English. This paper describes a theoretical and empirical investigation about the position of adjectives in the Italian language. The long term goal which oriented the study is the formalization of this information into a natural language generation system. Providing that adjectives mainly occur within noun phrases, we focused on them and we collected data from corpora representing very different text genres, i.e. social media and standard ones, in order to compare the theoretical predictions with the real use of the adjective in Italian. The results obtained by confirm the previsions of the modern linguistic theories but also show the different behaviour of adjectives in the distinct analysed genres.
Italiano. Questo lavoro presenta un’analisi teorica ed empirica sulla posizione degli aggettivi nella lingua Italiana. L’orientamento del lavoro e` dato dalla necessita` di formalizzare questa informazione nell’ambito di un sistema di generazione automatica della linguaggio. Poiche´ gli aggettivi si presentano principalmente nei sintagmi nominali, ci si e` concentrati su questi, raccogliendo dati da corpora che rappresentano generi di testo diversi, ovvero social media e standard, al fine di confrontare le previsioni teoriche con l’uso reale dell’aggettivo in Italiano. I risultati ottenuti confermano le previsioni delle moderne teorie linguistiche ma mostrano anche il diverso comportamento degli aggettivi nei diversi generi analizzati.
Corpus linguistics is a methodological approach
based on the extraction from a set of texts of data
useful for the study of language. Even if in
principle any collection of texts can be called corpus,
the term assumes a more precise connotation in
the context of modern linguistics, where a corpus
is featured by sampling, representativeness, finite
size, machine-readable form and a standard
reference
In this work we have applied a corpus-based approach and we considered two different corpora which represent two different text genres: one concerning social media language (PoSTWITA corpus) and one concerning balanced standard Italian (UD-it corpus). Indeed, while social media texts have recently gained great attention from the NLP community since they have many peculiar properties, standard texts can give a more accurate view on the status of some linguistic notions in “traditional” written text.
These above mentioned corpora allowed us an
in depth investigation about the position of the
adjective in the nominal phrase. Indeed, even if this
grammatical category is described in several
traditional Italian grammars
This work aims at achieving two major goals.
The first is to empirically confirm with the
analysis of corpora the theoretical predictions given
in
The paper is organized as follows: in Section 2 we review the linguistic literature concerning the position of the adjective within the Italian noun phrase. In Section 3, we explain the details of our corpus linguistic investigation. In Section 4, we describe the use of the empirical data in the SimpleNLG-IT realizer. Finally, the Section 5 closes the paper with conclusions and some pointers to future work. 2
We take into account the adjective in its primary
use
These different behaviour of the adjective also
carry different semantic values: nominal phrases
where the adjective precedes the noun indicate
more subjectivity or more stylistic refinement if
compared to the more neutral and objective
expressions where the adjective follows the noun,
as in the following examples (extracted from
Moreover, a descriptive function is usually
attributed in literature to pre-nominal adjectives,
while a restrictive function is attributed to
postnominal ones, e.g. in
In
The adjectives collocated in the lower part of the
hierarchy are more prone to assume post-nominal
positions, where those in the higher part more
frequently assume the pre-nominal ones. For
instance, the relational adjectives, that are at the
lower level of the hierarchy, are predominantly
post-nominal. The others can be freely positioned
before or after the noun, but those occupying lower
positions within the hierarchy have a stronger
tendency for post-nominal positions, while those in
higher part of the hierarchy are more freely placed
before or after noun
In order to validate the assumptions made in literature, and described in section 2 about the behaviour of the adjective, we selected corpora where Italian is annotated for what concerns morphology and syntax and representing also different text genres. We applied scripts in Python and SQL queries for detecting the presence of adjectives and noun phrases in both the reference corpora, but their classification is manually done, for carefully dealing with cases where ambiguity occurs.
We found a substantial help for finding a decisionmaking criterion for the classification of adjectives in the examples proposed in the Treccani online vocabulary. For instance, we tagged as evaluative the adjective pericoloso (dangerous), which is derived from the noun pericolo (danger), according to the vocabulary example un viaggio pericoloso (a dangerous journey). We tagged instead as relational the adjective solare (solar), like in the example luce solare (solar light), considering that the adjective is derived from the noun sole (sun), indicating an entity rather than a quality.
A particular attention must be paid to homonymous adjectives, like e.g. reale that may mean ’royal’ or ’real’. In this case, two different entries in the vocabulary must be introduced, one for each meaning of the adjective: the first tagged as relational, for the meaning derived from the noun re (king), and the second tagged as evaluative, for indicating the meaning ’actually existing’.
In the rest of this section the resources we used in our investigation are described also showing the differences that make them especially interesting for validating our results in two different contexts and text domains.
The data sets we used are respectively extracted
from two different corpora: PoSTWITA
PoSTWITA characterised by short texts (140 characters maximum) and a typical social media Italian jargon that is featured by a frequent use of creative expressions and incorrect words like in the following example: ho un disparato bisogno di soffocati di coccole. <3 ti amo piccola mia. ([I] have a desperate need 2http://universaldependencies.org/it/ overview/introduction.html 3www.universaldependencies.org to suffocate you with pampering. <3 [I] love you my baby.) where two incorrect words occur: disparato instead of disperato and soffocati instead of soffocarti.
Also distinctive graphic practices due to the
particular medium are symbols are very frequent
in Twitter posts, like e.g. acronyms and
abbreviations and elements without a clearly
defined syntactic function like hashtags, mentions
and emoticons
Each word of PoSTWITA is associated with a tag showing its grammatical category selected within the inventory of tags proposed for the part of speech tagging within the Universal Dependency project; only a few tags extends this inventory for better describe typical social media elements, like EMO for emoticons or URL for web addresses. Within our corpora we focused only on the words tagged as ADJ (adjectives), NOUN (common nouns) and PROPN (proper nouns), that is those involved in the noun phrase structures. Nevertheless, it must be observed that since PoSTWITA corpus is only tagged morphologically, a proper notion of noun phrase is not marked in it. In order to detect adjectives that are syntactically linked to nouns within noun phrases, we considered the adjectives that were immediately before or after nouns or proper nouns. According to this strategy, the number of adjectives occurring in prenominal position is 1,519, while the number of those in postnominal position is 1,740. 3.2
UD-it
UD-it corpus is tagged both morphologically and
syntactically. It is derived from the conversion
of different resources developed by Turin and
Pisa University’s Computer Science Departments
and Pisa CNR’s Computational Linguistics
Institute. This corpus is composed by legal texts
(Italian Constitution and part of the Civil Code),
Wikipedia and newspaper articles. We can
therefore say that, unlike PoSTWITA corpus, UD-it
corpus is representative of the so-called Standard
Italian, that is encoded, over regional, elaborate,
belonging to the upper classes, invariant and
written
It must be observed that the availability of the syntactic annotation of the UD-it corpus has allowed more reliable results with respect to that obtained from PoSTWITA. Indeed we can not be sure that an adjective is related to a specific noun just because it is near that noun, providing that an adjective can refer to a noun even if distant from it, as the following example shows, where an adverbial modifier is collocated between the noun and the adjective that modifies it: adottare principi il piu` possibile semplici VERB NOUN DET ADV ADJ ADJ (adopting principles the most possible simple) amod 3.3
The pie charts (Fig. 1) show the data extraction results. The largest percentage of the post-nominal adjectives provides some hints about the markedness of the pre-nominal position for both PoSTWITA and UD-it.
For what concerns the distribution in pre- and post-nominal position of the categories of adjectives described in sec. 2.1, it is represented in the histograms as detected in Figure 3 (PoSTWITA) and Figure 2 (UD-it). We collected these data by applying to our datasets scripts in Python and SQL queries running on a database version of the resources.
The diagrams show how the adjectives in the lower portion of the hierarchy (relational, colour and physical property) are predominantly in post-nominal position within the noun phrase, whereas the adjectives in the higher portion of the hierarchy (age and dimension) are in majority in the pre-nominal one. Evaluative adjectives are the most equally distributed. These results confirm the theoretical tenets presented in the previous part of the paper and collocate the behaviour of the adjective within a context that can be used for modelling in a computational perspective this grammatical category. 4
The formalization of linguistic properties is a
fundamental process both for NL processing as well
as for NL generation systems. In particular, a
widespread architecture for NLG assumes a
specific module for the linguistic realization, that is
essentially an algorithmic implementation of a
formal grammar
Nevertheless, providing that a more correct
modelling of the behaviour of words has a
positive impact on the human-machine interaction,
in SimpleNLG-IT we devised a new version of
the lexicon by following the procedure described
in
All the resource developed are made available on a free access repository5. 5
The paper presents a study about the behaviour of the adjective within the noun phrase. Providing that the qualitative description given by tradi4https://dizionario.internazionale.it/ nuovovocabolariodibase
5https://github.com/alexmazzei/ SimpleNLG-IT Category dimension age physical property colour relational evalutativepre evalutativepost
NVdB/UD-it 15/16 7/7 4/4 10/11 111/121 33/35 61/68 tional grammars does not allow the definition of a formal model, we considered a recent study that classifies the descriptive adjectives. The long term goal which oriented this study is to contribute to the development of a natural language generation system for Italian featured by a more careful modelling of the behaviour of words within sentence structures.
Assuming a corpus-based perspective we tested on two corpora for Italian the tenets of this study. The results confirm and validate the theory thus opening the window for a definition of a formal model that can be exploited in our computational framework.
Future work is planned to extend the validation of our model on larger datasets, where a wider variety of adjectives is used and also more complex noun phrase structures are taken into account with respect to the simple <adjective - noun>or <noun - adjective>associations here considered. In particular, providing that more than one adjective can occurs within a noun phrase and can be syntactically linked to a single noun, we intend to investigate on the preference order also in these cases.