The paper describes the creation of a manually validated dataset of Italian multiword expressions, building on candidates automatically extracted from corpora of written Italian. The main features of the resource, such as POS-pattern and lemma distribution, are also discussed, together with possible applications.
The computational treatment of multiword
expressions (henceforth, MWEs) is notoriously a major
challenge in NLP
With this work, we contribute to this line of
research by providing a dataset of 1,682 validated
Italian multiword expressions, obtained through
the manual annotation of candidates automatically
extracted from corpora of written Italian within the
CombiNet project
Copyright c 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). report on preliminary analyses carried out with respect to MWE features and distribution. 2
For the creation of the dataset we built on data
extracted within the CombiNet project, where the
computational task of extracting candidate word
combinations from corpora was aimed at
supporting the creation of an online lexicographic
resource for Italian
As a consequence, two different extraction methods – both based on the technique of searching corpora1 with sets of patterns, and ranking retrieved candidates using frequency and association measures – were used.2 More precisely, the search was performed using, in turn, shallow part-ofspeech (POS) sequences and syntactic relations: the former method performs better with fixed and adjacent word combinations, whereas the latter is more efficient for syntactically flexible combinations. Since for the present work we focus more on MWEs proper rather than combinatorics in general, we opted to use the data previously gathered with the POS-based method.
Candidates were obtained by feeding the EXTra
software
2For a full description of the methods and their
assessment, see
As a first step, we selected top-ranked results by cutting at LL 7,500, which we observed to be a good balance between precision (high chance of being a MWE) and recall (enough variety), yielding 7,045 candidates. Then we manually annotated this list of candidates to obtain the gold standard inventory of Italian MWEs released and described in the present paper. Each candidate was validated independently by two annotators, and a third annotator judged the conflicted cases,3 which amounted to 673 (less than 10%). We validated sequences that were deemed to display some type of conventionality (fixedness, idiomaticity, high familiarity of use). We included only MWEs in their ‘full’ form (e.g., punto di partenza ‘starting point’, in breve tempo ‘in a short time’), thus excluding sequences that were clearly part of incomplete MWEs (e.g. scanso di equivoci, lit. avoidance of misunderstandings, as part of the larger adverbial MWE a scanso di equivoci, lit. at avoidance of misunderstandings, ‘to avoid misunderstandings’). 3
The final list of valid MWEs amounts to 1,682 (about 24% of the candidates), and is made available to the community.4 The resource contains the following information: (i) lemmatized MWE;5 (ii) corresponding POS-pattern;6 (iii) corpus/corpora where the MWE was found; (iv) LogLikelihood; (v) raw frequency. 3.1
In order to make our resource re-usable on the very same corpora employed for the extraction, 3All annotations were performed by the authors.
4DOI: 10.6092/unibo/amsacta/6506. http://amsacta.unibo.it/id/eprint/6506 5MWEs are lemmatized because the extraction was performed using lemmas. A consequence of this is that we may have two identical lemmatized sequences that however differ in POS-tagging. For instance, cambio di guardia (lit. change of guard) occurs twice: in one case di ‘of’ is tagged as a bare preposition, in the other as an articulated preposition (della ‘of the’), giving rise to two partially different MWEs (the latter may mean both ‘changing of the guard’ and ‘changeover of leaders’, whereas the former can refer only to the second of these meanings).
6The tagset is available here: http://medialab. di.unipi.it/wiki/Tanl_POS_Tagset we kept all data in their original form. This means that lemmatization and POS-tagging were retained, even if erroneous.
Examples of errors and anomalies include: (a) inconsistent lemmatization, especially for prepositions (e.g. radere al suolo ‘raze to the ground’ occurs twice, lemmatized as radere a suolo and radere al suolo, although the preposition is correctly tagged as an articulated preposition in both cases) and conjunctions (e.g. carne e ossa ‘flesh and blood’ and the almost identical carne ed ossa, with the euphonic -d on the conjunction e ‘and’, are two separate items);
(b) wrong lemmatization and tagging, especially for participial-like forms (e.g. centro abitato ‘residential area’, lit. center inhabited, lemmatized as centro abitare, lit. center to inhabit; or posta elettronica ‘electronic mail’ lemmatized as porre elettronico, lit. to put electronic, since posta is interpreted as the feminine past participle of porre ‘to put’ and not as the noun posta ‘mail’), but not only (e.g. lavori di costruzione ‘construction works’ lemmatized as lavorio [instead of lavoro] di costruzione; or meccanica quantistica ‘quantum mechanics’ where meccanica is tagged as an adjective);
(c) multiple tagging for the same form (essere vero ‘be true’ occurs twice because vero is tagged sometimes as an adjective, sometimes as an adverb).
Tricky cases also include lexicalized forms (guarda caso ‘strangely enough’, where guarda is – correctly, from the technical point of view – lemmatized as guardare ‘look’ and tagged as verb, although it is no longer a verb within that lexicalized expression) and pronominal verbs (like sentirsi in dovere ’to feel obliged’, where the verb is lemmatized as sentire ’to feel’, and not as its reflexive form sentirsi, although the MWE requires the reflexive form). 3.2
The validated MWEs in this first release instantiate 82 POS patterns out of the 122 used for the extraction (cf. Section 2). Non-represented patterns (over 30% of the original set) include e.g. Prep-Adj-Verb (e.g. per quieto vivere ‘for a quiet life’) as well as more complex – and arguably less frequent – patterns such as N-Prep-ArtDef-N-Adj (e.g. lotta contro la criminalita` organizzata ‘fight against organized crime’).
V-Adv N-N V-Adj
Fq. 165 152 110
Example punto di vista ‘viewpoint’ valere la pena ‘to be worth’ scendere in campo ‘to take the field’ 83 avere paura
‘to be afraid’ 83 correre un rischio
‘to run a risk’ 80 tavola rotonda
‘round table’ 79 vigile del fuoco
‘fireman’ 77 di fronte a
‘in front of’ 75 al fine di
‘with the aim of’ 63 di parte
‘biased’ 62 andare avanti
‘to go on’ 62 piano terra
‘ground floor’ 55 essere presente
‘to be there’ 47 entrare nel merito
‘to address’ 35 dietro le quinte
‘behind the scenes’
Overall, most attested patterns are 2- or 3grams. The first 4-slot pattern V-Prep-ArtIndef-N only appears at rank 36, corresponding to 8 different MWEs (e.g. rispondere a una domanda ‘to answer a question’).
In terms of lexical categories, expectedly, most
frequent patterns pertain to the nominal and
verbal domains. The N-Prep(Art)-N type is the most
common pattern for complex nominals, in
agreement with theoretical literature
Example 1 in bianco e nero
‘in black and white’ 1 dare il via libera
‘to give green light’ 1 difficile a dirsi
‘difficult to say’ 1 mettere a dura prova
‘to put a strain (on)’ 1 degno di nota
‘noteworthy’ The single-word lemmas that concur to form the MWEs in our list amount to 1,235.
Not surprisingly, among the most used lemmas we find function words like prepositions (di ‘of’ fq.421; in ‘in’ fq.227; al ‘at/to the’ fq.124, a ‘at/to’ fq.55 and ad ‘at/to’ fq.10; per ‘for’ fq.50; da ‘from’ fq.34; su ‘on’ fq.24; con ‘with’ fq.20) and determiners (il ‘the’ fq.208; un ‘a’ fq.71 and una ‘a’ fq.41), which appear in many POS-patterns. Conjunctions are instead less frequent (e ‘and’ fq.21 and ed ‘and’ fq.4; o ‘or’ fq.4), like quantifiers (e.g. ogni ‘each’ fq.11).
Quite expectedly, top-ranked verbs (essere ‘to be’ fq.67; fare ‘to do/make’ fq.46; avere ‘to have’ fq.36; mettere ‘to put’ fq.35; prendere ‘to take’ fq.27; andare ‘to go’ fq.19; dare ‘to give’ fq.17) and top-ranked nouns (tempo ‘time’ fq. 32; mano ‘hand’ fq.26; parte ‘part’ fq.23; posto ‘place’ fq.17; giorno ‘day’ fq.16) are lexemes carrying a generic meaning, which favors their combinatory power. Among the mostly used words we also find numerals like primo ‘first’ (fq.30) or secondo ‘second’ (fq.18), and adverbs like non ‘not’ (fq.29).
A cursory comparison between the lemmas of
the MWEs in our list and the Vocabolario di Base
The distribution of MWEs in the two corpora used for the extraction is shown in Table 3.
We retrieved more MWEs from la Repubblica
la Repubblica (total) PAISA` (total) la Repubblica (only) PAISA` (only) Both than PAISA`, which is expected given that the latter is smaller in size (250M tokens vs. 380M). What is less expected is the rather low number of MWEs shared by the two corpora, amounting to 372, hence 22%. Although la Repubblica is a journalistic source and PAISA` is a web corpus containing more varied text genres (especially from Wikimedia Foundation projects), we expected a larger convergence, considering that they both contain written (mid-)formal texts and that PAISA` also contains texts from the news.
Some POS-patterns seem to be definitely more typical of one corpus over the other. As Table 4 illustrates, the N-Prep-N pattern, for instance, is much more typical of la Repubblica, whereas the N-Adj pattern is more attested in PAISA`.
la Repubblica (only) PAISA` (only) Both
Among top-ranked MWEs for both LogLikelihood and raw frequency we find in grado di ‘able to’ and per la prima volta ‘for the first time’, in both corpora. The highest ranked MWEs in PAISA` is voce correlata ‘see also’, which is obviously due to the texts that form this resource. Generally, topranked MWEs for LogLikelihood also have high frequency, but not in all cases: essere in essere ‘to exist’, for instance, turns out to be highly significant in terms of LogLikelihood but has a very low frequency in both corpora.
The sequences contained in this release are obviously quite heterogeneous.
Semantically speaking, some are very idiomatic in meaning (e.g. braccio di ferro ‘arm wrestling’, colpo di scena ‘coup de the´aˆtre’, mandare in onda ‘to broadcast’), some other (much) less so (e.g. prendere le distanze ‘to distance (oneself)’, andare in pensione ‘to retire’, di servizio ‘service (adj.)’), their specialty lying more in their familiar, conventional status (e.g. sapere benissimo ‘to know (damn) well’, essere favorevole ‘to be in favour’, nella storia ‘in history’). Still others may have more than one meaning, with different degrees of figurativity (e.g. mettere in scena, which can mean both ‘to stage’ and ‘to enact’).
From a formal point of view, some look rather
fixed and do not admit lexical insertion (e.g. vero
e proprio ‘proper’) or inflection (e.g. tra l’altro
‘by the way’, ordine del giorno ‘agenda’), whereas
others seem more flexible (e.g. essere certo ‘to
be sure’, andare bene ‘to be OK, to go well’,
posto di lavoro ‘workplace’). MWE variability
is one aspect that we did not address here but
definitely deserves to be investigated more
thoroughly (cf. e.g.
We also aim at broadening this initial list by
exploring more candidates from the CombiNet data,
which are obviously still rich of relevant
material. This first release, although limited, is
meaningful since it is the first list of commonly used
MWEs available for the Italian language, except
for domain-specific resources such as PANACEA