Can a full name resolve every case?

No. Different people can share a full name, and one person can publish under several forms.

Is disambiguation always performed by machine learning?

No. Identifiers, explicit rules, author confirmation and manual curation can contribute alongside learned methods.

Does a matching topic prove the author is the same?

No. Researchers in the same field often study similar topics. Topic similarity is supporting evidence, not proof.

AI Glossary / Data engineering and architecture / Intermediate

Author name disambiguation

Definition

Author name disambiguation identifies which author mentions in publication records refer to the same person. It separates different people who share a name and connects the varied names or spellings used for one person's work.

What it means

A search for “J. Patel” may retrieve papers written by several researchers. Conversely, one researcher may appear under a full name, initials, a former name or a different transliteration. A name string is therefore useful evidence, but it is not a dependable personal identifier by itself.

The task concerns author mentions attached to particular publications. A paper with five authors contains five such mentions. The goal is an accurate mapping from those mentions to people.

This matters when a research portal assembles a bibliography, an organization looks for expertise or an AI assistant summarizes a researcher's contributions. A mistaken association can turn a fluent summary into a confident account of someone else's work.

How it works

Many systems first generate plausible candidate matches, reducing the need to compare every mention with every other mention. They then examine evidence such as name variants, coauthors, affiliation history, publication dates and research topics. Rules or learned models can score pairs, while a clustering step groups mentions believed to belong to one author.

Persistent identifiers provide another route. ORCID offers an identifier independent of a person's name. Its integration guidance distinguishes identifiers collected through an authenticated workflow from identifiers entered manually or obtained with unknown authentication status. That provenance should accompany the link.

The final system needs a way to leave uncertain cases unresolved and correct mistakes. A shared institution is weak evidence when many researchers work there; changing institutions is not evidence that a person has become someone else. Decisions should consider combinations of evidence and retain the records supporting each association.

Three practical examples

A university repository brings together publications recorded under a staff member's initials and full name, while checking possible namesakes.
A research discovery service prevents two unrelated authors with the same name from sharing one publication profile.
A business research assistant verifies the author grouping behind an expert summary before attributing papers or proposing interview candidates.

Think of it this way

Think of a librarian sorting unsigned folders whose cover labels contain abbreviated names. The librarian compares the documents, collaborators and dates before deciding which folders belong together.

The analogy captures evidence gathering, but software scores are not a librarian's certainty. Sparse or incorrect metadata can support several plausible groupings. A responsible catalog preserves uncertainty instead of treating a tidy set of folders as proof that every association is correct.

Worked example

Counting the cost of a false merge

Imagine six invented author mentions in a small review set. The verified grouping is {A, B, C} for one person, {D, E} for a second and {F} for a third. There are four true same-person pairs: AB, AC, BC and DE.

A proposed system instead returns {A, B, C, D, E} and {F}. The first cluster contains 5 × 4 ÷ 2 = 10 predicted same-person pairs. Four are correct and six wrongly connect the two people. Pairwise precision is therefore 4/10 = 40%, while pairwise recall is 4/4 = 100%.

Perfect recall in this example does not make the result acceptable: it was achieved by merging two people's publication records. These teaching calculations use pairwise metrics, not the B-cubed metrics reported in some research benchmarks. A real assessment should also inspect clusters, difficult name variants and the consequences of mistaken attribution.

Illustrative case study

A fictional industry research team builds an assistant that prepares expert profiles. Its first prototype combines every record with the same surname and initial. During review, a materials researcher appears to have also written a collection of unrelated education papers.

The team treats the combined profile as a hypothesis and checks publication-level evidence. It separates the two author groups, records the supporting identifiers and preserves an unresolved older paper for manual review. The assistant is changed to show the sources behind attributed works and to avoid presenting unresolved matches as confirmed publications.

The scenario shows why a readable generated biography needs a reliable identity layer.

Common misconceptions and mistakes

Merging records because the names are exactly equal. Namesakes can publish in related fields.
Splitting records solely because affiliation or topic changed. Researchers move and develop new interests.
Treating a high similarity score as confirmed identity. The score reflects the model and available evidence.
Judging the system only on easy records. Common names, initials, transliterations and incomplete metadata can expose different failure patterns.
Confusing publication deduplication with author disambiguation. Removing two records of one paper does not identify all its authors.

Business use

For organizations building research products, define how mistaken merges and mistaken splits affect users. A false merge can misattribute achievements; a false split can hide relevant work. Choose review thresholds with those consequences in mind rather than optimizing a single headline score.

Preserve publication identifiers, source metadata and the reason for each author association. Provide a correction route and ensure corrections reach downstream summaries. Use authenticated identifiers when the workflow supports them, but still manage missing or conflicting metadata. Benchmark results from another discipline or collection should be tested against the records the product will actually process.

Check your understanding

A system groups all publications by “M. Lee” together and retrieves every genuine paper by the target researcher. Is the grouping necessarily accurate?

Answer: No. It may also include papers by other people. Coverage of the target's work must be considered alongside incorrect inclusions and the evidence for identity.

Related terms

Record linkage
Clustering
Precision
Recall
Entity resolution
Bibliographic database
Persistent identifier
ORCID
Data provenance
Human review

Learn next

Compare a publication-level author mention with a person-level profile. Trace the evidence linking the two, then examine one false merge and one false split. Study entity resolution and clustering evaluation before selecting a metric for a research catalog.

Clustering

Sources

Primary sources: Subramanian and colleagues, S2AND: A Benchmark and Evaluation System for Author Name Disambiguation, introduction and method overview: https://arxiv.org/abs/2103.07534 ; ORCID identifier explanation: https://support.orcid.org/hc/en-us/articles/360006897334-What-is-an-ORCID-iD-and-how-do-I-use-it ; ORCID guidance distinguishing authenticated and unauthenticated identifiers: https://info.orcid.org/documentation/integration-guide/orcid-id-display-guidelines/ . Relevant source material reviewed; the pair-count example and research-team scenario are original.

Previous: Authenticity Next: Authority