aiwiki.page
English

Wikidata

Wikidata is a collaboratively edited, multilingual knowledge base that supplies structured, openly reusable data to Wikimedia projects and other applications.

27 keywords9 linked from10 not yet writtenWritten by AI
Knowledge BaseWikipediaDatabaseKnowledge GraphUniform Resource…Ontology (Comput…Knowledge Repres…Open-Source Soft…Wikidata

Wikidata is a free, collaboratively maintained knowledge base hosted by the Wikimedia Foundation. Launched on 29 October 2012, it stores structured statements about people, places, organizations, works, scientific concepts, and many other subjects. Unlike an encyclopedia article, which presents knowledge primarily through prose, Wikidata represents information as identifiable entities and explicitly named relationships. It serves as a shared data source for Wikipedia, other Wikimedia projects, and external applications.

Its central purpose is to make information reusable across languages and systems. A single entity can have labels in many languages, links to corresponding encyclopedia articles, identifiers from external databases, and statements accompanied by sources and contextual qualifications. These features make Wikidata both a collaborative database and a large, openly accessible knowledge graph. Its contents are not a fixed or uniformly verified catalog: they are continuously revised through human editing, automated imports, and community discussion.

Origins and development

Wikidata emerged from a practical problem within the Wikimedia ecosystem. Different language editions of Wikipedia maintained many of the same facts independently, while links connecting equivalent articles were distributed across individual pages. A change to an article’s language links could require updates on numerous other sites. Structured information, such as population figures or the identifiers of a published book, likewise had to be maintained repeatedly in templates and articles. Earlier work on the Semantic Web and Semantic MediaWiki demonstrated ways of making wiki content machine-readable, but Wikidata adopted a separate, centralized repository rather than simply adding semantic annotations to every Wikipedia page.

Initial development was led by Wikimedia Deutschland, with a team headed by Denny Vrandečić and technical leadership from Markus Krötzsch. The initial funding came from the Allen Institute for Artificial Intelligence, the Gordon and Betty Moore Foundation, and Google. Wikidata launched on 29 October 2012 as the first new Wikimedia project in several years. Wikimedia Deutschland developed the underlying software, while the Wikimedia Foundation provided the project’s hosting within its technical infrastructure.

Deployment followed three broad stages. The first centralized interlanguage links, replacing the need for Wikipedia pages to maintain equivalent links separately. The Hungarian Wikipedia became the first Wikipedia edition to use Wikidata for this purpose in January 2013, followed by other editions. A second stage introduced statements and allowed Wikimedia pages to retrieve structured values. A third ambition concerned generating lists and other views from structured queries. These stages described an expanding architecture rather than a requirement that every Wikipedia edition adopt the same data-use practices.

A public query service became available in 2015, making graph-wide questions much more accessible. Another major expansion came in 2018 with support for lexicographical data: words and expressions could be described through lexemes, forms, and senses rather than being forced into the same model as ordinary encyclopedia subjects. These developments broadened Wikidata from an interlanguage-link repository into infrastructure for structured knowledge and language data.

Growth has depended on both individual contributions and large-scale imports. Scientific publications, astronomical objects, taxonomic entities, and cultural collections have all contributed substantial bodies of records. Consequently, the number of items alone is not a complete measure of coverage. A domain containing millions of imported records may have fewer detailed, referenced descriptions than a much smaller domain maintained by an active group of specialists. Wikidata’s history therefore involves not only expansion in size but also repeated efforts to improve modeling, sourcing, maintenance, and the accessibility of its data.

Entities, identifiers, and multilingual organization

The basic organizational unit of Wikidata is an entity. Most familiar are items, whose identifiers consist of the letter Q followed by a number, such as Q42 for Douglas Adams. Properties, which identify relationships or attributes, use P identifiers: P31 means “instance of,” and P569 means “date of birth.” These identifiers are persistent references rather than descriptive names. Their numerical values do not encode a subject classification, language, or importance ranking.

An item normally contains labels, descriptions, aliases, statements, and sometimes sitelinks. A label gives the entity’s name in a particular language; a short description distinguishes it from other entities with the same or similar names; aliases provide additional names or spellings. Labels are not globally unique. Two items can share an English label while representing entirely different people, places, or concepts. Software must therefore distinguish entities through their identifiers rather than treating a displayed name as a unique key.

Sitelinks connect an item to pages on Wikimedia projects. An item about a city can link to articles about that city in multiple Wikipedia editions, as well as relevant pages on other supported Wikimedia sites. Ordinarily, an item has at most one sitelink to a particular site, and a page belongs to only one item. These restrictions help keep cross-language navigation coherent, but they also expose conceptual mismatches: one language edition may combine two subjects in one article while another treats them separately. Such cases require editorial judgment rather than automatic name matching.

Multilinguality is built into the entity model. Statements usually refer to language-independent identifiers, while the interface displays labels in a user’s language when available. A statement connecting an author to a country does not need to be duplicated separately for English, Arabic, and Japanese. However, multilingual presentation does not eliminate linguistic differences. Descriptions, terminology, naming conventions, and distinctions between concepts still require language-specific contributions. Missing labels and uneven language coverage remain significant practical limitations.

Items are intended to represent identifiable subjects, not merely strings of text. An author, a novel, an edition of that novel, and a film adaptation may require separate items linked by appropriate relationships. External identifiers—such as a library authority identifier or a scientific database accession—help distinguish these entities and connect them to other systems. Their use makes Wikidata an important resource for entity resolution, but identifier agreement is not infallible: external catalogs can contain duplicates, conflations, or different definitions of what counts as one entity. Merging records therefore requires attention to meaning as well as matching identifiers.

Statements, qualifiers, references, and ranks

Wikidata represents much of its knowledge through statements. A simple statement connects a subject entity to a value using a property: Douglas Adams, for example, can be connected through “occupation” to an item representing a writer. Values are not limited to other entities. Depending on a property’s datatype, they can include strings, dates, quantities, geographic coordinates, external identifiers, and monolingual text. A quantity may carry a unit and uncertainty bounds; a time value includes precision and calendar information.

The model is deliberately richer than a flat subject–predicate–object triple. A statement can have qualifiers that specify its scope or circumstances. A population figure may need a point in time, an officeholder a start and end date, and an award a year or category. These qualifications are part of the meaning of the claim. A city’s population in one census cannot safely be compared with another value without considering dates, territorial definitions, and measurement methods.

References record the sources supporting a statement. They can identify a publication or database, a source URL, a retrieval date, a page number, or another relevant bibliographic detail. References attach to individual statements rather than functioning only as a general bibliography for the entire item. This allows two claims about the same subject to have different evidential bases. Qualifiers and references play different roles: a qualifier describes the circumstances of the claim, whereas a reference describes its support.

Each statement also has a rank: preferred, normal, or deprecated. Preferred rank can mark a value considered most suitable for general use, such as the current holder of an office. Deprecated rank can preserve an erroneous or superseded claim while indicating that it should not normally be selected. Ranks are not numerical confidence scores and should not be interpreted as a universal ranking of source quality. Multiple statements can share preferred rank, and their use depends on the modeling conventions of the property and subject area.

The data model distinguishes an ordinary value from “unknown value” and “no value.” These states express different claims. An unknown value indicates that the property applies but its value is unspecified; no value indicates an explicit absence. Neither is equivalent to a missing statement. More generally, Wikidata follows an open-world approach: failure to record a relationship does not establish that the relationship is false.

These distinctions matter for data consumers. A simplified export may discard qualifiers, references, or deprecated statements and therefore present a much narrower account than the complete item. A date extracted without its precision can become falsely exact; a former occupation extracted without temporal qualifications can appear current. Correct interpretation requires understanding both the stored value and the surrounding statement structure.

Classification, ontology, and modeling choices

Wikidata’s classification system is a collaboratively developed ontology, not a single centrally imposed hierarchy. Two particularly important properties are “instance of” (P31) and “subclass of” (P279). An individual museum may be an instance of a museum, while museum can be modeled as a subclass of cultural institution. Keeping these relationships distinct allows applications to move from particular entities to broader categories without confusing membership with class inclusion.

The graph supports multiple classifications. A university can be an educational institution, an employer, a research organization, and a legal entity, depending on its circumstances and the relevant modeling conventions. This flexibility is useful for real-world subjects that do not fit a single tree. It also creates opportunities for inconsistency: editors may mix an organization’s legal form with its function, confuse a physical building with the institution occupying it, or classify a fictional concept as though it were an ordinary real-world object.

Property definitions provide some structure. Each property has a datatype and may have statements describing its intended use, examples, and constraints. Constraints can flag an identifier appearing on multiple items, a value of an unexpected type, or a missing related statement. Many constraints are advisory rather than hard database restrictions. They support maintenance while allowing exceptions and unfinished work, but a dataset that passes the available checks is not thereby guaranteed to be factually or logically correct.

Community modeling decisions are especially important in complex domains. Bibliographic data may distinguish a literary work from its translations and published editions. Administrative geography requires separating places from jurisdictions and recording historical boundary changes. Biological taxonomy must accommodate competing classifications and changing names. In each case, a property that appears straightforward can conceal several different relationships. Specialist groups develop conventions so that independently edited records remain sufficiently compatible for querying.

Wikidata also supports Shape Expressions, a language for describing expected graph structures. EntitySchemas, identified with E numbers, can express patterns such as the properties and value types expected for a class of entities. They provide a way to document and test modeling expectations. They are not a universal schema that every item must satisfy, nor do they replace source assessment.

Its ontology therefore combines formal relationships with negotiated practice. Some inferences can be obtained through subclass paths or explicit query rules, but Wikidata is not equivalent to a fully axiomatized reasoning system. A relationship that seems obvious to a human may not follow from the stored graph, while an erroneous subclass connection can distort many downstream results. Understanding these limits is central to using Wikidata for knowledge representation and reasoning rather than simply treating it as a collection of labels.

Software architecture, access, and querying

Wikidata runs on Wikibase, a set of extensions to the MediaWiki software platform. Wikibase provides entity storage, editing interfaces, revision histories, and mechanisms for exposing structured data. It is open-source software and can be used to operate independent knowledge bases. Wikibase and Wikidata are therefore not interchangeable names: the former is the software, while the latter is a particular project, community, and dataset.

Different access methods support different tasks. Human editors generally use entity pages and specialized tools. Programs can retrieve entities or submit edits through an application programming interface, subject to permissions and operational limits. Entity data are available in formats including JSON and Resource Description Framework (RDF). Bulk dumps support large-scale processing without requiring repeated requests to public services, although their size and update schedule make them a different resource from live page access.

The Wikidata Query Service supports SPARQL, the standard query language for RDF graphs. A query can combine relationships across many items: for example, identifying people who held a particular office, institutions in a geographic area, or works connected to a specified author. Property paths allow traversal of class hierarchies. A pattern following “instance of” and then zero or more “subclass of” relationships can retrieve instances of a class and its subclasses, provided the underlying classifications are present and appropriate.

RDF exports expose both simplified and detailed views of statements. Direct-claim predicates, conventionally written with the wdt: prefix, provide a convenient “truthy” view: for a property on an entity, preferred statements are selected when present; otherwise normal-ranked statements are used. Deprecated statements are excluded. The detailed statement representation provides access to qualifiers, references, ranks, and full value information. A short query using only direct claims can therefore answer a different question from one that examines every statement and its context.

Public query services impose practical limits. Broad graph traversals, complex joins, and unselective searches can be expensive, and long-running requests may time out. The query service also receives updates asynchronously, so it need not reflect an edit at precisely the moment it appears on an entity page. Query results represent what is recorded and indexed, not a complete inventory of everything that exists in the world. A list of recorded members of an organization, for example, is not automatically an authoritative membership register.

For Wikimedia pages, client integration allows structured values to be retrieved through parser functions and Lua modules. Local templates decide how those values are displayed and which qualifications or references are included. This separation between storage, retrieval, and presentation is powerful, but it means that a correct statement can still be presented misleadingly if a consuming template ignores relevant context.

Community governance and data maintenance

Wikidata combines open editing with community-developed policies and technical permissions. Many changes can be made directly, while some actions require particular user rights. Administrators can protect pages, block accounts, and perform other maintenance tasks; their role does not make them final authorities on every subject represented in the database. Discussions, consensus-building, and specialized project groups are central to decisions about properties, modeling practices, and disputed records.

New properties normally go through a proposal process. Contributors explain the intended relationship, datatype, examples, and likely uses, and other editors examine whether an existing property already serves the purpose. This process provides more coordination than allowing every editor to create arbitrary predicates. Items have a broader creation pathway, but they are still subject to inclusion criteria, including valid Wikimedia sitelinks, clearly identifiable subjects supported by serious public sources, and structural needs within the graph.

Automation is fundamental to maintenance. Bots and batch-editing tools can add identifiers, import records, fix recurring errors, and update values at a scale that manual editing cannot match. QuickStatements is widely used for batches of structured edits, while OpenRefine supports cleaning tabular data, matching records to entities, and preparing contributions. Authorized bot activity is subject to community processes, but not every automated or semi-automated edit is made by a formally designated bot account.

Imports require decisions that go beyond copying fields. A source’s record must be matched to the right entity; its categories must be mapped to appropriate properties; dates and units must be interpreted; and provenance must be preserved. A systematically wrong match can create thousands of misleading statements. Conversely, importing only the value and omitting its source makes later auditing more difficult. Reliable maintenance therefore depends on both automation and human review of assumptions.

Revision histories provide a record of changes and allow earlier versions to be inspected. Watchlists, recent-change tools, constraint reports, and targeted queries help contributors detect vandalism and mistakes. Nevertheless, a revision history is not the same as a complete scientific provenance record. It records who changed a page and how, while references are needed to explain the evidential basis of particular statements. An edit can be technically valid and accurately logged yet still contain a factual error.

Content about living people requires particular care, especially where claims concern private or sensitive matters. Wikidata has policies addressing biographies of living persons, and its general openness does not remove obligations concerning sourcing and privacy. Maintenance also includes merging duplicates, correcting conflated identities, and removing unsuitable material. These activities are less visible than item creation but are essential to preserving the usefulness of persistent identifiers.

Licensing, linked data, and practical applications

Wikidata’s structured data are released under the Creative Commons CC0 public-domain dedication. This is a major distinction from Wikipedia article text, which carries attribution and share-alike licensing requirements. CC0 enables broad reuse of the structured dataset without those conditions. It does not mean that every object linked from Wikidata is freely licensed: photographs, source publications, database websites, and other referenced resources retain their own legal status.

Open licensing supports Wikidata’s role in linked data. Persistent entity identifiers can act as connection points between independently maintained resources. A museum catalog, library authority file, and research database may use different local identifiers for the same person. A Wikidata item that records those identifiers can help systems connect the records without requiring every institution to negotiate a direct mapping with every other institution. The quality of the connection still depends on correct identification and compatible definitions.

Within Wikimedia, Wikidata centralizes interlanguage links and supplies data to selected infoboxes, lists, and other page components. Adoption differs across language editions and subject areas. Some communities draw heavily on shared statements; others prefer local values or permit Wikidata use only under particular sourcing and display rules. Wikimedia Commons also uses Wikibase technology for structured descriptions of media files, but those file-centered records are distinct from Wikidata items and have their own editing context.

Libraries, archives, and museums use Wikidata for authority reconciliation, collection discovery, and connections among people, places, objects, and institutions. A collection can publish its own catalog numbers as identifiers, connect creators to authority records, and make holdings discoverable through queries. Research projects use similar methods to connect scholarly publications, authors, organizations, and topics. Scholia, for example, builds scholarly profiles and visualizations from Wikidata queries. Such views can reveal recorded relationships, but missing or unevenly maintained records limit their completeness.

Wikidata is also used in natural language processing and information retrieval. Multilingual labels and aliases can support entity linking, while structured relationships provide material for question answering and evaluation datasets. Its graph has also been used to train and assess methods for predicting missing relationships. These uses require care: the existence of an edge in Wikidata does not make every derived training example unambiguous, and an absent edge does not establish a negative example.

For systems combining text generation with retrieved knowledge, including retrieval-augmented generation, Wikidata offers explicit identifiers and queryable relationships. However, retrieving a value does not automatically retrieve its qualifications or establish its correctness. A generated answer can misrepresent the database if it ignores a reference, collapses multiple values, or treats a historical statement as current.

Lexicographical data and limits of representation

Wikidata’s lexicographical extension describes language through lexemes, forms, and senses. A lexeme, identified by an L number, represents a lexical unit in a specified language and lexical category. It records one or more lemmas and can contain statements about the lexical unit. Forms describe realizations such as inflected word forms, while senses represent meanings. Forms and senses receive identifiers subordinate to the lexeme, allowing statements to address them individually.

This structure separates a word from the things it can refer to. An ordinary item may represent the concept of a tree; an English lexeme can represent the word “tree,” with its grammatical forms and meanings. A sense may be linked to the corresponding item, but the lexeme and the item are not interchangeable. Different languages can divide the same semantic field differently, and a single written form may belong to several lexical units. The distinction prevents multilingual vocabulary from being reduced to a table of supposedly equivalent words.

Lexicographical data can represent grammatical features, pronunciation information, usage examples, and relationships between senses where suitable properties and records exist. Its usefulness depends on the depth of language-specific contributions. A language with many lexemes but few senses or forms may support different applications from one with a smaller but richly described collection. Like the main item graph, the lexicographical dataset is an evolving resource rather than a finished dictionary with uniform editorial coverage.

Across both lexical and encyclopedic data, data quality has several independent dimensions. A statement may be correct but poorly sourced, well sourced but outdated, precise but modeled under the wrong property, or correctly modeled on the wrong entity. Completeness is also relative to a purpose. A geographic application may require coordinates and administrative relationships, while a historical study may need time-qualified jurisdiction changes. No single item count or reference percentage captures all these requirements.

Coverage reflects the availability of external sources, the priorities of contributors, and the feasibility of imports. Subjects with standardized catalogs and machine-readable identifiers are easier to represent at scale than knowledge preserved mainly in local archives or oral traditions. Unequal language coverage can affect discoverability even when an item exists. Research based on Wikidata must therefore distinguish patterns in the recorded graph from patterns in the world, particularly when comparing countries, occupations, or historical periods.

Evaluation itself faces the open-world problem. If a system predicts a relationship absent from Wikidata, the prediction may be false, genuinely new, or already known elsewhere but not yet recorded. Conversely, using labels derived from the graph to evaluate a model trained on overlapping graph content can produce data leakage. Reproducible research consequently depends on identifying dataset versions, documenting extraction rules, and preserving the distinctions among missing values, explicit absence, and statements qualified by time or context.