The paper introduces the LiITA Knowledge Base of interoperable linguistic resources for Italian. After describing the principles of the Linked Data paradigm, on which LiITA is grounded, the paper presents the lemma-centred architecture of the Knowledge Base and details its core component, consisting of a large collection of Italian lemmas (called the Lemma Bank) used to interlink distributed lexical and textual resources. tive impact on the empirical study of the language and resource usability. Indeed, diferent resources may proWhen considering the number of digital linguistic re- vide diferent information or use diferent granularity sources, either lexical or textual, Italian is among the rich- of information about the same common object, namely est languages: e.g., at the time of writing, a search on the words, which appear as occurrences in corpora and as CLARIN Virtual Language Observatory,1 filtered for the entries in dictionaries or lexicons. Making this wealth Italian language, returns more than 8 000 results. Like of information interact represents one of today's main other high-resource languages, Italian is provided with a challenges, to best leverage the huge asset of (meta)data large set of fundamental resources, including WordNets collected over decades of work. ([1] and [2]), a few treebanks available from the Univer- As a consequence, a very active line of research cursal Dependencies collection2, historical corpora 34 and rently focuses on the so-called Linguistic Linked Open reference corpora of written (e.g., CORIS/CODIS [3]) and Data (LLOD), aiming to define common practices for the spoken language (e.g., KIParla [4]). representation and publication of linguistic resources acHowever, as is the case for many other languages, most cording to the principles of the Linked Data paradigm, linguistic resources for Italian vary in terms of data for- which underpins the Semantic Web5. mat, annotation criteria, and/or adopted tagsets. Such A recently concluded COST Action (Nexus Linvariation hinders full interaction between the (meta)data guarum6) resulted both in the creation of a large and provided by the many available resources, with a nega- cohesive scientific community and in the definition of a set of shared vocabularies for linguistic knowledge description. Some of these vocabularies have been widely applied in the LiLa Knowledge Base (KB), which is probably the main LLOD use case currently available. LiLa (Linking Latin) is a KB of Latin linguistic resources made interoperable through their representation and publication according to the Linked Data principles. Thanks to its streamlined and language-independent architecture, LiLa is today a reference model for projects aiming to achieve online interoperability between distributed linguistic resources. Building on the experience of LiLa and reusing its ar-
chitecture, the LiITA (Linking Italian)7 project has started the creation of a KB of interoperable linguistic resources for Italian published as Linked Data. This paper describes the development of the fundamental component of the LiITA KB, which consists of a collection of Italian lemmas (called the Lemma Bank) that serves as the connection point between word occurrences and their entries in the corpora and lexical resources that will be published in the KB.
Language)9 is a query language for (meta)data represented in RDF; 4. Include links to other URIs to allow people (and
machines) to discover more things.
Applying the principles of the Linked Data paradigm to
(meta)data derived from linguistic resources and
publishing them on the Web ofers several benefits [
This structuring is implemented in the form of Linked exchanged information using a common set of classes
Data, which are the pillars of the Semantic Web. Unlike and data categories defined in ontologies and
vocabua web made of hypertexts, where links are not semanti- laries [
The Linked Data paradigm is founded on four princi- This Section introduces the fundamental architecture ples defined by Berners-Lee himself 8: of the LiITA KB and details its core component, i.e., a collection of canonical forms of citations (lemmas) for the Italian language13. The base URI of the resource is http://www.liita.it/data/, a namespace we reserved by buying the domain from a registrar to use also as a URL, e.g., for the project website. 1. Use URIs as “names for things” to identify them uniquely and persistently. The “things” dealt with when handling linguistic (meta)data in Linked Data are linguistic objects, such as occurrences of words in texts, lexical entries in dictionaries, or sets of parts of speech; 3.1. The Architecture of LiITA 2. Use HTTP URIs to allow people (and machines)
to look up things on the Web; The architecture of the LiITA KB resembles that of the
3. Use standards such as RDF and SPARQL to pro- LiLa KB for Latin14, which is based on the assumption
vide useful information about what is identified that the sources of the (meta)data that the KB makes
by a URI, for the purpose of representation and re- interoperable are all related to words. These sources are
trieval of (meta)data. RDF (Resource Description linguistic resources and specifically:
Framework) [
Web. SPARQL (SPARQL Protocol And RDF Query 7http://www.liita.it/ 8https://www.w3.org/DesignIssues/LinkedData 9https://www.w3.org/TR/rdf-sparql-query/ 10http://hdl.handle.net/20.500.11752/ILC-1007 11http://hdl.handle.net/20.500.11752/ILC-66 12http://hdl.handle.net/20.500.11752/ILC-558 13https://github.com/LiITA-LOD 14https://lila-erc.eu/data-page/
Lexical entries and word occurrences coming from distributed resources are made interoperable in LiITA by linking them to their respective lemmas. This makes it possible to perform federated searches on the diferent linguistic resources that LiITA makes interoperable.
For example, one can search for all occurrences (tokens) of the same lemma in multiple textual corpora; or extract from multiple corpora all those tokens that have certain lexical properties provided by one or more lexical resources.
Given the central role played by lemmas in the architecture of LiITA, the core component of the KB is a collection of conventional citation forms (lemmas) of Figure 1: The OntoLex-Lemon model. Italian words, called the Lemma Bank.
In the LiLa KB lemmas are described with the help of
custom ontology.15 This ontology, on the one hand,
provides detailed information on some morphological and OntoLex-Lemon model.
linguistic features of the lemmas (e.g. the part of speech, In Figure 1, the Classes of OntoLex-Lemon are
graphthe gramatical gender for nouns and the inflectional class) ically represented within rectangles. The relationships
relying on the OLiA annotation model [
Lexical senses are lexicalised senses: a sense belongs exactly to one lexical entry. Semantic aspects that can 3.2. The LiITA Lemma Bank be expressed by multiple words are represented through Data modelling lexical concepts, which can therefore have more than one lexicalisation. A typical example of a lexical concept is the The Lemma Bank of LiITA consists of a collection of lem- synset in a resource like WordNet, which groups multiple mas of the Italian language, i.e., lexical citation forms words related by a conceptual synonymy relationship. adopted (more or less conventionally) in linguistic re- Forms can have one or more graphical varisources. These lemmas are the names of entries in (most) ants (written representations), represented through lexical resources and the forms chosen to gather all oc- the Data Property ontolex:writtenRep22, and currences of a particular word in (lemmatised) textual possibly one or more phonetic variants (Property resources. As mentioned above, the Lemma Bank plays a ontolex:phoneticRep23). One of these forms, the fundamental role in the LiITA KB, acting as the connec- object of the ontolex:canonicalForm Property24, tion point between entries in various lexical resources is the form that is conventionally chosen to represent and word occurrences in textual resources. the entire set of inflected forms of a lexical entry. The
Following the principles of the Linked Data paradigm, Lemma Bank of LiITA is a collection of such forms,
conceptual interoperability among the distributed re- modelled as individuals of the Class lila:Lemma25,
sources connected in LiITA is achieved by applying a which is a subclass of ontolex:Form, originally created
vocabulary for knowledge description commonly used for the LiLa project, and adopted in the LiITA Lemma
in the world of Linguistic Linked Open Data. In the spe- Bank accordingly. The lemmas of the LiITA Lemma
cific case of the Lemma Bank, this means adopting the
vocabulary defined by OntoLex-Lemon [
LiITA as a collection of lemmas used to make resources Regardless of whether two resources lemmatise particiinteroperable. ples according to diferent criteria (namely, one under the
The LiITA Lemma Bank makes textual resources for participial lemma and the other under the verbal lemma), Italian interoperable through the lila:hasLemma Prop- the two diferent lemmatisations are harmonised in the erty26, which links a token in a corpus with its lemma Lemma Bank. in the Lemma Bank. Lexical resources, on the other hand, are connected to the Lemma Bank through the Data acquisition ontolex:canonicalForm Property, which links a lexical entry in the resource to its corresponding lemma in The lemmas and PoS that constitute the Lemma Bank the Lemma Bank. is based on the lexical base of an online version of the
By using the Property lila:hasPos27, each lemma in dictionary Nuovo De Mauro32, which amounts to about
the Lemma Bank is assigned one part of speech, following 145 000 entries; out of these, 13 000 multi-word
expresthe Universal PoS tagset [
In the case of words that are assigned multiple PoS essary, as lemmatisers usually deal with single tokens. tags in lexical resources, multiple lemmas are created in About 94 000 lemmas were derived from the remaining the Lemma Bank. For instance, the word sopra ‘over’ is 131 000 entries. The most numerically abundant PoS usually assigned four PoS: preposition, adverb, adjective with which the Lemma Bank was populated are listed in and noun. Thus, four distinct lemmas are created in the Table 1.
Lemma Bank with four diferent PoS represented via the lila:hasPos Property. To harmonise diferent lemmatisation criteria that may be 56 575 Nouns found in linguistic resources, the Lemma Bank of LiITA 19 912 Adjectives includes two specific Properties. The symmetric Prop- 15 885 Verbs erty lila:lemmaVariant28 connects diferent forms of 359 Proper Nouns the inflectional paradigm of a word that can be used as 311 Adverbs lemmas. A typical case is that of pluralia tantum, which 111026 PCroonnjouunnctsions can be lemmatised either in the plural form or in the sin- 40 Prepositions gular form. This model allows, for example, for both the 58 Articles lila:Lemma pantaloni and pantalone, which are linked to each other by the lila:lemmaVariant Property.
While lila:lemmaVariant links lemmas that This population process was not an easy task for
are assigned the same part of speech, the Prop- two main reasons. Firstly, the online version of Nuovo
erty lila:hasHypolemma29 (and its inverse De Mauro is tailored for visualisation: data is mixed
propertylila:isHypolemma30) connects lemmas with graphical information. Secondly, Nuovo De Mauro
that can be used for the same word but have diferent stems from one of the greatest eforts in Italian
lexicoparts of speech. This is the case for the adjectives used graphic history, namely GRADIT (Grande dizionario
italas adverbs, e.g. veloce which can be interpreted (and iano dell’uso [
Past participles are another kind of hypolemma (e.g. ples, etc.) but also frequency, semantic domain, grouping caduto ‘fallen’), which in the Lemma Bank are assigned of senses, multi-word expressions and more. The extraction of data is in practice hindered by information that must be filtered out because it is not relevant for our purposes of building a lemma bank or is provided in some non-homogeneous forms. Therefore, in order to ease this 26http://lila-erc.eu/ontologies/lila/hasLemma 27http://lila-erc.eu/ontologies/lila/hasPOS 28http://lila-erc.eu/ontologies/lila/lemmaVariant 29http://lila-erc.eu/ontologies/lila/hasHypolemma 30http://lila-erc.eu/ontologies/lila/isHypolemma 31http://lila-erc.eu/ontologies/lila/Hypolemma 32https://dizionario.internazionale.it/. PoS tags were converted automatically into the Universal tagset, adopted in the Lemma Bank. initial work, we decided to preliminary extract the afore- eased by a graphical interface which will help with the mentioned PoS, leaving out a part of the minor lexical task of writing complex SPARQL queries. categories like acronyms (e.g. NASA, FBI ), exclamation Finally, given its language-independent architecture marks, or unit symbols (e.g. cm, kg) setting them aside and the use of common vocabularies for knowledge defor future developments of LiITA. scription, LiITA promises to have a substantial method
For the time being, the Nuovo De Mauro’s PoS cat- ological impact on how linguistic resources are published egorisation rationale was adopted with some in-house and made interoperable as Linked Data. adjustment. In fact, the Nuovo De Mauro’s PoS categorisation rationale was mapped to the UPOS tagset. The original tagging was that of the Italian grammarian tradi- Acknowledgments tion, hence we had to adapt some tags, for example conjunctions. As a matter of fact, De Mauro’s conjunctions This contribution is funded by the European Union didn’t distinguish between subordinate and coordinate, - Next Generation EU, Mission 4 Component 1 CUP so, we aligned manually each of the dictionary’s conjunc- J53D2301727OOO1. The PRIN 2022 PNRR project “Litions to the UPOS tags. For the rest of De Mauro’s PoS ITA: Interlinking Linguistic Resources for Italian we have manually found the correspondence with UPOS via Linked Data” is carried out jointly by the Università tagset. Cattolica del Sacro Cuore, Milano and the Università di Torino.
In this paper we presented the first steps towards the publication as LLOD of a collection of canonical forms of citation (lemmas) for Italian. Such Lemma Bank is the core component of LiITA, a knowledge base of interoperable linguistic resources for Italian inspired by the LiLa knowledge base for Latin. LiITA aims to compensate the current lack of interoperability between Italian resources, as well as to become the pivot to interlink all the present and future lexicons and corpora for Italian. To this aim, the Lemma Bank is modelled such that it can harmonise diferent lemmatisation criteria found in lexical and textual resources, following a bottom-up approach rather that a top-down one.
Building a Lemma Bank to make distributed resources interoperable in Linked Data is an open-ended process. As the linking of more and more resources to the KB might require the inclusion of new lemmas, the LiITA Lemma Bank will keep on growing, both through the extraction of lemmas from other lexical sources and in a resource-driven fashion.
Beside extending the Lemma Bank and linking the first
resources, the LiITA project will develop online services,
following what has been done for LiLa [