Wikidata and LiLa for Latin: Enabling Interoperability and Access to Inflected Forms and Corpus Attestations

Discussion paper Wikidata and LiLa for Latin: Enabling Interoperability and Access to Inflected Forms and Corpus Attestations David Lindemann*a, Matteo Pellegrinib, Francesco Mambrinib, Marco Passarottib aDepartment of Linguistics and Basque Studies, EHU University of the Basque Country; bCIRCSE Research Centre, Università Cattolica del Sacro Cuore, Milan, Italy *Corresponding author: David Lindemann; [email protected] Author roles David Lindemann: Conceptualization, Methodology, Software, Visualization, Writing – review & editing Matteo Pellegrini: Conceptualization, Formal analysis, Methodology, Resources, Writing – original draft, Writing – review & editing Francesco Mambrini: Conceptualization, Methodology, Writing – review & editing Marco Passarotti: Conceptualization, Methodology, Writing – review & editing Abstract This paper presents an approach to integrating Latin inflected forms and corpus attestations within a Linked Open Data (LOD) framework, enhancing interoperability between Wikidata and the LiLa knowledge base. Building on the PrinParLat lexicon of Latin verb principal parts, we generate the complete set of inflected forms for over 8,000 verbs, encoded as RDF in a dedicated Wikibase instance. These forms are linked to the Index Thomisticus Treebank (ITTB), whose morphologically annotated tokens are related to corresponding forms based on segmental identity, lemma alignment, and mapped morphological features. Our generation and linking process achieves over 95% coverage of ITTB verbal tokens, demonstrating the robustness of our pipeline even for Medieval Latin data. By aligning Paralex, Wikidata, and LiLa ontologies, we ensure semantic interoperability and facilitate future integration into Wikidata. Beyond Latin, this workflow provides a reproducible model for linking inflectional paradigms and corpus attestations in other languages. Keywords: Wikidata; Wikibase; LiLa; Latin morphology; inflectional lexicon; linked open data (1) Context and motivation (1.1) Context: Wikidata and LiLa Wikidata (Vrandečić & Krötzsch, 2014; Erxleben et al., 2014) is a knowledge base of structured data coded according to the Resource Description Framework (RDF; cf. Lassila and Swick, 1998), following the principles of the Linked Open Data paradigm (LOD; cf., with a focus on linguistic data, Cimiano et al., 2020). Each item in Wikidata has its own Uniform Resource Identifier (URI), and information about it is stored as RDF statements that connect the item to a value using a defined property. Several classes and properties – also with dedicated URIs – are defined to allow for the representation of the variegated information recorded in the original data. Among else, rich information is provided on lexemes using the OntoLex-lemon vocabulary (McCrae et al., 2017), by now a de facto standard for the modelling of lexical information as RDF LOD. Following this standard, on Wikibase, lexemes are assigned the class ontolex:LexicalEntry,1 and are linked to their senses (instances of ontolex:LexicalSense) through the property ontolex:sense, and to their forms (instances of ontolex:Form) through the property ontolex:lexicalForm (cf. Lindemann, 2025 for a more detailed discussion on the implementation of OntoLexLemon in Wikibase). Furthermore, in line with the spirit of LOD, Wikidata items can also feature links to identifiers of the corresponding items in external identifiers. For instance, for Latin lexemes, the URI of the corresponding lemma in the knowledge base of the Linking Latin (LiLa) project (Passarotti et al., 2020) is given. From there, many other pieces of information can be retrieved on that lexeme, since LiLa acts as a hub to make lexical and textual resources on Latin available as RDF LOD, and hence interoperable. To achieve interoperability between resources available for Latin, a lemma bank was created that currently2 includes 230,402 forms that can be used as lemmas for Latin words, and both tokens of textual resources and entries of lexical resources are linked to the corresponding lemma in there. To date, the lemma bank contains links to a few digital dictionaries (such as the Lewis and Short Latin-English dictionary, cf. Mambrini et al., 2021a; the Lexicala Latin-French dictionary, cf. De Paoli et al., 2025; Velez’s Latin-Portuguese dictionary, cf. 1 See the Appendix for a list of the namespaces of compact URIs used in this paper. 2 As of November 3, 2025. The same holds for the other counts shown in this paper. Dezotti et al., 2024), other lexical resources (such as a Wordnet, cf. Mambrini et al., 2021b; a sentiment lexicon, cf. Sprugnoli et al., 2021; a word-formation lexicon, cf. Pellegrini et al. 2022), as well as corpora pertaining to different epochs and genres (such as the LASLA corpus for the preand post-classical period, cf. Fantoli et al., 2022; the CompHistSem corpus for the late and medieval period, cf. Pedonese et al., 2023; the CLaSSES collection of inscriptions, cf. De Felice et al., 2023). To date, almost 12M corpus tokens are linked to the LiLa knowledge base. Thanks to the presence of links between Wikidata lexemes and LiLa lemmas (Lindemann et al., 2023),3 the two projects already stand out as a virtuous example of the integration of data from different sources guaranteed by the RDF technology and the LOD paradigm. In this paper, we discuss an enhancement of this integration, concerning inflected forms of lexemes and their attestation in corpora. (1.2) Motivation: other forms and their occurrence in texts Inflectional lexicons are lexical resources that list (potentially all) inflected forms that are available for each lexeme, and code the cell that forms occupy in a lexeme’s paradigm – i.e., the morphosyntactic properties that they convey. Such resources are increasingly being developed, because of their potential use for both theoretical and applied linguistics. On the one hand, they can be used to train and test Natural Language Processing (NLP) systems addressing specific tasks: for instance, the Unimorph project (Kirov et al., 2016; Batsuren et al., 2024) scraped Wiktionary data to obtain lexicons consisting of formlemma-cell triples in tabular format for a wide range of languages (to date, 169); those lexicons have been used to train and test systems for the tasks of morphological inflection (predicting inflected forms from lemmas), reinflection (predicting inflected forms from each other), analysis and other shared tasks proposed in various editions of the Special Interest Group on Morphology and Phonology of the Association for Computational Linguistics (SIGMORPHON, cf. Cotterell et al., 2016, Nicolai et al., 2024). On the other hand, inflectional lexicons have been widely used in the literature on theoretical morphology, to perform systematic quantitative analyses of predictability in implicative relations between forms (see, e.g., Pellegrini, 2023 on Latin, Herce, 2025 on Spanish, and Beniamine, 2018 on a small but typologically diverse sample of languages). The Paralex project (Beniamine et al., 2023), an effort towards the standardisation of inflectional 3 See https://www.wikidata.org/wiki/Property_talk:P11033. lexicons more complex than the ones released in the Unimorph project, allows for the creation of lexicons that include phonological transcriptions that are crucial for the quantitative analyses mentioned above. Besides a recommended standard format, Paralex features an OWL (Web Ontology Language) ontology that introduces classes and properties that can be used to release those lexicons as RDF LOD (cf. Pellegrini et al., 2025). As hinted above, for many lexemes, Wikidata lists sets of inflected forms, and the morphosyntactic properties they convey, thus essentially providing users with inflectional data in RDF for many languages, including Latin. However, no information is provided on the actual usage of Latin forms, as reflected in corpus attestations. Indeed, especially in large paradigms, like the ones of Latin verbs, it will often be the case that only a few forms for each lexeme are actually found in texts, while many others never appear (Bonami & Beniamine, 2016). This information can be obtained from LiLa, looking at the attestations of such forms in the corpora linked to the lemma bank that feature a fine-grained annotation of morphological features. This is the topic that we aim to address in this paper: we discuss our efforts to map corpus tokens linked to the LiLa lemma bank to forms of lexemes expressed in RDF in the same format as in Wikidata, to allow for a distinction between forms that are possible, but never used in LiLa corpora, and forms that are attested in texts. (1.3) Content of this work Rather than working on Wikidata Latin lexemes directly, we create our own Wikibase instance.4 This allows us to identify potential issues and test different options to solve them. Furthermore, it makes it possible to showcase these different options in Wikidata community discussions about the data model to adopt towards Latin lexemes, and about how to overcome some limitations of the Wikidata lexicon, such as the following ones. On the one hand, there are gaps in the coverage of Wikidata with respect to tokens in LiLa corpora: missing lexemes, such as PRAEEXISTO ‘pre-exist’; lexemes for which at present no forms are listed, such as LAUO5 ‘wash’; specific forms that are systematically not listed, such as contracted perfects, e.g., PRF.ACT.INF praecipitasse for PRAECIPITO6 ‘precipitate’; 4 See https://lilamorph.wikibase.cloud. The instance is hosted by the Wikibase Cloud hosting service provided by Wikimedia Deutschland, see https://wikibase.cloud. 5 See https://www.wikidata.org/wiki/Lexeme:L1056162. 6 See https://www.wikidata.org/wiki/Lexeme:L281899. and other less systematic gaps. On the other hand, Wikidata forms are sometimes underspecified for some of the inflectional properties that they convey, such as gender. For instance, for the perfect participle of the verb CANDIDO7 ‘make white’, the form candidatis is only coded as being the dative plural, with no information on gender, since that form can convey masculine, feminine or neuter gender. However, tokens matching that form in texts can be specified for gender, according to the gender of the controller noun: the adjective form will appear annotated as masculine if agreeing with a masculine noun, feminine if agreeing with a feminine noun, neuter if agreeing with a neuter noun. Some Latin corpora linked to LiLa – e.g., the Index Thomisticus Treebank (Passarotti, 2019) – accordingly tag the token as either masculine, feminine or neuter. If such tokens were linked to the underspecified form recorded in Wikidata, there would be loss of information, that would make it impossible to check, for instance, the frequency of that form occurring as masculine, feminine or neuter. Also, a direct link added to the token pointing to the form URI would lose granularity in that regard. In general, a more granular approach, representing orthographically identic but morphologically ambiguous forms as separate entities, makes querying for specific forms more straightforward. To overcome these limitations, rather than using Wikidata forms, we exploit another inflectional resource linked to LiLa, PrinParLat (Pellegrini et al., 2025), to generate full paradigms including fully specified inflected forms, and link those forms to fully specified annotations of tokens in corpora. Based on that data, we could for example propose to represent all attested forms in Wikidata, while morphologically possible but unattested forms would remain unlisted. Since at the time of the writing of this paper PrinParLat only included verbs, in the experiments presented here, we only work on verb paradigms. As reference corpus, we use the Index Thomisticus Treebank, as it is the only corpus linked to LiLa (Mambrini et al., 2022) that features an annotation of morphosyntactic properties as fine-grained as for distinguishing, for example, masculine, feminine and neuter forms with identical orthographical representation. (2) Dataset description Repository location – https://doi.org/10.5281/zenodo.17553591 7 See https://www.wikidata.org/wiki/Lexeme:L262071. Repository name – Zenodo Object name – Lilamorph Wikibase: Latin Inflected Forms and Corpus Attestations Format names and versions – CSV, python 3.12 Creation dates – 01-09-2025 to 07-11-2025 Dataset creators – David Lindemann (programming), Matteo Pellegrini, Francesco Mambrini and Marco Passarotti (source data) Language – Metalanguage: English. Object language: Latin License – GNU General Public License v3.08 Publication date – 07-11-2025 (3) Experiments (3.1) Generating forms: PrinParLatInfLexi The point of departure for the generation of forms for this work is PrinParLat (Pellegrini et al., 2025).9 It is a Paralex compliant inflectional lexicon, also released as RDF LOD and linked to the LiLa knowledge base. It lists Principal Parts of Latin verbs – i.e., sets of inflected forms of a lexical unit from which its full paradigm can be inferred (see Stump & Finkel, 2013). The cells that are used as principal parts in the resource are PRS.ACT.INF, FUT.ACT.IND.3SG (or the corresponding morphologically passive forms for deponents), PRF.ACT.IND.1SG (3SG for impersonals), PRF.PASS.PTCP.NOM.N.SG, and FUT.ACT.PTCP.NOM.N.SG. The forms of PrinParLat are assigned not only to lexemes, identified on semantic grounds (i.e., forms that share the same lexical meaning), but also to flexemes (Fradin & Kerleroux, 2003), identified on formal grounds. For instance, the verb LAVO ‘wash’ can be inflected according to either the 1st or the 3rd conjugation, and its principal parts are 8 The ITTB is under a CC BY-NC-SA 4.0 license. There is no one license common to all the resources interlinked in LiLa; as for the Lemma Bank, which is what is relevant here, it is CC BY-SA 4.0. The GNU General Public License v3.0 is the license for code that is consistent with the “as open as possible, as closed as necessary” approach pursued in LiLa and with the licenses chosen for the third-party sources mentioned here (ITTB, LiLa). 9 More specifically, the version used for this work is v2.0.0 (DOI: 10.5281/zenodo.17815898). accordingly assigned to different flexemes,10 as shown in Table 1. lexem e flexem e (conj.) PRS.ACT. INF FUT.ACT. IND.3SG PRF.ACT. IND.1SG PRF.PASS.PTCP. NOM.N.SG FUT.ACT.PTCP. NOM.N.SG ‘wash’ LAVO (1st) lauare lauabit lauaui lauatum lauturum LAVO (3rd) lauere lauet laui lautum lauturum Table 1 – Lexemes, flexemes, and principal parts for the verb LAVO ‘wash’ Thanks to this structure, it is possible to capture the systematic relationship between variants that share some formal feature – be it the conjugation, as in this example, or the stem variant on which the form is built, e.g., abalienvs. abalenfor the verb ‘separate’ (PRS.ACT.INF abalienare/abalenare, FUT.ACT.IND.3SG abalienabit/abalenabit, etc.). Furthermore, in this way, each flexeme can be assigned a unique descriptor of its inflectional behaviour (e.g., the conjugation), even if the lexeme is compatible with different conjugations – as happens for LAVO in Table 1. This allows each PrinParLat flexeme to be linked to its own corresponding lemma in the LiLa Knowledge Base – e.g., in this case, the PrinParLat flexeme pertaining to the 1st conjugation is linked to the LiLa lemma 110084,11 which is itself stated to be a 1st conjugation verb , while the PrinParLat flexeme pertaining to the 3rd is linked to the LiLa lemma 110085,12 which is itself stated to be a 3rd conjugation verb. Indeed, each flexeme of PrinParLat features tags for both its traditional coarse-grained conjugation among the ones of the classification of traditional grammars, as coded on the corresponding LiLa lemma through the property lila:inflectionType, and a finegrained inflection micro-class (Dressler et al., 2008), grouping together only lexemes that display the same inflectional behaviour in all forms (see Pellegrini et al., 2025 for further details). For instance, the verbs LAUDO ‘praise’, AMO ‘love’ and CUBO ‘lie’ are all traditionally assigned to the same conjugation, the first, with PRS.ACT.INF -are (amare, laudare, cubare) and the same set of inflectional patterns in the present system (e.g., 10 See http://lila-erc.eu/data/lexicalResources/prinparlat/id/flexeme _l0416_6480 for the flexeme of the 1st conjugation, http://lila-erc.eu/lodview/data/lexicalResources/prinparlat/id /flexeme_l0416_6479 for the flexeme of the 3rd conjugation. 11 See http://lila-erc.eu/data/id/lemma/110084. 12 See http://lila-erc.eu/data/id/lemma/110085. PRS.ACT.IND.3.SG amat, laudat, cubat). However, only LAUDO and AMO are also assigned to the same micro-class, since they also share the inflectional patterns of the perfect (e.g., PRF.ACT.INF laudauisse, amauisse), perfect participle and supine (laudatum, amatum), while CUBO is assigned to a different micro-class that displays other inflectional patterns in those cells (see PRF.ACT.INF cubuisse, SUP cubitum). In this context, we exploit the information provided by PrinParLat – namely, principal parts of flexemes and their conjugation and micro-class – to generate the full paradigms of the Latin verbs recorded in PrinParLat, including all adjectival and nominal forms (participles, gerund(ive)s, supines). This is achieved by means of rules, whose bases are specific principal parts of flexemes of specific conjugations, and whose outputs are other forms produced through replacements coded by regular expressions. Table 2 shows a sample of such rules, and their output when applied to forms of the verb LAVO shown above in Table 1. rule inflectionType base form output form replacement output (LAVO) 1 v1r PRS.ACT.INF PRS.ACT.IND.3SG are$ à at lauat 2 v3r PRS.ACT.INF PRS.ACT.IND.3SG ere$ à it lauit 3 v1r,v3r PRF.ACT. IND.1SG PRF.ACT. IND.3SG i$ à it lauauit, lauit 4 v1r,v3r PRF.PASS.PTCP. NOM.N.SG PRF.PASS.PTCP. NOM.F.SG um$ à a lauata, lauta Table 2 – Rules to generate forms of Latin verbs The outcome of the application of this process is a new resource providing information on the content of the 254 paradigm cells available for the 11K flexemes included into PrinParLat –amounting to a total of more than 2,5M cells. We call this PrinParLat-based inflected lexicon PrinParLatInfLexi and we release it on Zenodo in the Paralex community.13 13 DOI: 10.5281/zenodo.17819184. Another inflected lexicon of Latin verbs and nouns that we created in the past for other purposes is already available, LatInfLexi (Pellegrini & Passarotti, 2018). There is a remarkable amount of overlap between the two resources, so in the future we plan to merge them into a single reference resource for inflected forms in Latin. However, this cannot be done yet because there are pieces of information that are provided in only one of the two (e.g., information on flexemes in PrinParLatInfLexi only, information on vowel length in LatInfLexi only). For an ITTB token to be linked to a PrinParLatInfLexi form, the following requirements must be met. i. The token and the form must be segmentally identical. Since the ITTB tokens do not display any coding of vowel length, differently than the forms of PrinParLatInfLexi, where vowel length is coded at least on endings, we use a normalised version of the latter representation where vowel length is removed. ii. The token and the form must link to the same LiLa lemma. Only for tokens for which no direct match with any LiLa lemma could be found in PrinParLatInfLexi forms, we also checked whether a match could be found with lemmas that are stated to be variants in the lemma bank, through the property lila:lemmaVariant. For instance, the ITTB token epulemur20 is linked to the deponent lemma EPULOR.21 While no PrinParLat(InfLexi) entry is linked to that lemma, there is a matching entry linked to non-deponent EPULO,22 which in the LiLa lemma bank is stated to be a variant of EPULOR; we thus link the token to the corresponding form of that entry (epulēmur).23 20 See https://lilamorph.wikibase.cloud/wiki/Item:Q25520. 21 See http://lila-erc.eu/data/id/lemma/101349. 22 See http://lila-erc.eu/data/id/lemma/101350. 23 See https://lilamorph.wikibase.cloud/wiki/Lexeme:L29533#F152. Figure 8 – The token lavatur in our Wikibase instance24 24 See https://lilamorph.wikibase.cloud/wiki/Item:Q29683. The fine-grained morphological tags of the token and of the form must be compatible. Since the two resources use different annotation schemes for morphological properties – a custom Paralex-compliant coding using the abbreviations of the Leipzig Glossing Rules in PrinParLatInfLexi, the UD tagset in the ITTB – we needed to explicitly map the values of the two annotations to one another, as shown in Table 4.25 Note that in a couple of cases, a single tag in the PrinParLatInfLexi set corresponds to two values of the ITTB one. This is because tags like ‘future perfect’, common in traditional Latin descriptions, cannot be considered as expressing a value of a single feature, but rather distinct values of distinct features of tense and aspect. The link from the corpus token to a matching ontolex:Form is represented on our Wikibase by the property P21. Each of those links is referenced by a link to the exact version of the matching algorithm script used for producing the matches, i.e., the general matching algorithm, and the one asking for matches through lemma variants, respectively.26 feature value PrinParLatInfLexi (Leipzig Glossing Rules) ITTB (Universal Dependencies) case ablative abl Case#Abl case accusative acc Case#Acc case dative dat Case#Dat case genitive gen Case#Gen case nominative nom Case#Nom case vocative voc Case#Voc gender feminine f Gender#Fem gender masculine m Gender#Masc gender neuter n Gender#Neut mood imperative imp Mood#Imp mood indicative ind Mood#Ind mood subjunctive sbjv Mood#Sub number plural pl Number#Plur number singular sg Number#Sing 25 See https://lilamorph.wikibase.cloud/wiki/Main_Page#PrinParLat_mor phological_cell_descriptors_(Leipzig_abbreviations) for the mapping of Leipzig abbreviations to UDP features and to Wikidata entities, and, through Wikidata, to entities of the LexInfo ontology. 26 We are using Wikibase references, and as value for the reference link, the ID of the Github commit of the version used in the matching procedure, see, e.g., https://lilamorph.wikibase.cloud/wiki/Item:Q49188. person first 1 Person#1 person second 2 Person#2 person third 3 Person#3 tense-aspect fut Tense#Fut tense-aspect fprf Tense#Fut, Aspect#Perf tense-aspect iprf Tense#Imp tense-aspect prf Tense#Past, Aspect#Perf tense-aspect pprf Tense#Pqp tense-aspect prs Tense#Pres verb form gdv VerbForm#Gdv verb form ger VerbForm#Ger verb form inf VerbForm#Inf verb form ptcp VerbForm#Part verb form sup VerbForm#Sup voice act Voice#Act voice pass Voice#Pass Table 4 – Mapping between the PrinParLatInfLexi and ITTB tagsets for morphological feature values (4) Outcome and discussion As the outcome of the process described in the previous section, in our Wikibase instance we have on the one hand a collection of the 8,018 lexemes of PrinParLat, each of them supplied with a full list of generated forms from PrinParLatInfLexi (up to 1,812 per lexeme, depending on the amount of form variants available for the same cells); on the other hand, a collection of 71,195 verbal tokens extracted from the ITTB. Table 1Table 5 summarises the outcome of the process of linking tokens to forms. linked tokens unlinked tokens total 1:1 links 1:N links 67,740 (95.1%) 969 (1.4%) 2,486 (3.5%) 71,195 68,709 (96.5%) Table 5 – Outcome of the process of linking ITTB forms to PrinParLatInfLexi forms The overall coverage of the generated forms with respect to ITTB tokens is very high, above 95%. This result is even more remarkable if we consider that the data source from which our forms are generated, PrinParLat, is intended to document systematically only words that are found in Classical Latin dictionaries (cf. Pellegrini et al., 2025), and thus does not cover Medieval Latin words systematically, while our tokens come exactly from a corpus of Medieval Latin, consisting of works of Thomas Aquinas (13th century). Indeed, many of the tokens that cannot be linked are occurrences of verbs that are only attested from the Middle Ages on, and are thus not found in Classical Latin dictionaries, like CONDITIONO ‘to condition’. If we exclude such cases of tokens that are not mapped to any form because there is no lexical entry in PrinParLatInfLexi that is linked to the corresponding lemma, only 1,769 unlinked tokens remain. These can be due to forms that are not represented in PrinParLatInfLexi, like comparatives and superlatives of participles (e.g., NOM.M/F.SG convenientior ‘more convenient’, NOM.N.SG convenientissimum ‘very convenient’), adverbs derived from participles (e.g. consequenter ‘consequently’ from CONSEQUOR), or simply spelling variants (e.g., PRS.ACT.INF abiicere instead of standard abicere ‘throw away’). In other cases, the reason lies in different choices in the annotation, e.g. present participles of deponents that are tagged as passive in the ITTB vs. active in PrinParLatInfLexi. In most cases, ITTB tokens are unambiguously linked to only one form in PrinParLatInfLexi. There are cases of tokens that can be linked to different forms, but ultimately, all of these ambiguities are due to issues that are not related to our procedure to link tokens to forms. On the one hand, this can happen because of ambiguity in the linking of ITTB tokens to LiLa lemmas in the original resource: for instance, the lemmatization of the token deserere27 is not disambiguated between the two available options, namely DESERO ‘sow’28 vs. DESERO ‘abandon’.29 As a consequence, that token is compatible with forms of both corresponding lexemes in PrinParLatInfLexi. This ambiguity in the linking of tokens to LiLa lemmas explains about two thirds of the ambiguous token-to-form links in our data (628 out of 969). On the other hand, as we have seen, tokens in the ITTB are not coded for vowel length, while forms of PrinParLatInflexi are, and there are forms that only differ in vowel length. The most systematic – and hence most quantitative impactful – example is the PRF.ACT.IND.3PL cell, that can be formed by attaching to the perfect stem either -erunt or -ērunt. While these variants differ in their frequencies and contexts of occurrence, both are systematically given as possible variants for all lexemes in PrinParLatInfLexi, that aims at the maximum possible coverage. Consequently, a token 27 See https://lilamorph.wikibase.cloud/wiki/Item:Q11916. 28 See http://lila-erc.eu/data/id/lemma/98883. 29 See http://lila-erc.eu/data/id/lemma/98882. like dedicaverunt30 can be linked to two different PrinParLatInfLexi forms, dedicaverunt31 and dedicavērunt,32 without the possibility of disambiguation (except possibly for poetry, where the length can sometimes be inferred from metre). This unavoidable ambiguity due to lack of information on vowel length in corpora accounts for about one third of the cases of ITTB tokens that are linked to more than one form in PrinParLatInfLexi (343 out 969). To summarise, the coverage of ITTB tokens with our generated forms is very high, and almost no case of 1:0 and 1:N links is due to actual gaps in our generated forms or bugs in the linking process. Consequently, we consider our work to be a successful example of enrichment and integration of data for Latin. The writing operations to Wikibase described in section 3 have been carried out by own python scripts, deploying the wikibaseintegrator libary33 for the construction of JSON objects, as required by the MediaWiki API, describing Wikibase entities34 of type lexeme (for PrinParLat lexemes and flexemes) and of type item (for corpus tokens). (5) Applications and perspectives for future work Recently, we have aligned Wikidata Latin lexemes to the LiLa lemma bank and have also created new Wikidata lexeme entities.35 With the different forms lexicons built on our Wikibase instance, we are now in the position to contribute to a discussion in the Wikidata community, comparing different options of representation of inflected forms. We would like to highlight corpus token linking as central use case for Wikibase forms, which entails to adopt the data model that caters best for that application, namely a separate listing of orthographically identical but morphologically ambiguous forms. Having chosen Wikibase as platform for the experiments presented here, all datasets remain now ready for intervention of human or algorithmic users, who would mark ambiguous links (from token to form, or from token to LiLa lemma), as “preferred” or “deprecated”, so that the ambiguity is resolved. For such task, the software provides the necessary means for querying the content and introducing rank values to ambiguous 30 See https://lilamorph.wikibase.cloud/wiki/Item:Q11272. 31 See https://lilamorph.wikibase.cloud/wiki/Lexeme:L28639#F172. 32 See https://lilamorph.wikibase.cloud/wiki/Lexeme:L28639#F174. 33 See https://wikibaseintegrator.readthedocs.io. 34 See https://www.mediawiki.org/wiki/Wikibase/DataModel. 35 See https://www.wikidata.org/wiki/Property_talk:P11033. statements. These edits would be recorded in the users’ edit history, for revision and evaluation statistics.36 The data that results from this work, and the procedure outlined to obtain it, are also of interest for theoretical phonology and morphology, that can benefit from the possibility of having accurate frequency counts of wordforms. More specifically, this can be useful for quantitative studies on predictability in morphology, such as the ones cited in Section 1. Such studies normally use inflectional lexicons listing all forms of lexemes, regardless of their attestation. However, there is a growing awareness – see, e.g., Boyé and Schalchli (2019) – that a more ecological setting for the task of predicting forms should be adopted, e.g., using attested forms as predictors, and non-attested forms as the ones to be predicted, or at least taking into account information on token frequency in some way. This work provides data can be exploited for that purpose, and it outlines a procedure that can be applied to obtain more data on this aspect. Indeed, in the future, we plan to extend the application of this procedure to other lexical categories that display different inflected forms in Latin – namely, nouns and adjectives – as well as to other corpora that feature the necessary fine-grained morphological tagging – such as the LASLA corpus. Furthermore, we envisage such an effort to be undertaken also for other languages. Appendix We provide here a list of the namespaces of the compact URIs used in this paper. lila: <http://lila-erc.eu/ontologies/lila/> ontolex: <http://www.w3.org/ns/lemon/ontolex#> paralex: <https://www.paralex-standard.org/paralex_ontology.xml#> wikidata: <http://www.wikidata.org/entity/> 36 For example, at https://lilamorph.wikibase.cloud/wiki/Item:Q15797#P16, upon the decision whether that token “dico” should be linked to the first conjugation verb or the third conjugation verb (grounded, for instance, on the inflection paradigm of unambiguous forms of the same group of lemmas appearing in the same text), the user would mark the correct link as “preferred rank”, or, alternatively, the other one as “deprecated rank“, so that the matching algorithm in the next round would not again propose ambiguous token-to-form links. Competing interests The authors have no competing interests to declare. References Aronoff, M. (1993). Morphology by Itself: Stems and Inflectional Classes. Cambridge: MIT Press. Batsuren, K., et al. (2022). UniMorph 4.0: Universal Morphology. In Proceedings of the Thirteenth Language Resources and Evaluation Conference (pp. 840-855). Marseille, France: European Language Resources Association. URL: https://aclanthology.org/2022.lrec-1.89/ Beniamine, S. (2018). Classifications flexionnelles. Étude quantitative des structures de paradigms. Doctoral dissertation, Université Sorbonne Paris Cité-Université Paris Diderot (Paris 7). Beniamine, S., Anderson, C., Carroll, M., Guzmán Naranjo, M., Herce, B., Pellegrini, M., Round, E., Sims-Williams, H, & Tresoldi, T. (2023). Paralex: a DeAR standard for rich lexicons of inflected forms. In The Fourth International Symposium of Morphology. Bonami, O., & Beniamine, S. (2016). Joint predictiveness in inflectional paradigms. Word structure, 9(2), 156-182. DOI: 10.3366/word.2016.0092 Boyé, G., & Schalchli, G. (2019). Realistic data and paradigms: The paradigm cell finding problem. Morphology, 29(2), 199-248. DOI: 10.1007/s11525-018-93351. Ciccarese, P., Ocana, M., Garcia Castro, L. J., Das, S., & Clark, T. (2011). An open annotation ontology for science on web 3.0. Journal of biomedical semantics, 2(Suppl 2), S4. DOI: 10.1186/2041-1480-2-S2-S4 Cimiano, P., Chiarcos, C., McCrae, J. P., & Gracia, J. (2020). Linguistic linked data. Springer International Publishing. DOI: 10.1007/978-3-030-30225-2 Cotterell, R., Kirov, C., Sylak-Glassman, J., Yarowsky, D., Eisner, J., & Hulden, M. (2016). The SIGMORPHON 2016 shared task—morphological reinflection. In Proceedings of the 14th SIGMORPHON workshop on computational research in phonetics, phonology, and morphology (pp. 10-22). DOI: 10.18653/v1/W16-2002 De Felice, I., Tamponi, L., Iurescia, F., & Passarotti, M. (2023). Linking the Corpus CLaSSES to the LiLa Knowledge Base of Interoperable Linguistic Resources for Latin. In Proceedings of the Ninth Italian Conference on Computational Linguistics (CLiC-it 2023) (pp. 172-178). Venice, Italy: CEUR Workshop Proceedings. URL: https://aclanthology.org/2023.clicit-1.22/ De Paoli, A., Passarotti, M. C., Ruffolo, P., Moretti, G., & Kernerman, I. (2025). Linking the Lexicala Latin-French Dictionary to the LiLa Knowledge Base. In Proceedings of the 5th Conference on Language, Data and Knowledge (pp. 197-207). URL: https://aclanthology.org/2025.ldk-1.21/ Dezotti, L. C., Passarotti, M., & Mambrini, F. (2024). Modelling and Linking an Old Latin-Portuguese Dictionary to the LiLa Knowledge Base. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 11537-11547). URL: https://aclanthology.org/2024.lrec-main.1008/ Erxleben, F., Günther, M., Krötzsch, M., Mendez, J., & Vrandečić, D. (2014). Introducing wikidata to the linked data web. In International semantic web conference (pp. 50-65). Cham: Springer International Publishing. DOI: 10.1007/978-3-319-11964-9_4 Fantoli, M., Passarotti, M., Mambrini, F., Moretti, G., & Ruffolo, P. (2022). Linking the LASLA Corpus in the LiLa Knowledge Base of Interoperable Linguistic Resources for Latin. In Proceedings of the Linked Data in Linguistics Workshop@ LREC2022 (pp. 26-34). URL: https://aclanthology.org/2022.ldl-1.4/ Fradin, Bernard, & Kerleroux, F. (2003). Troubles with lexemes. In G. Booij, J. DeCesaris, A. Ralli & S. Scalise (eds.), Selected papers from the third Mediterranean Morphology Meeting (pp. 177-196). Barcelona: IULA – Universitat Pompeu Fabra. Herce, B. (2025). VeLeSpa: An inflected verbal lexicon of Peninsular Spanish and a quantitative analysis of paradigmatic predictability. Language Resources and Evaluation, 59(2), 1705-1718. DOI: 10.1007/s10579-024-09776-2 Lassila O., Swick, R.R. (1998) Resource Description Framework (RDF) Model and Syntax Specification. URL: https://www.w3.org/TR/1999/REC-rdf-syntax19990222/ Kirov, C., Sylak-Glassman, J., Que, R., & Yarowsky, D. (2016). Very-large scale parsing and normalization of Wiktionary morphological paradigms. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16) (pp. 3121-3126). URL: https://aclanthology.org/L161498/ Lindemann, D. (2025). Ontolex-Lemon in Wikidata and other Wikibase instances. In Proceedings of the 5th Conference on Language, Data and Knowledge: The 5th OntoLex Workshop (pp. 35-45). DOI: 10.5281/zenodo.15471514 Lindemann, D., Ahmadi, S., Khan, A. F., Mambrini, F., Iurescia, F., & Passarotti, M. C. (2023). When OntoLex Meets Wikibase: Remodeling Use Cases. CEUR Workshop Proceedings, 2773. https://ceur-ws.org/Vol-3640/paper14.pdf Mambrini, F., Litta, E., Passarotti, M., & Ruffolo, P. (2021a). Linking the Lewis & short dictionary to the LiLa knowledge base of interoperable linguistic resources for Latin. In Proceedings of the eighth Italian conference on computational linguistics (CLiC-it 2021) (pp. 216-222). URL: https://aclanthology.org/2021.clicit-1.34/ Mambrini, F., Passarotti, M., Litta, E., & Moretti, G. (2021b). Interlinking Valency Frames and Wordnet Synsets in the LiLa Knowledge Base of Linguistic Resources for Latin. In Further with Knowledge Graphs (pp. 16-28). IOS Press. DOI: 10.3233/SSW210032 Mambrini, F., Passarotti, M., Moretti, G., & Pellegrini, M. (2022, June). The Index Thomisticus Treebank as Linked Data in the LiLa Knowledge Base. In Proceedings of the Thirteenth Language Resources and Evaluation Conference (pp. 4022-4029). URL: https://aclanthology.org/2022.lrec-1.428/ McCrae, J. P., Bosque-Gil, J., Gracia, J., Buitelaar, P., & Cimiano, P. (2017). The Ontolex-Lemon model: development and applications. In Proceedings of eLex 2017 conference (pp. 19-21). Nicolai, G., Chodroff, E., Mailhot, F., & Çöltekin, Ç. (2024). Proceedings of the 21st SIGMORPHON workshop on Computational Research in Phonetics, Phonology, and Morphology. Mexico City, Mexico: Association for Computational Linguistics. URL: https://aclanthology.org/2024.sigmorphon-1/ Passarotti M. (2019). The Project of the Index Thomisticus Treebank. In Berti, M. (ed), Digital Classical Philology. Ancient Greek and Latin in the Digital Revolution