computational lexicon, lexical resources, morphology, morphological harmonization
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License will populate the morphological layer of the computatasks can take advantage of lexical resources, for exam- tion of a new computational lexicon for the Italian lanCLiC-it 2023: 9th Italian Conference on Computational Linguistics, sion of the morphological layer, carried out through the
A significant number of digital lexical resources are
available for many languages. In CLARIN Virtual Language
Observatory (VLO)1, a search for “lexicalResource” of
Italian provides 52 results. Two resources appear in
several versions and updates: Parole-Simple-Clips (PSC)2
[
However, ICWs includes hundreds of thousands of forms generated automatically and not manually revised, ∗Corresponding author. †These authors contributed equally. nEvelop-O LGOBE
0000-0002-7998-9768 (F. Sciolette); 0000-0003-4320-6466 (S. Marchi); 0000-0002-0716-1160 (E. Giovannetti) CEUR Workshop Proce dings htp:/ceur-ws.org ISN1613-073
CEUR
Workshop Proceedings (CEUR-WS.org) 3https://dspace-clarin-it.ilc.cnr.it/repository/xmlui/handle/20.500. 11752/ILC-977 4https://lindat.mff.cuni.cz/repository/xmlui/handle/11372/LRT-2893 5https://lindat.mff.cuni.cz/repository/xmlui/handle/11372/LRT-2894 CEUR
ceur-ws.org which, despite being morphologically correct, have very low, if not zero, usage frequency.
Although not listed in Clarin’s VLO, we also mention
SIMPLELex-it, since built similarly to our lexicon by
combining various existing resources [
Even in lexical resources which have been manually developed and revised, however, the linguistic coverage of entries can pose problems, both in terms of lexical coverage and content of entries. Hence, integrating information from diferent sources, as we did in this work, can be efective in filling the gaps, though it can present several challenges in terms of harmonization of distinct formats and models.
guage that we called CompL-it. We started from the
enrichment of an existing resource, LexicO6 [
In particular, this first phase was focused on the
expanrated values (or TSV)8.
integration of two other resources: a list of lemmatized
forms generated by the morphological analyzer MAGIC7
[
This core of forms, lemmas, and morphological traits
tional lexicon CompL-it under construction, which will
later be released in the form of Linguistic Linked Open
6https://dspace-clarin-it.ilc.cnr.it/repository/xmlui/handle/20.500.
7https://dspace-clarin-it.ilc.cnr.it/repository/xmlui/handle/20.500.
11752/ILC-1002
8https://github.com/klab-ilc-cnr/CompL-it_morphological_layer
Data (LLOD) (see Section 5). For this reason, it was cho- the following treebanks: i) ISDT [
morphological layer
The sources we considered for building the morphological layer of CompL-it difer from each other for model, 3.1. Harmonization vocabularies, and aims; for this reason, it was first neces- Morphological data are represented in the considered sary to carry out a harmonization process to make the resources in diferent ways. The vocabulary labels of each resources comparable to each other, as shown in Sec- resource was mapped into LexInfo11, the data category tion 3. ontology for OntoLex-Lemon model12, de facto standard
Regarding the choice of sources, we opted to include for representing lexical resources in the Semantic Web.
only resources for which manual revision was docu- In the case of M-GLF and LexicO, it involved the direct
mented. In this sense, we chose not to delve into the conversion of their custom tagsets - specific for Italian
data at this initial stage. Corrective actions, aimed at into the nomenclature of LexInfo.
preventing the generation and propagation of errors, fo- The LexicO and M-GLF vocabularies also follow a
difcused on issues that could be resolved through automatic ferent theoretical approach compared to the UD used
processes and were independent of data evaluation, such in treebanks. In the first two cases, the vocabulary is
as redundancy or the comparison of entries to assess designed for lexical resources, and the POS tags are
fineinformation richness (see 3.3). grained, often finding a direct counterpart in LexInfo, as
LexInfo serves as an ontology for this type of resource. In
2.1. LexicO the case of UD, used for corpus annotation, word
descriptions are assigned a “universal” POS tag, further specified
LexicO is available on CLARIN as a relational database, by features defined in the Universal Features vocabulary.
and shares the same linguistic model of PSC, which is The cases addressed in the mapping can be classified
based on the theory of Generative Lexicon by James into three types: i) perfect correspondence; in these cases,
Pustejovsky [
MAGIC is a morphological analyzer which includes three modules: a lexicon compiler for Italian, the morphological analyzer itself, and the morphological generator.
With an ad hoc script, we extracted all forms generated by the morphological analyzer. The generated output consists of a series of linguistic objects called “words”, for each of which lemmas, morphosyntactic types, and features are specified. This resource was made available on CLARIN as “MAGIC - Generated Lemmatized Forms” (M-GLF).
Treebanks are collected and listed in the Universal Dependencies (UD) repository9. We excluded non-manually revised treebanks from the selection. Additionally, we excluded treebanks aimed at representing specific case studies that could introduce sparsely attested forms into the lexicon or introduce excessive “noise”. We considered
Once the vocabularies were harmonized, each resource was converted into a file in CoNLL-like format.
The choice of this format is primarily due to two reasons: i) UD treebanks are already in tabular format (CoNLL is a TSV); ii) the information from LexicO and M-GLF does not have a specific output format (the former is stored in a relational database, while the latter is in a textual format that does not adhere to any standard) and can be easily transformed into a TSV format. 10https://www.clarin.eu/parlamint 11https://github.com/ontolex/lexinfo 12https://www.w3.org/2016/05/ontolex/ 13https://github.com/acoli-repo/olia 14https://github.com/klab-ilc-cnr/Tables-for-mapping-of-Italian-Lexicon-ComplIt
In a first phase, from each of the aforementioned re- Similarly, Figure 2 shows the distribution of lemmas sources, a list of forms with lemmatization and morpho- per resource. logical traits was extracted. Subsequently, the obtained lists were converted in distinct CoNLL-like files, using ad hoc developed Perl scripts. In this phase, the tagsets were also converted according to the mappings in LexInfo mentioned in the previous section.
The merging process of the three resources represented by the CoNLL-like files was divided into two phases: i) initially, two resources were compared and combined in a partial merge; ii) subsequently, the third resource was added to the comparison to obtain the final output. The algorithm compared two entries at a time. If two entries were equal in terms of form, POS, and lemma, then their morphological features were compared. If the features of the first entry constituted a subset of those belonging to the second entry, the latter was considered for the final output, being richer in linguistic information. The algorithm, developed in Java, was made available on GitHub15.
The resulting output consists of 790,758 forms associated with 102,000 lemmas and the relative traits.
In CompL-it, each form is associated with the following data: lemma, part of speech (POS), and morphological features specific to the considered POS. Table 1 shows an example of form for the lemma “gatto” (cat).
Despite the evident larger size of M-GLF compared to LexicO, it is important to specify that the choice to use this latter as the reference base was mainly qualitative, in particular for its multilevel structure16, which will be exploited for enriching CompL-it in subsequent works (Section 5).
In Table 2, the number of forms per POS in LexicO is compared to the final CompL-it resource, along with the respective percentage increase.
It is worth noting the significant increase in values, particularly for adjectives and adverbs, which have a lower coverage in LexicO17.
To conclude this section, we provide in Table 3 a quantitative comparison between CompL-it and some of the lexical resources mentioned in the introduction, specifically PSC, ItalWordNet, and SIMPLELex-IT. We excluded ICWs from the comparison due to the mentioned issue of overgenerating forms (which doesn’t align well with the need to represent lexically precise data) and IFWs, as it contains many multiword entries that we have chosen to exclude from our lexicon at this time.
In this article, we documented a first step towards building a new computational lexicon of Italian. A set of approximately 800 thousand lemmatized forms with morphological features was created, through the integration of existing resources. In the next phases, this lexical core will be converted as a LLOD based on the OntoLex-lemon model, making the resulting lexicon more easily shareable, interoperable, and compliant with Semantic Web standards. Additionally, new linguistic layers will be added, starting from semantics, by using the information already available in LexicO and by integrating data from WordNets for Italian.
This work was conducted in the context of the TALMUD project and the scientific cooperation between S.c.a r.l. PTTB and CNR-ILC.
Actionable: the Construction of the ParlaMint-IT Corpus, in: Proceedings of the Workshop ParlaCLARIN III within the 13th Language Resources and Evaluation Conference, 2022, pp. 117–124. URL: https://aclanthology.org/2022.parlaclarin-1.17/. [13] N. Ruimy, M. Monachini, R. Distante, E. Guazzini, S. Molino, M. Ulivieri, N. Calzolari, A. Zampolli, Clips, a multi-level Italian computational lexicon: A glimpse to data, in: Proceedings of the Third International Conference on Language Resources and Evaluation (LREC02), 2002.