingualism.

Blackwell, Oxford.

Rosario, Barbara and Marti Hearst. 2001.

Classifying the semantic relations in noun
compounds. In Conference on Empirical
Métodos en el procesamiento del lenguaje natural,
pages 82–90, pittsburgh, Pensilvania.

Rosario, Barbara, Marti Hearst, and Charles
Fillmore. 2002. The descent of hierarchy,
and selection in relational semantics. En
The 40th Annual Meeting of the Association
para Lingüística Computacional (LCA),
pages 247–254, Filadelfia, Pensilvania.
Saint-Dizier, Patrick. 2005a. PrepNet: A

framework for describing prepositions:
Preliminary investigation results. In The
6th International Workshop on Computational
Semántica, pages 25–34, Tilburg.

Saint-Dizier, Patrick, editor. 2005b. Syntax
and Semantics of Prepositions. Saltador,
Dordrecht.

Selkirk, Elisabeth. 1982a. Some remarks on

noun phrase structure. In Peter W.
Culicover, Thomas Wasow, and Adrian
Akmajian, editores, Formal Syntax.
Prensa académica, Londres.

Selkirk, Elisabeth. 1982b. Syntax of Words.

CON prensa, Cambridge, MAMÁ.

Sp¨arck Jones, karen. 1983. Compound
noun interpretation problems. En
Frank Fallside and William A. Woods,

227

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

Ligüística computacional

Volumen 35, Número 2

editores, Computer Speech Processing.
Prentice Hall, Acantilados de Englewood, Nueva Jersey,
pages 363–381.

Tatu, Marta and Dan Moldovan. 2005. A

semantic approach to recognizing textual
entailmant. In Proceedings of Human
Language Technology Conference and
Jornada sobre Métodos Empíricos en Natural
Language Proceesing (HLT/EMNLP 2005),
pages 371–378, vancouver.

Turney, Peter. 2006. Expressing implicit

semantic relations without supervision. En
The Computational Linguistics Conference /
Asociación de Lingüística Computacional
Conferencia (COLING/ACL), pages 313–320,
Sídney.

tyler, Andrea and Vyvyan Evans. 2003.
The Semantics of English Prepositions:
Spatial Sciences, Embodied Meaning, y
Cognición. Prensa de la Universidad de Cambridge,
Cambridge, MAMÁ.

Vandeloise, Claude, editor. 1993. La couleur
des pr´epositions, volumen 110. Larousse,
París.

Villavicencio, Aline. 2006. Verb-particle

constructions in the World Wide Web. En
Patrick Saint-Dizier, editor, Syntax and
Semantics of Prepositions, volumen 29 of Text,
Speech and Language Technology. Saltador,
Dordrecht, pages 115–130.

Volk, Martín. 2006. How bad is the problem

of PP-attachment? A comparison of
Inglés, German and Swedish. In The
European Association for Computational
Lingüística (EACL), the ACL-SIGSEM
Workshop on Prepositions, pages 81–88,
trento.

zorros, Peter. 1998. EuroWordNet: A
Multilingual Database with Lexical
Semantic Networks. Kluwer Academic
Publishers, Verlag.

Winston, Morton, Roger Chafﬁn, y

Douglas Hermann. 1987. A taxonomy of
part-whole relations. Ciencia cognitiva,
11:417–444.

Zelinski-Wibbelt, Cornelia, editor. 1993. El

Semantics of Prepositions. Mouton de
Gruyter, Berlina.

D
oh
w
norte
oh
a
d
mi
d

F
r
oh
metro
h

t
t

pag

:
/
/

d
i
r
mi
C
t
.

metro

i
t
.

mi
d
tu
/
C
oh

yo
i
/

a
r
t
i
C
mi
–
pag
d

F
/

3
5
2
1
8
5
1
7
9
8
6
2
4
/
C
oh

yo
i
.

0
6
–
7
7
–
pag
r
mi
pag
1
3
pag
d

b
y
gramo
tu
mi
s
t

oh
norte
0
8
S
mi
pag
mi
metro
b
mi
r
2
0
2
3

228
Descargar PDF