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

(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 OntoLex-Lemon 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. 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 pre- and 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., 2022) scraped Wiktionary data to obtain lexicons consisting of form-lemma-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; 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 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’; 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 (ITT; 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 (November 2025) PrinParLat only included verbs, we had to work solely on verb paradigms in the following experiments. 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.

Repository location

https://doi.org/10.5281/zenodo.17553591

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

(3.1) 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 1^st or the 3^rd conjugation, and its principal parts are accordingly assigned to different flexemes,¹⁰ as shown in Table 1.

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., abalien- vs. abalen- for 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 1^st conjugation is linked to the LiLa lemma 110084,¹¹ which is itself stated to be a 1^st conjugation verb, while the PrinParLat flexeme pertaining to the 3^rd is linked to the LiLa lemma 110085,¹² which is itself stated to be a 3^rd 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 fine-grained 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., 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.

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

(3.2) Uploading PrinParLatInflexi forms to the Wikibase

We then use the data of PrinParLatInfLexi to upload data in a Wikibase instance that we created. In the RDF version of PrinParLat linked to LiLa, it was necessary to have flexemes as the anchor point: as we have seen in Section 3.1, flexemes can be assigned a unique tagging of inflectional behaviour, and they can consequently also be unambiguously linked to lemmas of the appropriate inflection type, differently than lexemes, that can display forms inflected according to different conjugations. However, for the purposes of this work (the corpus token linking) we decided to use lexemes as the main lexical entries, to avoid an undesired duplication of forms that would give rise to widespread ambiguity in the linking of tokens to forms, and would also consequently impact the ability to retrieve accurate form counts.

Such duplication can happen because there are cases in which some forms of a lexeme are compatible with more than one flexeme. For instance, the verb ‘clean’ displays both forms inflected according to the 2^nd conjugation (prs.act.inf tergēre, fut.act.ind.3sg tergēbit) and forms inflected according to the 3^rd conjugation (prs.act.inf tergere, fut.act.ind.3sg terget), that are accordingly assigned to different flexemes. However, forms like prf.act.ind.1sg tersi, prf.pass.ptcp.nom.n.sg tersum and fut.act.ptcp.nom.n.sg tersurum cannot be assigned to one conjugation specifically: the morphological processes yielding these forms – namely, -s- suffixation for the formation of the perfect stem ters- and of the so-called “third stem” (Aronoff, 1993) ters- – are found both with verbs of the 2^nd conjugation (e.g., mulgeo ‘milk’, with prf.act.ind.1sg mulsi and prf.act.ptcp.nom.n.sg mulsum) and with verbs of the 3^rd conjugation (e.g., spargo ‘scatter’, with prf.act.ind.1sg sparsi and prf.act.ptcp.nom.n.sg sparsum), and can thus be considered to be compatible with both identified flexemes, as Table 3 illustrates.

In a first experiment, we uploaded one lexical entry for each flexeme into our Wikibase instance and generated full paradigms for those entries. Therefore, all forms generated from the prf.act.ind.1sg, prf.pass.ptcp.nom.n.sg and fut.act.ptcp.nom.n.sg are generated twice: such duplication happens also, e.g., for prf.act.ind.2sg tersisti, prf.pass.ptcp.nom.f.sg tersa, fut.act.ptcp.nom.f.sg tersura, and so on. This would create redundancy and ambiguity in the linking of tokens to form – as any occurrence should be linked to both generated forms, with no possibility for disambiguation. To avoid this issue, in a second experiment we pivoted on lexemes instead: we thus generated different forms for the same cell only when those are segmentally different (like tergēre vs. tergere and tergēbit vs. terget), while there is only one form for the prf.act.ind.1sg tersi and prf.pass.ptcp.nom.n.sg tersum (and all forms generated from such principal parts).

The following figures show how lexical entries are represented in our Wikibase instance¹⁴ and illustrate the difference between the two approaches. Figure 1 shows the entry for the flexeme tergo, with prf.act.ind tersī and prs.act.inf tergere only (see Figure 2). Figure 3 shows the entry for the flexeme tergeo, with another prf.act.ind tersī with its own URI and prs.act.inf tergēre only (see Figure 4). Figure 5 shows the lexical entry for the lexeme tergo, listing both tergere and tergēre (see Figure 6), as they are segmentally different, but with only one form when the shape is the same as in tersī. In Figure 5, the lexeme is linked to its flexemes and to the corresponding lemma in LiLa through dedicated properties. If more than one LiLa lemma is compatible with a lexeme, the links to those are qualified by the flexeme they correspond to, to ease disambiguation. A list of forms with their grammatical features is then provided, and each individual form is also linked to the flexemes it is compatible with (Figure 6). Furthermore, the form and its segmentation as provided in PrinParLatInfLexi are given. Lastly, there is a link to a Wikibase item that represents the coding of the paradigm cell that the forms fill. Such a reification is recommended in the Paralex ontology, where paradigm cells are assigned to individuals of a dedicated class paralex:Cell. In PrinParLatInfLexi (like in PrinParLat), the paradigm cell is expressed through a custom Paralex-compliant coding that uses the labels common in traditional descriptions of Latin, and their abbreviations defined in the Leipzig Glossing Rules¹⁵ – e.g., “prf” for perfect tense-aspect.

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

More generally, as we have seen, the Paralex ontology provides classes and properties that are envisaged to be used for the release of Paralex data as RDF LOD. Since the resource we start from, PrinParLatInfLexi, is constructed to be a Paralex lexicon, to guarantee interoperability with other Paralex lexicons released as RDF LOD, the correspondence between classes and properties that we introduce in our Wikibase instance and those of the Paralex ontology, on one side, and to Wikidata, on the other, is established through dedicated properties, as shown in Figure 7 for the Paralex orth form property, linked to paralex:orth_form and wikidata:P7243. This enables us to export a version of the datasets in Wikibase using the aligned Wikidata or Paralex OWL property, respectively, for a seamless integration into these environments.

Figure 7

(3.3) Uploading ITTB tokens to the Wikibase

The other pieces of information that we need to fulfil the aim of this work are corpus tokens. To that end, we use the Index Thomisticus Treebank (ITTB). It is a corpus of the work of Thomas Aquinas featuring the fine-grained morphological tagging that we need for this work. Since PrinParLatInfLexi only includes verbs, we only extract verbal tokens of the ITTB. We then generate items in our Wikibase instance for each verbal token, as shown in Figure 8, preserving the links to the IDs of the token in the LiLa Knowledge Base and of the lemma to which the token is linked through dedicated properties. We also record the fine-grained morphological annotation provided in there using the tagset of the Universal Dependencies (UD) project (Petrov et al., 2012) and expressed as RDF LOD using the Web Annotation Ontology (Sanderson et al., 2013). In our Wikibase instance, we introduce a property that points to the different feature-value pairs in the UD tagset (e.g., “Tense#Past”).

Figure 8

(3.4) Mapping ITTB tokens to PrinParLatInfLexi forms

The final step is linking the tokens of the ITTB to the forms of PrinParLatInfLexi.

For an ITTB token to be linked to a PrinParLatInfLexi form, the following requirements must be met.

1. 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.

2. 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

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