One name can point to many things.Disambiguation finds the intended entity.
On this page, entity disambiguation means resolving an ambiguous mention among multiple plausible referents by comparing candidate identities against contextual evidence. The task is a focused part of the broader mention-to-identity resolution layer covered in ENTITY / 03.
The objective is not to force every mention into a known identity. A robust system should preserve ambiguity when evidence is insufficient and should also allow a no-link or NIL-style outcome when the intended referent is not represented in the candidate catalog.
Disambiguation resolvesidentity competition.
The same surface form can refer to several valid entities. Resolution requires evidence that separates the intended candidate from alternatives.
Which referent does the mention denote?
ENTITY / 03 owns the broader mention-to-identity layer: mention handling, type, aliases, identifiers, attributes, relationships and identity evidence can all contribute to resolving a referent.
BROAD IDENTITY-RESOLUTION LAYERWhich candidate survives when identities compete?
This node owns the ambiguity subproblem: candidate competition, elimination, contradictions, abstention and no-link outcomes when more than one referent is plausible.
AMBIGUITY / CANDIDATE-COMPETITION LAYERResolve the mentionagainst competing candidates.
Select an ambiguous surface form. The lab is illustrative: it shows how type, local context, attributes and relationships can support or eliminate candidate referents. It is not a Google Search resolver or a public confidence model.
Resolution begins witha candidate set.
Many entity-linking and disambiguation pipelines generate a plausible candidate set and then rank or eliminate candidates using contextual evidence. End-to-end systems may not expose these stages explicitly, so the pipeline below is a conceptual decomposition rather than a universal architecture.
Strong resolution usesmultiple compatible signals.
No single cue is universally decisive. Practical systems often combine several compatible signals, but those signals are not necessarily statistically independent.
Entity Type
Person, organization, location, product, work, event or another semantic class.
FILTER IMPOSSIBLE TYPESLocal Context
Nearby words, verbs, modifiers and topics establish what interpretation fits the passage.
READ THE CONTEXT WINDOWAttributes
Occupation, location, date, product category, nationality and other properties separate candidates.
MATCH THE FINGERPRINTRelationships
Connections to people, products, places and organizations can strongly constrain the identity.
VERIFY GRAPH EDGESAliases
Legal names, abbreviations, alternate names and common references may point to the same entity.
UNIFY SURFACE FORMSTemporal Context
Dates and historical periods can eliminate candidates that did not exist or did not hold a role at that time.
CHECK TIME VALIDITYGeographic Context
Country, city, market and local references can disambiguate similarly named places and organizations.
CONSTRAIN LOCATIONEvidence / Metadata
Official URLs, structured identity fields, identifiers and corroborating metadata can strengthen a selected candidate when they refer to the same entity.
VALIDATE THE MATCHType can removeentire candidate classes.
Google Cloud Natural Language documentation explicitly states that entity type can help distinguish or disambiguate similarly named entities. That statement describes the Cloud Natural Language API, not a disclosed Google Search ranking or entity-resolution pipeline.
Lawrence
Could reference a person, place or work depending on context.
AMBIGUOUS SURFACE FORMPERSON vs WORK_OF_ART
A sufficiently supported type classification can eliminate interpretations that are incompatible with the surrounding sentence.
CANDIDATE SPACE SHRINKSContext selects the identity
Type does not always finish the job, but it can dramatically reduce the candidate set.
TYPE → CONTEXT → ENTITYAliases and coreference can feedthe same identity decision.
Alias normalization and coreference can supply additional mention-to-entity evidence, but they are not identical to entity disambiguation. Alias handling maps alternate names to the same identity; coreference links mentions inside discourse; disambiguation then evaluates competing referents when ambiguity remains.
International Business Machines Corporation
Formal registered identity.
CANONICAL FORMIBM
Widely used abbreviation.
ALIASthe company
Coreference that may point back to IBM inside a passage.
MENTION LINKOne organization referent
The legal name and abbreviation can map directly to the same entity; “the company” requires discourse-level coreference evidence before it is attached to that entity.
KEEP TASK BOUNDARIES VISIBLEAttributes createa candidate fingerprint.
When names collide, a small set of discriminating properties can separate otherwise similar identities.
Professional footballer
Occupation: athlete · Sport: football · National team: United States · Birth year: 1989.
FINGERPRINT ADifferent same-name person
A hypothetical namesake could have a different occupation, employer, location and identifier set. Those differences prevent facts from being merged into the athlete profile.
ILLUSTRATIVE CONTRASTGraph relationships can becomeidentity evidence.
A candidate that fits the name but conflicts with known relationships is a weak resolution.
“Python code uses type hints…”
The sentence implies a programming-language ecosystem rather than an animal.
Python → type-hint feature context
The clue is semantically compatible with the programming-language referent and incompatible with the animal interpretation.
Python programming language
The candidate aligns with type, attribute and relationship evidence.
Resolution can depend onwhen the statement applies.
People change roles, organizations rename, products gain new versions and corporate structures change. Temporal context can help distinguish current from historical names, roles and relationships. A merger or reorganization may preserve, replace or terminate an entity identity depending on the underlying legal and operational facts.
Resolution is more defensible whenthe identity can be corroborated.
Official identity surfaces, applicable identifiers and well-matched references can provide additional support after contextual resolution. Structured data is one explicit representation layer, not identity proof by itself.
Official website or profile
A first-party surface can establish how an organization, person or product represents its own identity, subject to the claim being checked.
FIRST-PARTY IDENTITY EVIDENCEname / alternateName / url / sameAs
For applicable entity types, structured fields can make identity information more explicit. Use sameAs only for another page that represents the same entity, not for merely related pages.
Independent matching references
Compatible evidence from additional sources can reduce remaining ambiguity.
CONFIRM THE CANDIDATEA robust resolver canabstain from linking.
False certainty is worse than explicit uncertainty. Keep ambiguity among known candidates separate from a no-link/NIL case where the intended referent is not represented in the available entity catalog. The labels below are a conceptual editorial model, not universal API states.
One candidate dominates
Type, context, attributes and relations consistently support one entity.
COMMIT IDENTITYOne candidate leads
The evidence is directional but not yet strong enough for high confidence.
PRESERVE UNCERTAINTYEvidence is insufficient
Multiple candidates remain plausible after available filtering.
DO NOT FORCE A MATCHNo candidate represents the referent
The mention may denote a real entity or concept that is absent from the available catalog, index or knowledge base.
DO NOT LINK TO A WRONG KNOWN ENTITYNot all resolution difficulty isthe same problem.
The operational categories below are not a formal linguistic taxonomy. They separate common causes of mention-to-identity difficulty so namesake collision, acronym collision, alias normalization, coreference and temporal context are not handled as if they were one task.
One string can denote conceptually different referents, such as Mercury the planet and mercury the chemical element.
USE TYPE + DOMAIN CONTEXTDifferent people, companies or places can share an identical or near-identical proper name.
USE ATTRIBUTES + RELATIONSThe same abbreviation can expand into unrelated organizations, systems or concepts.
USE EXPANSION + FIELD CONTEXTDifferent strings can refer to one entity and should not be fragmented into separate identities.
MERGE ONLY WHEN IDENTITY MATCHESPronouns and descriptions such as “the company” can refer back to an entity introduced earlier in the passage.
TRACK MENTION CHAINA title, owner, role or organizational name can change while the entity continues to exist.
USE DATE + STATEGood resolution removes candidatesfor explicit reasons.
This conceptual matrix uses the mention “Jordan” with Chicago Bulls and player-role context. It shows how candidate elimination can be more rigorous than simply assigning the most familiar entity.
Resolution needs evidence forand against the candidate.
A robust resolver should not collect only confirming signals. Contradictions, missing qualifiers and neutral evidence help prevent premature commitment.
The candidate type, attributes or relationships fit the passage and independently support the same identity.
INCREASE SUPPORTA material fact conflicts with the candidate’s type, date, location, role or known relationship.
REOPEN CANDIDATE SETThe clue is true for several candidates and should not be treated as if it separates them.
DO NOT OVERWEIGHTA necessary distinction cannot be tested from the available passage or sources.
PRESERVE UNCERTAINTYThe entity can persist even whenthe surface form changes.
Coreference resolution is a related discourse task rather than a synonym for entity disambiguation. Pronouns, role descriptions and abbreviated references can point back to an already resolved entity and thereby provide evidence to a broader identity-resolution system.
Sometimes the best resolution isnot to resolve.
These states are an internal editorial model for handling uncertainty. They are not Google confidence thresholds, public machine scores or universal decision boundaries. A real system may also distinguish ambiguity among known candidates from a NIL/no-link state.
Several candidates fit the name and there is not enough context to narrow the field.
RETURN UNRESOLVEDType or domain context removes some candidates, but more than one plausible identity remains.
SEEK MORE EVIDENCEOne candidate leads, but a material alternative cannot yet be excluded.
QUALIFY THE CLAIMIndependent signals converge and competing candidates conflict with the evidence.
LOWER AMBIGUITYThe same identity is also supported by applicable canonical references or identifiers.
RESOLUTION SUPPORTEDEntity resolution fails whenweak signals are treated as proof.
Most disambiguation errors come from premature commitment, missing context or incorrectly merged identities.
Assuming identical strings always refer to the same entity.
Ignoring that one surface form can represent different semantic classes.
Treating alternate names of one entity as separate identities.
Using a current relation to interpret an older historical statement.
Selecting a candidate whose known graph relations contradict the passage.
Resolving the mention despite insufficient evidence.
Selecting the most famous candidate rather than the candidate best supported by the local evidence.
A pronoun or descriptive mention becomes attached to the wrong antecedent as the passage develops.
Supporting clues are collected while contradictory evidence is ignored.
The intended referent never enters the candidate set, so later ranking has no chance to select it.
The intended referent is absent from the known entity inventory, but the system forces a link to an unrelated known candidate.
NER, alias normalization, coreference and disambiguation outputs are treated as if they were interchangeable evidence states.
Disambiguation can constrainwhich evidence is relevant.
In retrieval and AI-search environments, resolving the wrong entity can send later stages toward irrelevant documents, passages or relationships. Actual retrieval, reranking, synthesis and citation behavior remain system-dependent.
Audit whether identity isactually resolvable.
This is a conceptual editorial and knowledge-modeling checklist, not a Google score.
Are plausible alternate identities acknowledged before resolution?
CHECKDoes the candidate type fit the mention and surrounding context?
CHECKDo nearby words and topics support the chosen identity?
CHECKDo distinguishing properties match the selected candidate?
CHECKAre graph relations compatible with the passage?
CHECKAre alternate names consolidated without merging different entities?
CHECKAre dynamic roles and relationships valid for the relevant date?
CHECKCan the selected entity be connected to authoritative identity evidence?
CHECKHave conflicting candidate signals been investigated rather than ignored?
CHECKCan the system preserve uncertainty when no candidate is sufficiently supported?
CHECKAre material contradictions used to eliminate or reopen candidates?
CHECKDo pronouns and descriptive mentions remain attached to the correct antecedent?
CHECKWas the intended referent actually included in the candidate set before ranking?
CHECKCan the system represent a referent that is absent from the known entity catalog without forcing a wrong link?
CHECKAre mention detection, typing, alias normalization, coreference and candidate disambiguation kept conceptually distinct?
CHECKIs repeated derivative evidence distinguished from genuinely independent corroboration?
CHECKResearch entity resolution.Explore identity, evidence and ambiguity.
Explore the complete Entity SEO research layer from foundation and identification through attributes, relationships, disambiguation, salience, contextual relevance, coverage, representation depth, authority and AI Search. Each node owns a distinct analytical function.
What Is Entity SEO?
Define the entity-oriented layer through identity, attributes, relationships, context and disambiguation.
OPEN NODE → ENTITY / 02 · DISTINCTIONEntities vs Keywords
Separate lexical strings from identifiable referents while preserving the role of keyword and query analysis.
OPEN NODE → ENTITY / 03 · IDENTITYEntity Identification
Own the broader mention-to-identity layer through type, aliases, identifiers, attributes, relationships, context and identity evidence.
OPEN NODE → ENTITY / 04 · PROPERTYEntity Attributes
Model literal or descriptive properties, values, qualifiers, temporal state and evidence while keeping entity-valued relations separate.
OPEN NODE → ENTITY / 05 · GRAPHEntity Relationships
Connect resolved entity nodes through explicit predicates, direction, scope, time, vocabulary control and evidence.
OPEN NODE →Entity Disambiguation
Resolve competing candidate referents, test contradictory evidence and preserve unresolved or no-link states when a unique identity is not justified.
CURRENT NODEEntity Salience
Analyze which identified entities are central, supporting or peripheral to a document or passage.
OPEN NODE → ENTITY / 08 · CONTEXTContextual Relevance
Evaluate how task, domain, surrounding concepts and local evidence affect an entity’s relevance in context.
OPEN NODE → ENTITY / 09 · COVERAGEEntity Coverage
Evaluate whether the information architecture represents the important entities and semantic neighborhoods required by the topic.
OPEN NODE → ENTITY / 10 · REPRESENTATIONEntity Representation Depth
Evaluate how deeply an entity is represented through attributes, relationships, context, distinctions and supporting information.
OPEN NODE → ENTITY / 11 · AUTHORITYEntity Authority
Examine broader recognition, reputation, corroboration and source context without inventing a public Google entity-authority score.
OPEN NODE → ENTITY / 12 · FRONTIEREntity SEO & AI Search
Explore how resolved identity can support retrieval and source interpretation without guaranteeing selection, synthesis or citation.
OPEN FRONTIER →Ground the model indocumented entity systems.
These sources support specific parts of the page and should not be conflated. Google Cloud Natural Language documents entity-analysis output and disambiguation-related type handling; Google Search documents Organization structured data and general structured-data eligibility. None exposes a public Google Search “entity disambiguation score” or Google Search’s internal entity-resolution architecture.
Documents entity analysis and explicitly states that entity type can help distinguish and/or disambiguate similarly named entities. This is Cloud Natural Language documentation, not Search ranking documentation.
OPEN SOURCE →Shows entity names, types, mentions, metadata and salience returned by entity analysis; current examples also show mention-level probability in the API output.
OPEN SOURCE →Google states that Organization markup can help it understand administrative details and disambiguate an organization in Search results, and documents fields such as name, alternateName, url, sameAs and identifiers.
OPEN SOURCE →Documents content-quality and eligibility rules for structured data and reinforces that correct markup does not guarantee that a Google Search feature will appear.
OPEN SOURCE →A surface form creates ambiguity.Evidence narrows the candidate set.
Entity disambiguation is not about selecting the most famous interpretation. It is about choosing the best-supported referent when competing candidates exist, preserving uncertainty when evidence is insufficient, and allowing a no-link state when the intended referent is absent from the available catalog.