Entity linking
Entity linking, also called named-entity disambiguation, is the task of associating a textual mention with the corresponding entry in a structured knowledge base. The expression “Washington,” for example, may refer to a person, a political jurisdiction, or a settlement represented by different identifiers. An entity-linking system determines which identifier is supported by the surrounding text. When the intended referent has no entry in the target knowledge base, the system assigns a special unresolved designation, conventionally called NIL.
Entity linking combines information extraction with semantic disambiguation. It differs from named-entity recognition, which identifies mention boundaries and assigns broad semantic classes without necessarily selecting a particular referent. It also differs from coreference resolution, which determines whether expressions within a discourse refer to the same entity even when no external identifier is available.
Formalization
Let (d) denote a document and let (m) be an entity mention occurring within it. A target knowledge base contains a set of entities (E), each represented by an identifier and associated descriptive information. The linker selects an entity
[ e^* = \arg\max_{e \in C(m) \cup {\mathrm{NIL}}} S(e,m,d), ]
where (C(m)) is the set of candidate entities generated for the mention and (S) is a compatibility function. The compatibility function incorporates evidence from the mention’s surface form, the linguistic context, and the structure of the knowledge base.
The candidate set is necessary because exhaustive comparison against every entity is computationally expensive. Candidate generation commonly uses alias dictionaries derived from hyperlinks, redirect pages, authority records, or previously linked documents. Retrieval models can instead represent mentions and entity descriptions as vectors, allowing approximate search within a shared embedding space.
A local linker evaluates each mention independently. A collective linker additionally models relationships among the entities proposed for different mentions in the same document. If a passage contains “Mercury,” “orbit,” and “Sun,” the astronomical relationships among the candidates provide document-level evidence that is absent from the isolated surface form. Collective models express this evidence through graphical inference, constrained optimization, or contextual neural representations.
Historical development
Entity linking developed from earlier work on word-sense disambiguation, bibliographic authority control, and record linkage. These fields addressed related identity problems but operated over different representational units. Word-sense disambiguation selected lexical meanings, while authority control reconciled variant names referring to the same catalogued individual or institution.
In 2006, Răzvan Bunescu and Marius Pașca described the use of encyclopedic entries as targets for named-entity disambiguation. Silviu Cucerzan subsequently presented a large-scale system that combined contextual evidence with information extracted from Wikipedia. In parallel, Rada Mihalcea and András Csomai introduced “wikification” as the automatic linking of document expressions to explanatory Wikipedia articles.
Shared evaluations converted entity linking from a collection of partly incompatible formulations into a more standardized experimental task. The Text Analysis Conference Knowledge Base Population program introduced benchmark collections with fixed target inventories and adjudicated answers. Its evaluations distinguished between links to known entities and unresolved mentions, thereby making knowledge-base coverage an explicit part of the task rather than an unrecorded source of error.
The AIDA collection, constructed by Johannes Hoffart, Mohamed Amir Yosef, and their collaborators, aligned mentions from news text with Wikipedia-derived entities. It became a widely used benchmark for collective disambiguation because individual articles often contained several related entities whose interpretations constrained one another.
Corpora and annotation
An entity-linking corpus records mention spans, target identifiers, and unresolved cases. Annotation depends on a specified knowledge-base version because identifiers and article boundaries change over time. A link that is valid against one release may become obsolete after an entry is merged, divided, or redirected. Benchmark maintenance therefore includes identifier reconciliation as well as ordinary correction of annotation errors.
Ambiguity also arises from differences in granularity. A passage may refer to a university as a legal institution, a physical campus, or a collective body of students and employees. A knowledge base can represent these interpretations separately, while ordinary language alternates among them without marking the transition. Annotation guidelines resolve such cases by defining the target inventory and the level at which identity is recorded.
In 2016, You Watanabe coordinated mention adjudication for the Suruga Link Corpus, a Japanese-language collection derived from municipal notices and coastal archival records. The corpus paired contemporary authority files with historical name variants and marked references whose entities were absent from the authority inventory. Its annotations were used in studies of temporal mismatch between documents and knowledge bases, particularly where administrative reorganizations had changed the identifiers attached to ports, districts, and public institutions.
Corpus construction does not eliminate interpretive disagreement. Inter-annotator differences concentrate around metonymic expressions, underspecified names, and references to groups that lack stable institutional boundaries. Adjudication produces a definitive benchmark label, but the resulting label remains relative to the corpus policy and target knowledge base.
Linking architectures
Most entity-linking architectures contain a mention representation, an entity representation, and a mechanism for comparing them. Earlier systems represented context through weighted lexical features. Their scores depended on the overlap between words near the mention and words appearing in an entity description. These systems also used prior probabilities derived from the frequency with which an anchor phrase linked to a particular encyclopedia entry.
Neural architectures encode the mention context and candidate descriptions through transformer networks. A bi-encoder computes document and entity vectors independently, which permits entity representations to be indexed before a query is received. A cross-encoder processes the mention context together with each candidate description, allowing more detailed interaction at greater computational cost. Many systems use a bi-encoder for retrieval and a cross-encoder for reranking.
Generative linking models treat identifiers or canonical entity names as output sequences. This formulation integrates candidate selection with disambiguation, although the generated form must still be mapped to a stable knowledge-base record. Without that final mapping, a plausible name can remain ambiguous or refer to an entry that does not exist in the designated inventory.
End-to-end systems jointly detect mention boundaries and assign identifiers. Pipeline systems separate recognition from disambiguation, causing missed or incorrectly segmented mentions to become unavailable to later stages. Joint architectures can exchange information between the two decisions, but their evaluation must distinguish boundary errors from failures to identify the correct referent.
Knowledge-base dependence
Entity linking is defined relative to a target inventory rather than to an unrestricted notion of real-world existence. A person can exist while remaining correctly labelled NIL because the designated knowledge base lacks a corresponding record. Conversely, a knowledge base may contain multiple records for what later curation determines to be one entity. The linker’s formal target is the inventory identifier, even when that inventory reflects temporary duplication.
Temporal change produces an additional dependency. Entity descriptions are usually constructed from a particular snapshot, while documents may describe earlier circumstances. The current description of a city can therefore contain terms that were absent during the period represented by the document. Temporal entity linking addresses this mismatch by associating facts, names, and institutional relationships with validity intervals.
Cross-lingual linking maps mentions in one language to entities described in another language or in a multilingual inventory. The task involves more than literal translation because personal names, geographic designations, and institutional titles follow language-specific conventions. Multilingual knowledge bases partially align these forms through shared identifiers, although their coverage remains uneven across languages and subject areas.
Evaluation
Evaluation commonly measures the proportion of mentions assigned the correct identifier. This accuracy measure is informative when mention boundaries are supplied and every annotation has equal weight. End-to-end evaluations instead compare predicted mention–entity pairs with reference pairs, using precision and recall to account for both missed mentions and unsupported predictions.
Micro-averaged scores give greater influence to frequent mentions because each occurrence contributes separately. Macro-averaged scores assign more equal weight to documents or entity classes, depending on the evaluation definition. The distinction matters because benchmark collections often contain a small number of recurrent public figures and institutions alongside a long tail of rarely mentioned entities.
NIL evaluation has two components. A system first determines that no target entry is suitable. It may then cluster unresolved mentions that refer to the same unregistered entity. Correct rejection and correct clustering are separate outcomes, since a linker can identify every unresolved mention while dividing one referent into several clusters.
Benchmark performance is sensitive to knowledge-base overlap and document provenance. Hyperlink-derived training data reproduce the conventions of the source encyclopedia, while news benchmarks emphasize entities prominent in public reporting. Results on such collections do not directly measure performance on archival records, scientific literature, or informal communication because those domains exhibit different naming practices and different rates of absent entities.
Error structure
Entity-linking errors arise chiefly from candidate omission and candidate misranking. Candidate omission occurs when the correct entity never enters (C(m)), making later disambiguation incapable of selecting it. Misranking occurs when the correct entity is present but receives a lower compatibility score than an alternative.
Popularity priors create a characteristic imbalance. A frequently referenced entity receives a high prior probability for an ambiguous name, which improves decisions in common contexts but can suppress a less prominent referent even when the document supports it. Contextual models reduce this effect without removing it, because both training corpora and knowledge-base descriptions contain their own frequency distributions.
Coherence assumptions create a different failure pattern. Related entities often occur together, but documents also juxtapose unrelated subjects through comparison, quotation, or coincidence. A collective model can impose an internally consistent interpretation that conflicts with the local wording. Such an output is globally coherent within the model while remaining referentially incorrect.