Glossary
Entity disambiguation
Entity disambiguation selects which possible real-world referent a mention or record denotes in context.
Identity and entity resolution
Entity disambiguation resolves ambiguity among possible referents for a name, mention or record. The same name can denote several organizations, and one organization can appear under several names. Context such as location, role, date and surrounding relationships helps distinguish candidates. The correct outcome can be no identified referent when the evidence does not support one.
In practice, a mention is compared with candidate entities and the surrounding source text. A trial sponsor named in a document should not be linked solely because a similarly named company exists in a database. Cross-document evidence may help, but it can also propagate an earlier error. A reviewable decision keeps the mention, candidate set and decisive context.
For organizations defending a linked claim, disambiguation prevents evidence about one entity from being attributed to another. Entity resolution addresses the broader problem of matching records; disambiguation often focuses on interpreting a particular mention in context. Knowledge-base population depends on it when extracted claims are attached to existing entities and relationships.
Example
A fictional paper mentions Halden; context shows a fictional research institute, so the mention is not linked to Halden Therapeutics AG (fictional).
Candidate generation and ranking
Disambiguation is usually done in two steps. Candidate generation produces a short list of possible referents for a mention, typically from a dictionary of names and aliases, or from a vector index over entity descriptions. Ranking then chooses among the candidates using the context of the mention: the words around it, the other entities mentioned in the same document, and prior information such as how often a name refers to each candidate. The paper by Hoffart and colleagues on robust disambiguation is a well-known early system that combined these signals and enforced coherence among the entities chosen in a document. The task is also called entity linking, as in the Wikipedia article, and the textbook by Jurafsky and Martin covers the standard methods.
Recent systems use dense retrieval and language models for both steps. The quality of candidate generation limits everything that follows, as with blocking in record linkage: if the right entity is not in the short list, the ranker cannot choose it. A correct system also needs a way to say that the mention refers to nothing in the reference list, which is the case that most often goes wrong.
The not-in-list case and ambiguity that cannot be resolved
Two situations deserve explicit handling. In the first, the mention denotes an entity that is not in the reference list, for example a newly founded company. A system that must always return a candidate will return the nearest one and create a false link. The remedy is an explicit null outcome with a threshold, and a queue of unlinked mentions for review and possible addition. In the second, the context does not settle which of two entities is meant, and the honest output is a ranked list with scores, or an unresolved mark, not a confident pick.
Evaluation should reflect both. Report accuracy on linkable mentions and, separately, the rate at which unlinkable mentions are correctly left unlinked. Slice by entity popularity, since systems tend to do well on common entities and poorly on obscure ones, which are the ones that matter in a long-tail professional domain. The article on confidence-aware truth discovery for long-tail evidence describes a related problem of handling thin evidence.
A worked example
A news item says: Halden announced a partnership with a university in Basel. The mention Halden could denote Halden Therapeutics AG (fictional), a fictional Halden Pharma GmbH, or a place name. Candidate generation retrieves all three from the alias list. Context favours the first two, since the sentence concerns a partnership. The text elsewhere in the article names a drug programme that matches the first company's pipeline and not the second's, so the ranker prefers the first with a score of 0.78. The threshold for automatic linking is 0.85, so the mention is stored as a candidate link awaiting confirmation, with the supporting passage. An analyst confirms it. Had the mention been linked automatically at 0.78, a coincidence of names would have entered the record as an apparent fact.
Related terms and boundaries
Disambiguation links a mention to an entry; entity resolution and record linkage link records to each other. It is the step in information extraction that turns names into entities, and it supplies the entities for knowledge-base population. A mistake here is propagated as a wrong identity in every later step, including truth discovery.
Common misreadings
- Linking to the most famous namesake by default.
- Forcing a link when context is insufficient.
- Ignoring temporal or geographic clues.
Appears in
- Vector search vs entity lookup
Vector search finds semantically similar material; entity lookup tests which specific subject a record denotes.
- Resolving identity by rule
Make each same-entity or different-entity decision explicit, evidenced, reviewed and reversible.
- CUSIP
A nine-character identifier used for securities in North American markets.
- European Case Law Identifier (ECLI)
A common identifier format for European judicial decisions.
- Global Trade Item Number (GTIN)
A GS1 identifier for trade items at a defined packaging or presentation level.
- International Securities Identification Number (ISIN)
A twelve-character code for identifying a financial instrument under ISO 6166.
- ORCID iD
A persistent identifier that distinguishes contributors to research.
- Unique Ingredient Identifier (UNII)
An FDA-assigned identifier for a substance based on its defined identity.
References
- Shen, Wang and Han (2015), Entity Linking with a Knowledge Base: Issues, Techniques, and Solutions, IEEE Transactions on Knowledge and Data Engineering 27(2).
- Hoffart et al. (2011), Robust Disambiguation of Named Entities in Text, EMNLP.
- Wikipedia, Entity linking.
- Jurafsky and Martin, Speech and Language Processing (3rd ed. draft).
- Christophides et al. (2020), End-to-End Entity Resolution for Big Data: A Survey, ACM Computing Surveys.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Entity disambiguation" bear on a decision of yours?
This entry sits under identity and entity resolution. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.