In this paper, we describe the process of linking the corpus CLaSSES (which collects non-literary Latin texts of diferent periods and places) to the LiLa Knowledge Base of linguistic resources for Latin made interoperable through their publication as Linked Data. The paper details the RDF modeling of the (meta)data provided by CLaSSES and presents three queries on data from diferent resources that interact in LiLa. ontologies for the representation of linguistic (meta)data. Among the resources interlinked in LiLa is the The Latin language shows a large diversity, in the light CLaSSES corpus, which enhances the set of lexical and of its wide usage in terms both of diachrony (spanning textual data made interoperable by the Knowledge Base across two millennia) and diatopy (all over Europe and with a peculiar kind of non-literary Latin texts (such beyond). Such diversity is mirrored in the set of linguistic as inscriptions, writing tablets, and letters) written in resources currently available for Latin, ranging from col- diferent periods and provinces of the Roman Empire, lections of literary texts of the Classical era,1 to corpora of thus contributing to extend the coverage of LiLa with a documentary texts of the Medieval times,2 dictionaries,3 typology of texts not present so far in the Knowledge and glossaries.4 Base. Like for many other languages, one limitation that af- This paper details the process of linking CLaSSES to fects the wealth of resources for Latin is their sparseness, LiLa, and is organized as follows. Section 2 presents the which prevents the full exploitation of the data they pro- corpus CLaSSES. Section 3 describes the LiLa Knowledge vide. The LiLa Knowledge Base was built to overcome Base. Section 4 discusses the modeling and the linking of such limitation, making distributed resources for Latin CLaSSES into LiLa. Section 5 reports three examples of interact through their publication as Linked Data, by us- queries that exploit the interoperability of CLaSSES with ing a set of commonly used vocabularies provided by other resources in LiLa. Finally, Section 6 provides some conclusions on the results of the linking, and outlines directions of future work.
CLiC-it 2023: 9th Italian Conference on Computational Linguistics, Nov 30 — Dec 02, 2023, Venice, Italy * Corresponding author. $ irene.defelice@edu.unige.it (I. De Felice); lucia.tamponi@fileli.unipi.it (L. Tamponi); 2. CLaSSES federica.iurescia@unicatt.it (F. Iurescia); marco.passarotti@unicatt.it (M. Passarotti)
© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License CLaSSES (Corpus for Latin Sociolinguistic Studies on
EpiACPWrElooUrckRteshdohinpgs oIhStpN:/uc1e6u1r3-w-0s.o7gr3ghACtttEhribUiustRiopnWa4.p0oIenrrtekrwnsahtaioosnpalc(PCoCrnoBcYcee4i.0ev).deidngasn(dCdEiUscRu-sWseSd.ojrogi)ntly by the gorraaptohriyc otefxPtSh)oinseaticdsigaintadl PrhesoonuorlcoegycraetaPteidsabUyntihveerLsaitby-.
authors, solely for academic purposes scientific responsibility is to Freely accessible on the internet,5 it consists of over 3,400
be divided up as follows: I. De Felice wrote Sections 2 and 4.2; L. non-literary Latin texts such as inscriptions, private
letTamponi Sections 5, 5.1, 5.2, 5.3; F. Iurescia wrote Section 4.1; M. ters, ink tablets, ostraka and papyri from various
periaPuasthsaorrost.ti wrote Sections 1 and 3. Section 6 is to be attributed to all ods (6th century BCE to 6th century CE) and regions
1Such as the LASLA corpus: https://www.lasla.uliege.be/cms/c_ of the Roman Empire. The goal of CLaSSES is to use
8570411/fr/lasla-textes-latins. non-literary texts that exhibit (ortho-)graphic variants
2Such as the corpus of Computational Historical Semantics: https: as a source to study the sociolinguistic variation of Latin
3/S/uwcwhwas.cothmepbhiilsintsgeuma.loLrga/thino-mEne.ghltimshl.dictionary curated by Ch. T. [
Lemmas Lexical Entries
Tokens
NLP Output Lexical Resources - Latin Wordnet - Valency Lexicon - Dictionaries...
Textual Resources - Digital libraries - Treebanks - Textual corpora...
NLP Tools - Tokenizers - Taggers/parsers - Lemmatizers...
sumption that strikes a good balance between feasibility The aim of the “LiLa - Linking Latin” ERC project (2018- and granularity: textual resources are made of (occur2023)6 was to reach interoperability between the wealth rences of) words (“tokens”), lexical resources describe of existing lexical and textual resources that have been properties of words (in “lexical entries”), and NLP tools developed in the last decades for Latin. One of the main process words (producing “NLP outputs”).8 problems that LiLa solved is the fact that such resources The core of the Knowledge Base is the so-called Lemma and tools are often characterized by diferent conceptual Bank,9 a collection of about 200,000 Latin lemmas taken and structural models, which makes it dificult for them from the database of the morphological analyzer LEMLAT to interact with one another. [12]. Interoperability is achieved by linking all those
To this goal, LiLa undertook the creation of an open- entries in lexical resources and tokens in corpora that ended Knowledge Base, following the principles of the point to the same lemma.
Linked Data paradigm.7 All content involved or
referenced in the linguistic resources connected in LiLa is 4. CLaSSES into LiLa
made unambiguously findable and accessible by
assigning an HTTP Uniform Resource Identifier (URI) to each
data point. Data reusability and interoperability between 4.1. Modeling (Meta)data
resources are achieved by establishing links between dif- The Lemma Bank of the LiLa Knowledge Base is
modferent URIs and by using web standards such as: [a] the eled as a collection of Lexical Forms of Ontolex-Lemon.
RDF data model, which is based on triples: (i) a predicate- Lexical Forms are the inflected forms of Lexical
Enproperty connects (ii) a subject (a resource) with (iii) its tries and are assigned one, or more graphical variants
object (another resource, or a literal) [
Furthermore, the LiLa Knowledge Base makes refer- ontolex:canonicalForm, to model that it is the form
ence to classes and properties of already existing ontolo- that is canonically chosen to represent the entire lexical
gies to model the relevant information. The main ones entry, i.e., the lemma. As a consequence, the Lemma Bank
are POWLA for corpus data [
Within this framework, LiLa uses the lemma as the Forms that can be used as Canonical Forms in the remost productive interface between lexical resources, annotated corpora and NLP tools. Consequently, the archi- 8In Figure 1 the arrows going from and to the node for “NLP Output” tecture of the LiLa Knowledge Base is highly lexically based (Figure 1), grounding on a simple, but efective asrepresent the fact that tokens that are the output of a specific NLP tool (a tokenizer) become the input of further tools (like, for instance, a syntactic parser). 9http://lila-erc.eu/lodview/data/id/lemma/LemmaBank. 10http://www.w3.org/ns/lemon/ontolex#writtenRep.
7https://www.w3.org/DesignIssues/LinkedData.html. sources for Latin to be interlinked in the LiLa Knowledge Base.
In particular, textual resources are connected to the Lemma Bank through the property lila:hasLemma,11 which links a token in a corpus with its lemma in the Lemma Bank. In LiLa, textual resources are modeled as objects of the type Corpus from the POWLA ontology .12 Each Corpus includes one, or more powla:Document,13 which are the parts in which the corpus is divided, like for instance the diferent texts that it contains, or its sections. In the case of the Corpus entitled CLaSSES, there are 10 documents, corresponding to as many sections of the resource.14 Every document of CLaSSES is assigned two layers, namely (1) a Document Layer, which collects all the tokens of a section, and (2) a Citation Layer, which records the full citation path of each token of a section.
For instance, Figure 2 shows the modeling of one token from CLaSSES. The token (sacra) is linked to its lemma in the Lemma Bank (sacer) by the lila:hasLemma property, and to the Document Layer by the POWLA:hasLayer property.15 The properties lila:isLayer,16 lila:hasCitSubUnit17 and POWLA:hasChild18 link the Citation Layer to the token. In the example of Figure 2, the token sacra occurs in the inscription number 27 of volume S of the Document entitled Inscriptiones latinae liberae rei publicae, to which both its Document and Citation Layers are linked through the property POWLA:hasDocument.19
assigned a lemma are linked to the Lemma Bank of LiLa. Around 14k tokens of CLaSSES are not lemmatized due to the fragmentary nature of the texts contained therein. By exploiting the original lemmatization of the corpus, the automatic linking of the tokens of CLaSSES resulted in the following three output categories.
items), i.e. whenever the lemma-PoS couple in CLaSSES was linked to one single lemma-PoS couple in the LiLa Lemma Bank. For such cases, we conducted an evaluation of the mapping on 10% of the couples. The data were randomly selected; to ensure that the sample was representative, the original PoS distribution was maintained. 11https://lila-erc.eu/ontologies/lila/hasLemma. 12http://purl.org/powla/powla.owl#Corpus. 13http://purl.org/powla/powla.owl#Document. 14http://lila-erc.eu/data/corpora/CLaSSES/id/corpus. 15http://purl.org/powla/powla.owl#hasLayer. 16https://lila-erc.eu/lodview/ontologies/lila_corpora/isLayer. 17https://lila-erc.eu/lodview/ontologies/lila_corpora/ hasCitSubUnit. 18http://purl.org/powla/powla.owl#hasChild. 19http://purl.org/powla/powla.owl#hasDocument.
thus the evaluation was conducted on 349 couples.
Only 7 errors were found, all due to a wrong PoS tagging in the source data that caused a mapping error. Thus, the rate error is very low, i.e., 2%. 2. No match (or one-to-zero lemma; 5,366 items), i.e.
when the lemma in CLaSSES was not associated with any lemma in LiLa. In this case, with the addition of the new lemma in LiLa we have enriched the Lemma Bank. Proper names are the category more afected, since inscriptions typically feature a wide range of anthroponyms which can identify the committee of the text (e.g., in public texts), the honorand (e.g., in sacred inscriptions) or the name of the dead on epitaphs [13]. In addition, given the wide geographical extension of our corpus, CLaSSES features local proper names typical of specific areas (e.g., Sardinia, or Roman Britain) that do not occur easily in Classical texts; an example from Sardinia [14, 15] is Scribonissa in ANRW-B61-6 [15, p. 45]. A few lemmas pertaining to other parts of speech were also added to the Lemma Bank, consisting mainly of hapax, like ansata in BTT-196-47 (lemma ansatus ‘provided with handles’),20 infrascribo in CEL-I-232-8 (lemma infrascribo ‘to write lower down’),21 internumero in BTT-645-48 (lemma internumero ‘to reckon among other things’).22 3. Ambiguous match (one-to-many lemmas; 1,503 items), i.e. when the lemma in CLaSSES was associated with several possible lemmas in LiLa. In most cases, the correct lemma between two or more possible ones was identified manually by a 20http://lila-erc.eu/data/id/lemma/89148. 21http://lila-erc.eu/data/id/lemma/142756.
22http://lila-erc.eu/data/id/lemma/142757. disambiguation based on the linguistic context of Lemma Bank. This is particularly useful for both quanthe document; this happens, for instance, in the titative and qualitative linguistic analysis. For example, case of homographs, as for the word dico, linked among the forms found in CLaSSES, it is possible to find both to d˘ıco, ‘to proclaim’ or ‘to dedicate’23 and occurrences of the same lemma either in the form with to d¯ıco, ‘to name’, ‘to utter’.24 On rare occasions a single consonant or with a double consonant – such (29 tokens), it was however not possible to disam- as the name Mummius in the tituli mummiani, which biguate between the lemmas available in LiLa: as is displayed either in the forms Mumius, in CIL-I2-628, a consequence, we linked the ambiguous tokens or Mummius, in CIL-I2-627, 629 [13]. The presence of to all their corresponding lemmas. This was due the alternation between <C> and <CC> in these into the fragmentary nature of some texts, where scriptions can be interpreted as a sign of an incomplete an analyzable context for disambiguation was not generalization of consonant doubling at this stage. Howavailable. This is the case, for example, of BTT- ever, it is fundamental to exclude the possibility that 609-16 mallus (context: [...] mallus alu[...] [...]us), the form Mumius occurring in our corpus represents a for which two senses are equally possible, that of commonly attested variant of the proper name Mummius. ‘pole’25 and ‘appletree’.26 This information is not readily retrievable in the available sources, since such spelling variants of proper names are generally not recorded in the dictionaries. However, by 5. Querying CLaSSES in LiLa collecting the occurrences of the lemma Mummius in the textual resources interlinked through LiLa, it is possible Thanks to the interoperability of CLaSSES with the other to ascertain that the variant without consonant doubling resources for Latin linked to the LiLa Knowledge Base, is never attested in any of the texts provided by such research questions related to non-literary texts can be resources (e.g., Cicero’s De Lege Agraria, In Verrem and empirically investigated on the several diferent textual Tacitus’ Annales, included in the LASLA corpus).28 Thus, resources interlinked in the Knowledge Base by running we may assume that the form Mumius found in CIL-I2queries on the SPARQL endpoint of LiLa27. By focusing 628 is a hint of the incomplete generalization of geminatio on the question of spelling variants attested in the in- consonantium, in line with the chronology proposed in scriptions, in what follows we shortly consider two case the literature. studies, i.e., consonant doubling (see 5.1), and the writing of long /i:/ through the diphthong <EI> (see 5.2). Moreover, we report and briefly discuss a query that exploits 5.2. <EI> for /i:/ the information on derivational morphology recorded in the Lemma Bank (see 5.3).
of LiLa can also shed light on the writing of /i:/ through <EI> in Latin sources. It is known from the literature 5.1. Consonant doubling [19, 20, 21, 22] that, in the ‘urban’ Latin of the city of Rome, the monophthongization of the diphthong [ej] As is known, the spelling of Latin long consonants took place in two steps: (i) [ej] > [e:],29 between the 3rd through geminatio consonantium was introduced at the and mid-2nd century BCE; (ii) [e:] > [i:], between the 2nd end of the third century BCE [16, 17, 18]. Consonant dou- and 1st century BCE. The data from CLaSSES, obtained bling, however, generalized slowly, so it was seldom omit- through the function "Search for linguistic phenomena" ted in the second century BCE in inscriptions. For exam- (label "Diphthong - Classical <I> /¯ı/ = <EI>"), confirm ple, in the 2nd-century inscriptions included in CLaSSES, the traditional picture, indicating that the spelling <EI> 28 tokens (20 lemmas) display single for double conso- for /i:/ is either a conservative spelling retained in earlier nants over 72 spellings with geminatio consonantium. documents, or an archaizing feature that characterizes These tokens can be easily retrieved through the func- the solemn register of later public and oficial inscription "Search for linguistic phenomena" available in the tions. More in detail, in CLaSSES the spelling <EI> for CLaSSES online search interface, by selecting the label /i:/ is found in 225 occurrences (99 lemmas), mainly in "single pro double consonant". Thanks to the interoper- older public inscriptions, before the 1st century BCE (212 ability between distributed resources provided by LiLa, occurrences over 225). A more comprehensive view of it is possible to search the occurrences of the lemmas this phenomenon can be obtained thanks to the interopfor these tokens in the corpora interlinked through the erability between diferent Latin corpora made possible in LiLa. By running a query on the corpora interlinked in the Knowledge Base, it is possible to collect all the tokens linked to the 99 lemmas concerned and select those 23http://lila-erc.eu/data/id/lemma/99301. 24http://lila-erc.eu/data/id/lemma/99302. 25http://lila-erc.eu/data/id/lemma/111421. 26http://lila-erc.eu/data/id/lemma/111423. 27https://lila-erc.eu/sparql/. 28https://lila-erc.eu/data/corpora/Lasla/id/corpus. 29Possibly a long lax [i:] [20, 23]. where the spelling <EI> for /i:/ takes place.
For instance, of particular interest is the form sei for s¯ı ‘if’ that is found in Archaic Latin. By using LiLa, it is possible to find that out of the 22,161 occurrences of si in the corpora interlinked therein, 10 show the form sei. One relevant example is from Plautus’ Epidicus (Ep. 567, twice). These 2 occurrences of sei, which in LiLa are recorded as 2 tokens from the LASLA corpus, testify to the above-mentioned first step of the monophthongization process ([ej] > [e:]), which takes place in the age of Plautus and which is attested elsewhere in his works.
• con-, 440 occurrences (6.6%).
LASLA, in which only con- is among the most frequent afifxes, with 32,763 occurrences (7,9%), whereas -in counts just for 1.2% of all afixes extracted from the corpus (5,137 occ.), -(t)or for 2.3% (9,593 occ.), and -t for 1% (4,024 occ.).
Such diferences are largely due to a number of lexemes that are highly frequent in epigraphic texts, in particular dominus ‘master’ (198 occ.) for the sufix -in and imperator ‘general, emperor’ (153 occ.) for the sufix -(t)or, which are most frequent in public inscriptions, or libertus/liberta ‘freedman’ (281 occ.) for the sufix -t, which is most frequent in funerary inscriptions, where the epitaph often refers to the civil status of freed slaves. Therefore, even if there is a major diference in dimension between the two corpora, a query such as the one here illustrated can bring to light specificities of the corpus CLaSSES that go beyond the lexical level and that could not be observed without comparison with other resources.
So far, we have discussed some very easy queries on specific lexical items that can be performed to compare information provided by CLaSSES to that provided by other resources. However, LiLa allows not only to explore and compare single corpora at the lexical level (via the Lemma Bank), but also to conduct in-depth linguistic analysis, concerning, for instance, morphology. For example, it is possible to compare the type and number of 6. Conclusion and Future Work afixes found in CLaSSES, investigating how many (and which type of) derivational morphemes are represented The linking of CLaSSES into LiLa represents an added in non-literary texts. The list of afixes that build up the value for both the resources. As for CLaSSES, its lexicon of CLaSSES can be accessed with a SPARQL query (meta)data are now interoperable with the other rethat retrieves all the lemmas in the CLaSSES corpus that sources interlinked in the Knowledge Base. As for LiLa, feature an afix (either prefix, or sufix) in their morpho- the non-literary texts of CLaSSES increased significantly logical form, and reports the number of their occurrences its textual coverage, both in terms of size and in terms of therein (see Listing 130). register variation.
Morphological information was not annotated in In the near future, we plan to model and interlink in CLaSSES. Thus, the link to LiLa allows to conduct more LiLa other types of metadata provided by CLaSSES, such in-depth linguistic analyisis; most importantly, it also as information about the provenance and the dating of allows users to compare diferent corpora with relation the texts. We plan to start from metadata on the time span to specific linguistic features. For instance, it is possi- of the texts, that we will model as Linked Data using data ble to investigate to what extent the derivational mor- categories and properties from the CIDOC Conceptual phology found in non-literary texts deviates from that Reference Model.31 of Classical texts by performing the very same query on the LASLA corpus, by simply replacing the URI for CLaSSES in the SPARQL query (as subject of the Acknowledgments powla:hasSubDocument property) with that for the LASLA corpus: http://lila-erc.eu/data/corpora/Lasla/id/ corpus.
The afixes that most frequently occur in the CLaSSES corpus are three sufixes and a prefix:
from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme – Grant Agreement No. 769994. The work of Lucia Tamponi is partly funded by the PRIN Project “Ancient languages and writing systems in contact: a touchstone for language change”, prot. 2017JBFP9H. • -in, 486 occurrences (7.3% of afixes extracted
from the corpus); • -(t)or, 456 occurrences (6.9%); • -t, 442 occurrences (6.7%); 30The query outputs a table with four columns: the label of the lemma (?lemmaLabel), the type of afix, either prefix or sufix ( ?affixType), the label of the afix ( ?affixLabel) and the total number of tokens for the lemma in CLaSSES ((count(?tokenClasses) as ?count)). 31https://www.cidoc-crm.org/.
Listing 1: A SPARQL query on the LiLa Knowledge Base PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#> PREFIX lila: <http://lila-erc.eu/ontologies/lila/> PREFIX dc: <http://purl.org/dc/elements/1.1/> PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX powla: <http://purl.org/powla/powla.owl#> SELECT ?lemmaLabel ?affixType ?affixLabel (count(?tokenClasses) as ?count) WHERE { ?lemmaLiLa a lila:Lemma ; (lila:hasPrefix|lila:hasSuffix) ?affix ; rdfs:label ?lemmaLabel . ?affix rdfs:label ?affixLabel ;
rdf:type ?affixType . ?tokenClasses lila:hasLemma ?lemmaLiLa ; powla:hasLayer ?docLayer ; rdfs:label ?tokenClassesLabel . ?docLayer powla:hasDocument ?subDocLayer . ?subDocLayer dc:title ?titlesubDocLayer .
<http://lila-erc.eu/data/corpora/CLaSSES/id/corpus> powla:hasSubDocument ?subDocLayer . } GROUP BY ?lemmaLabel ?affixType ?affixLabel ORDER BY DESC(?count)
Bretschneider, Roma, 2019, pp. 13–53. [18] L. Tamponi, La geminatio consonantium: studio su un corpus di epigrafi latine anteriori al I secolo d.C.,
Studi e Saggi Linguistici 60 (2022) 29–50. [19] M. Niedermann, Phonétique historique du latin, Li
brairie C. Klincksieck, Paris, 1953. [20] M. Leumann, Lateinische Laut- und Formenlehre,
Beck, Munchen, 1977. [21] W. S. Allen, Vox latina: A Guide to the Pronunciation of Classical Latin, Cambridge University Press,
Cambridge, 1978. [22] M. Mancini, Dilatandis litteris: uno studio su Cicerone e la pronunzia ‘rustica’, in: R. Bombi, G. Cifoletti, F. Fusco, L. Innocente, V. Orioles (Eds.), Studi linguistici in onore di Roberto Gusmani, Edizioni dell’Orso, Alessandria, 2006, pp. 1023–1046. [23] M. Benedetti, G. Marotta, Monottongazione e geminazione in latino: nuovi elementi a favore dell’isocronismo sillabico, in: P. Molinelli, P. Cuzzolin, C. Fedriani (Eds.), Latin Vulgaire–Latin Tardif, Actes du Xe Colloque International sur le Latin Vulgaire et Tardif, Sestante, Bergamo, 2014, pp. 25–43.