Lexical Data

Lexical Data in Wikibase

This page offers an introduction to how lexical data — structured information about words, their forms, and their meanings — can be stored, edited, and linked using Wikibase. It explains the essentials, including how OntoLex, the W3C-recommended vocabulary for the representation of lexical data, is re-used in the Wikibase Ontology. It introduces to Wikidata lexemes, the first and today the largest use case for Wikibase lexemes.

Introduction

Lexical data is structured information about words: how they are written (forms), what they mean (senses), and how they relate to other words (such as compounds, synonyms, or equivalents in other languages), and to ontological items. In Wikibase, each lexical entry (a description of a word) has its own dedicated page and is described in a consistent way so that both humans and machines can understand, edit, and re-use it.

At the moment, the largest Wikibase lexemes use case is Wikidata, which includes a very large and ever-growing collection of lexical entries created by volunteers across many languages (over 1.4 million entries in 1,379 different languages by May 2025).

Core structure of lexemes in Wikibase

In Wikibase, each word is described on a single, editable lexeme page that brings together three essential building blocks:

• the lexical entry or lexeme (with a lemma, a language, and a lexical category or "part of speech"),
• its senses (the different meanings), and
• its forms (inflected forms).

This mirrors the entry–sense–form core structure of the OntoLex-Lemon model. These three OntoLex core classes, and also some OntoLex properties, are re-used in the RDF representation of a Wikibase lexeme. The three core classes, in Wikibase, are stored and displayed together, on one lexeme page. The lexical entry has a numeral universal identifier in the entity namespace of the Wikibase instance, preceded by the letter L (for lexeme); senses and forms are managed as sub-entities, named e.g. L1-S1, and L1-F1. This core structure is pre-set in the data model fundamental to any Wikibase instance and cannot be modified by the user, and the same is true for a small number of properties to be attached to the three core classes, listed in the table below; the three obligatory properties describing a lexeme must have a value, a sense must have a gloss (a short sense-disambiguating text, represented in RDF using skos:definition), and a form must have a representation (a value for ontolex:representation).

Obligatory basic classes and properties in their domain for describing lexemes in Wikibase

ontolex:LexicalEntry

wikibase:lemma

wikibase:lexicalCategory

dct:language

ontolex:sense

ontolex:lexicalForm

ontolex:LexicalSense

skos:definition

ontolex:Form

wikibase:grammaticalFeature

ontolex:representation

In detail, this obligatory structure entails the following restrictions:

• A lexeme a.k.a. lexical entry must point to exactly one item in the same Wikibase as value for lexical category.
• A lexical entry must point to exactly one item in the same Wikibase as value for its language.
• A lexical entry must have at least one lemma. Several lemmata can co-exist for the same entry when covering different spelling variants (e. g. British and American English). In the RDF representation, lexeme lemmata appear attached to the entry using a property named wikibase:lemma. Lemmata are indexed for the MediaWiki text search index, so that a user can search for lexemes, manually in the interface, or via API.
• A lexical sense must have at least one sense gloss, in any language, i. e. not necessarily or not only in the language associated to the entry. Purpose of the gloss is to provide the information necessary to discriminate word senses.
• A form must have at least one written representation. Alike different values for lemma on entry level, form representations in several spelling variants can be attached to the same form entity. A written representation in most cases will be a string as found in text of that language, but written representations also include code transcriptions, such as those describing sign language forms. To describe a form without providing any type of written representation is not foreseen.
• A form can have zero or more items in the same Wikibase as values for grammatical feature, i.e. the morphological description of the form.

The Wikibase lexeme, as illustrated in the Wikibase documentation

All lemmata, glosses, and also form representations are associated to a language code. The available codes are the same that are available throughout the mediaWiki instance, e. g. for labels, descriptions, and monolingualtext strings.

These restrictions guarantee that all lexemes in Wikibase are interoperable at a basic level. Beyond that, any additional relation involving entry, sense or form objects is not predefined and can be modeled according to the use case using self-defined properties in Wikibase statements, as illustrated in the figure. In opposition to that, all other relations added to lexeme, sense or form are added as Wikibase statements (see the section about the JSON/RDF representation) always involve a property defined in the namespace of the Wikibase instance itself, identified by a number preceded by the letter P, and that has a range restricted to one Wikibase datatype.

The Wikibase lexeme data model is also explained in the documentation of the Wikibase MediaWiki extension.

The Wikidata lexeme collection

Wikidata Lexemes logo

Wikidata lexemes form an open and editable collection, where anyone is welcome to collaborate. The data model is documented on the Wikidata pages and is based on three fundamental building blocks described above. Concerning advanced modeling questions beyond this basic structure, contributors get support from each other. A core group of more experienced and active users provides advice to newcomers and occasional contributors also through dedicated outreach events. A manually curated list of lexical entries provides examples in several languages of good and complete modeling practice.

Instead of following prescribed models for every possible linguistic detail (such as those recommended by the OntoLex community for the representation of morphology, etymology, multilingual equivalents, etc.), on Wikidata, a bottom-up grown set of properties is used for describing entries, senses, and forms; see the table below for the ten most frequently used properties for each of the classes, pointing to other entities on Wikidata, or to a data value (in the table, excluding external id properties). In the following, we point out some of them:

• Translation equivalence is expressed using wd:P5972, a property set (in both directions) directly between senses. In May 2025, Wikidata contained 119,847 translation links between senses of different languages.
• For representing etymology, a lexeme is linked directly to the lexical entry describing the etymon using wd:P5191. In May 2025, Wikidata contained 40,540 etymology links from lexeme to lexeme.
• For representing decomposition, wd:P5238, the combines lexeme property, links a compound or multiword entry to its constituents. This property is used in May 2025 207,799 times.
• Pronunciation is described in multiple ways, but always associated to forms; the value of a general pronunciation property wd:P7243 is in general identical to the form representation, but may include stress indicators and other diacritics, e.g. to indicate vowel length. It is recommended that pronunciation audio files and/or IPA transcriptions are attached as qualifiers to the pronunciation claim, but audio files can also be directly attached to form. Today, 74% of 252,652 pronunciation audio files are linked directly to a form. Only audio files hosted on Wikimedia Commons can be used.
• As seen in the table below, there are properties devoted to certain transcription and transliteration systems; each of them is, like all Wikidata properties, described through its entity data, and on an own discussion page. In parallel, a general property transliteration or transcription (wd:P2440) can be used, and its value qualified using the property determination method or standard (wd:P459). The former, defined as subproperties of wd:P2440, are heavily used on Wikidata in general, but in only about 1.2% of its use cases (i.e., not more than 3,808), that property is attached to forms, on which the general property is used around ten times more often: The 70,165 uses of wd:P2440, with a wd:P459 qualifier pointing to the transliteration system, are almost half-divided between items and forms.
• Usage examples (wd:P5831) are recommended to be attached to entry, and not to sense. However, about 7% of the 31,271 usage examples in Wikidata lexemes today remain attached to sense; the distribution across languages shows a diverse picture, but a clear preference for entry as subject of the property. If attached to entry, one or more senses can be declared subject sense (wd:P6072); this is done using that property as qualifier in the example statement. The advantage of this modeling, apart from being able to declare a usage example to be pertinent to more than one sense, lies in the ability to have examples attached to the entry also if in the moment of upload and without or before whatever sense disambiguation procedure it is not clear which sense should be marked as subject sense, e.g. when dealing with examples stemming from corpora, or if (still) no senses are described for the lexeme: As soon as the correct sense can be determined, the wd:P5831 claim is enriched with with a wd:P6072 qualifier, without having to delete and re-write the whole usage example statement (with its references), and attach it to sense. Another strong reason for attaching usage examples to entry is to enable their annotation with a subject form, i.e. the word form of the lexeme appearing in the example. In addition, having examples at both levels complicates their retrieval in queries.

The Wikidata lexeme collection can be explored through dedicated pages, and with external tools. For instance, Ordia (Nielsen 2019) generates dictionary-like exhibitions of lexical data; it also features a tool to look up Wikidata forms matching to tokens in text. Ordia also provides statistics on the Wikidata lexeme collection, such as, for example, lists of the most frequently used properties in the domains of entry (a. k. a. lexeme), sense and form. Another tool, Synia, also provides statistics on lexemes, e.g. counts of values for wd:P6191 language style in different languages.

The 12 most frequently used properties describing Wikidata lexemes, senses and forms (May 2025)

LexicalEntry
wd:P5185	grammatical gender
wd:P5911	paradigm class
wd:P5238	combines lexeme
wd:P31	instance of
wd:P5187	word stem
wd:P1552	has characteristic
wd:P5402	homograph lexeme
wd:P2348	time period
wd:P5191	derived from lexeme
wd:P5186	conjugation class
wd:P5831	usage example
wd:P7486	grammatical aspect
LexicalSense
wd:P5137	item for this sense
wd:P5972	translation
wd:P5973	synonym
wd:P1343	described by source
wd:P18	image
wd:P9488	field of usage
wd:P6191	language style
wd:P8394	gloss quote
wd:P9970	predicate for
wd:P6271	demonym of
wd:P6084	location of sense usage
wd:P10339	semantic gender
Form
wd:P7243	pronunciation
wd:P898	IPA transcription
wd:P443	pronunciation audio
wd:P5279	hyphenation
wd:P5825	ISO 15919 transliteration
wd:P8881	ITRANS (Indic scripts)
wd:P8530	alternative form
wd:P5276	Slavistic Phonet. Alphab. transcr.
wd:P10822	homophone form
wd:P1721	Hanyu Pinyin transliteration
wd:P11950	appears before phonolog. feat.
wd:P11951	appears after phonolog. feat.

The table below lists overall counts for the ten languages with best absolute coverage at the three levels. Asking for relative coverages, without data about the total amount of corpus lemmata, corpus types, and dictionary senses in a language, we might consider the relation between the coverage on Wikidata and the number of speakers of a language. A query like that (taking into account languages with more than 100,000 speakers) results in a different ranking, with Estonian, Breton, Basque and Danish leading lexemes, and Basque, Breton, Dagbani and Norwegian Bokmål as top four languages for senses.

The ten languages with best absolute coverage at the three levels (May 2025)

LexicalEntry	language	LexicalSense	language	Form	language
1,422,331	all Wikidata	585,893	all Wikidata	14,396,207	all Wikidata
239,465	German	48,308	Bokmål	2,802,106	Estonian
102,150	Russian	41,973	English	1,257,083	Basque
96,753	Danish	30,756	Basque	1,246,745	Russian
83,218	Estonian	29,790	Nynorsk	1,199,194	Latin
68,758	English	24,127	Czech	871,813	Czech
67,367	Malayalam	24,055	Swedish	753,118	Malayalam
64,445	Italian	22,000	Italian	692,671	Spanish
63,128	Spanish	21,036	Japanese	641,454	Danish
56,296	Latin	20,484	Persian	571,091	German
48,214	Swedish	18,851	Egyptian	522,107	Italian

Apart from those already mentioned, a range of tools has emerged around Wikidata lexemes, designed to help creating entries and append senses, for linking senses to ontological references, for creating forms collections through templates, searching for usage examples and adding them to entries, for massively recording pronunciation audios, or for graph visualisations of lexical relations.

One goal of the Wikidata lexemes collection is to enable natural language generation for drafting Wikipedia article text from abstract knowledge representations (Morshed 2024), and another possible application is corpus annotation (Lindemann 2024, Lindemann et al. 2025); all these applications of course depend on the degree the lexemes, senses and forms in the collection cover the languages to process.

Wikibase lexemes as linking hub: the case of Wikidata

Wikibase lets lexemes point to external resources, such as dictionary portals and other lexical databases. These links are stored using properties of datatype external id. For a human reader, they act like ‘see also’ links: one can jump from a Wikibase lexeme entry to a dictionary entry elsewhere. For machines, they make it possible to run federated queries, i.e. to search across several linked databases at once.. For example, a query can involve Wikidata and the LiLa Latin Knowledge Base, exploiting the LiLa URI attached to Wikidata Latin lexemes, and calling the LiLa SPARQL endpoint from within the Wikidata Query Service, or vice versa.

Wikidata lexemes, on entry level, by May 2025 count 2.3 Million external id statements. The table below shows usage counts for the most frequent 20 properties. As for sense, all lexical resources aligned to Wikidata on sense level today still lack significant coverage, although we point out the fact that e.g. English Wordnet synset identifiers are aligned not to lexical senses but to ontological Wikidata items, which, as explained, are referenced by a significant number of lexeme senses. An alignment of forms to external resources such as corpus-based form repositories is at large still not present, although it would be interesting, for instance in rich-morphology languages, where the morphologically possible forms outnumber the forms that are actually attested in corpora, so that forms attestation is very valuable information (Lindemann et al. 2025), with a similar value lemma attestation has in languages with a comparably reduced number of different inflected forms, like English.

Most most frequent external id properties on Wikidata lexemes (May 2025)

Count	Property	Property label (en)	Lang. (ISO-639-1)
231,859	wd:P9940	DWDS-Lemma-Identifikator	de
153,322	wd:P8376	Duden-Lexem-Identifikator	de
139,067	wd:P11519	elexiko ID	de
122,350	wd:P11138	Sõnaveeb entry ID	(45 lang.)
84,951	wd:P9947	WDG-Lemma-Identifikator	de
66,512	wd:P13258	Presisov večjezični slovar ID	fr de en sq sh
65,599	wd:P9529	Den Danske Ordbog article ID	da
61,333	wd:P5912	Oqaasileriffik online dictionary ID	da en kl nb
54,755	wd:P12630	Aragonario ID (6th version)	es an
51,566	wd:P11033	LiLa Linking Latin URI	la
49,987	wd:P9385	DWB Lemma ID	de
45,751	wd:P5275	OED Online ID	en
39,052	wd:P9962	Ordbog over det danske sprog ID	da
37,925	wd:P12420	Il Nuovo De Mauro ID	it
37,137	wd:P10042	Bokmålsordboka-ID	nb nn
36,535	wd:P11838	Svenska Akademiens ordlista ID	sv
35,124	wd:P9387	GWB Lemma ID	de
31,014	wd:P12690	New Oxford American Dictionary ID	en
29,803	wd:P11319	Little Academic Dictionary ID	ru
29,316	wd:P12828	DAKA Danish-Greenlandic Dictionary ID	da

Modeling lexical data in Wikibase

Apart from the very basic information about a lexeme, a sense, and a form, which is, as explained above, modeled using pre-set properties, users are free to model fine-grained linguistic descriptions according to their own criteria, that is, following the traditions emerged in Wikidata lexemes, or differently. In general, the following should be taken into account: In order to keep all data pertaining to the same lexical entry "together", so that it would all get stored in the same entity data JSON blob, and displayed on the same lexeme entity page, it is convenient to keep the modeling of the linguistic description as shallow as a list of Wikibase statements. That is what the Wikibase documentation recommends. Unlike in the OntoLex modules, where the linguistic description often involves several reification layers, the model followed in Wikidata, and, for the same reason, in any other Wikibase, will try, wherever possible, to stay with certain reification approaches, which can be combined, but will be limited to the following:

• Using statement qualifiers, i. e. semantic triples describing the main claim of the statement.
• Using references for provenance annotations.
• Using subproperties such as, on Wikidata, wd:P1721 Hanyu Pinyin transliteration, subproperty of wd:P2440 transliteration or transcription.

For reification, i.e. for being able to make claims about claims, OntoLex modules often rely on extra structural layers (e.g. special translation nodes); Wikibase keeps things shallower and simpler. Where OntoLex might introduce a dedicated entity for a translation and then annotate it, Wikibase simply records a translation link between two senses and attaches any extra details as qualifiers on that link. This approach avoids creating new ontological items for describing relations, while still letting editors specify details such as a translation type or direction.

Linking word senses to concepts, and facets of multilinguality

In Wikibase, each word sense can be linked directly to a conceptual item in the knowledge graph. In Wikidata, this is done through the property wd:P5137 item for this sense. For example, Wikidata’s lexeme wd:L3549, describing the English noun foot, has three senses, each of them pointing to a different conceptual item node in the graph, using that property. wd:L3549-S1, the first listed sense, is linked to an item describing a unit of length, while the second sense points to an item describing a furniture part, and the third sense, wd:L3549-S3, glossed as “anatomical structure found in vertebrates”, links to the anatomical entity. Each of the three linked Wikidata items is itself further described, for example, with links to Wikipedia articles in multiple languages (in the RDF representation, the property used here is schema:about), which have a title in that language, and which provide encyclopaedical descriptions of the concept. That means in general terms that Wikibase provides a framework where lexical and ontological (conceptual) descriptions converge, and where text pages (for Wikidata, Wikipedia articles) about concepts also have their habitat. Since, in addition, Wikibase items are annotated with multilingual labels and descriptions, the wd:P5137 reference of a lexeme sense into the ontological part of the Wikidata graph already provides three facets of multilinguality: The concept labels, the textual concept descriptions, and entire text pages attached to an ontological item referenced by lexeme senses — they all may cover multiple languages.

The German lexeme "Pferd" and relations in Wikidata

Since several languages’ word senses can be linked to the same item, wd:P5137 item for this sense provides translation (and, inside the same language, synonymy) information. As of May 2025, Wikidata contains 227,908 item for this sense claims that link senses to items. Calculating the number of translation links through shared references for every item, that sums 5.23 million translation (and intralingual synonymy) links between lexeme senses, if counting the connecting links twice, i.e. as translation link in both directions. This by far outnumbers wd:P5972 sense translation relations, which constitute an alternative without leaving the domain of lexemes, as needed for senses without an ontological reference in Wikidata (most prominently, senses of lexemes with a lexical category other than noun). As described above, multilingual sense glosses (very short sense descriptions attached using the built-in skos:definition) provide another facet of multilingual sense description.

The figure shows an example for a lexeme’s relations in the Wikidata graph: the lexical entry describing the German noun Pferd ("horse"), linked to entities of different types: ontological concepts (items), other lexical entries (lexemes), and lexical senses. Entities of type item are used to represent the language, which allows querying for lexemes according to language features such as the language family or countries where languages are spoken or native to, to represent the lexical category, and, as explained above, for ontological sense references. Label, definition and lemma values are always strings associated to a language code; the figure only shows English and German, while the cited entities on Wikidata cover more languages.

Wikibase instances that describe lexical data

Get a list of Wikibase instances that describe lexemes from Wikibase World using this SPARQL query:

#title: Wikibases that contain lexemes, from Wikibase World
PREFIX wwd: <https://wikibase.world/entity/>
PREFIX wwdt: <https://wikibase.world/prop/direct/>

SELECT ?wikibase ?wikibaseLabel ?wikibaseUrl ?inception (sample(?desc) as ?description) (group_concat(distinct str(?authorName); SEPARATOR=" & ") as ?authors)
WHERE { SERVICE <https://wikibase.world/query/sparql> {
  ?wikibase wwdt:P3 wwd:Q10; wwdt:P12 wwd:Q53. # is a Wikibase / describes lexemes
    optional {?wikibase wwdt:P5 ?inception.} 
    optional {?wikibase schema:description ?desc.}
    optional {?wikibase wwdt:P1 ?wikibaseUrl.} 
    optional {?wikibase wwdt:P17 ?author. ?author rdfs:label ?authorName.}
  ?wikibase rdfs:label ?wikibaseLabel. filter(lang(?wikibaseLabel)="en")}
} group by ?wikibase ?wikibaseLabel ?wikibaseUrl ?inception ?description ?authors order by ?inception

Try it!

Resources

• Wikidata lexemes documentation: https://www.wikidata.org/wiki/Wikidata:Lexicographical_data
• Wikibase lexemes data model documentation: https://www.mediawiki.org/wiki/Extension:WikibaseLexeme/Data_Model
• Nielsen (2020): "Lexemes in Wikidata: 2020 status": https://aclanthology.org/2020.ldl-1.12
• Lindemann (2025): "Ontolex-Lemon in Wikidata and other Wikibase instances": https://doi.org/10.5281/zenodo.15471514
• "Wikibase-Lex" workshop (2023), presentation slides and videos: https://wikibase-lex.sciencesconf.org/

Use the following query for retrieving bibliographical items indexed with domain "Lexical Data".

#title: All bibliographical items indexed with domain label "Lexical Data"
PREFIX dhwb: <https://dhwiki.wikibase.cloud/entity/>
PREFIX dhdp: <https://dhwiki.wikibase.cloud/prop/direct/>
PREFIX dhp: <https://dhwiki.wikibase.cloud/prop/>
PREFIX dhpq: <https://dhwiki.wikibase.cloud/prop/qualifier/>

select ?item ?typeLabel ?date ?itemLabel ?langLabel (group_concat(distinct ?authorname;SEPARATOR=", ") as ?authors) (iri(concat("https://www.zotero.org/groups/5639268/dariah_wg_dhwiki/items/",?zot,"/item-details")) as ?zotero) (iri(concat(str(wd:),?wd)) as ?wikidata)
       
where { 
  ?item dhdp:P5 dhwb:Q2; dhdp:P8 ?type; dhdp:P41 dhwb:Q635; # Q635 Lexical Data
        dhdp:P26 ?date; dhdp:P11 ?lang; dhdp:P6 ?zot; dhp:P29 [dhpq:P10 ?authorname].
  optional {?item dhdp:P1 ?wd.}
  
  SERVICE wikibase:label { bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }
} group by ?item ?typeLabel ?date ?itemLabel ?langLabel ?authors ?zot ?wd
order by desc(year(?date))

Try it!