Two researchers named Wei Zhang, both publishing in materials science, at two different Chinese universities: which one authored a given 2019 paper? A researcher who published under a maiden name before 2015 and a married name after: are those two people or one? A name transliterated three different ways across three publishers’ house styles: one author, or three fragmented publication records? This is the author name disambiguation (AND) problem, and it sits underneath nearly everything a research office does with publication data—grant biosketches, tenure dossiers, institutional reporting, bibliometric analysis. Three distinct layers of infrastructure now attack this problem in different ways: ORCID’s claimed, persistent identifier; the algorithmic/probabilistic name-matching built into citation databases like Scopus and Web of Science; and the person-record identity resolution performed inside institutional CRIS platforms. This guide explains how each layer works, where each one fails, and how they fit together in practice.
Why author names collide in the first place
Name-based identification breaks down for several distinct, well-documented reasons, and a workable disambiguation strategy has to account for all of them, not just the most obvious one:
- Common names. Surnames like Smith, Kim, Chen, or Wang are shared by thousands of active researchers, often within the same country and sometimes within the same institution or subfield. Initials-plus-surname formatting (a near-universal citation convention) makes the problem worse by discarding exactly the information—full given names—that would otherwise disambiguate many of these cases.
- Name changes. Marriage, divorce, and personal choice all produce researchers who have published under more than one surname across a career. A disambiguation system that only matches on the current name will silently split that person’s output into two disconnected records.
- Transliteration and script variation. A single name in a non-Latin script (Chinese, Cyrillic, Arabic, and others) can be romanized multiple valid ways across different publishers, databases, and even the same author’s own past submissions, producing several string-distinct name forms for one real person.
- Affiliation and field changes. Algorithms that lean on institutional affiliation or subject-area consistency as disambiguating signals can incorrectly split a legitimate career move (a researcher changing universities or shifting into an adjacent field) into what looks like two different people.
- Homonym vs. synonym errors. The literature on this problem distinguishes two opposite failure modes: a homonym error merges two different people who share a name into one record, while a synonym error splits one person’s real output across multiple records because of name variants. Any evaluation of a disambiguation system’s accuracy has to account for both failure directions, not just one.
ORCID’s approach: a claimed, persistent identifier
ORCID (Open Researcher and Contributor ID) takes a fundamentally different approach from every method described below: instead of trying to algorithmically infer which publications belong to which person, it gives the researcher a persistent, non-reassignable identifier—a 16-digit ORCID iD—that the researcher registers once, keeps for life regardless of institution or name changes, and asserts onto their own works directly (or authorizes a trusted party, such as a publisher or funder system, to assert on their behalf). Because the identifier is independent of the name string, it survives every one of the collision scenarios above: a name change, a transliteration variant, an affiliation move, or a common surname shared with a hundred other researchers. ORCID was created specifically to address this problem—the initiative’s own framing has consistently been that name-based identification is structurally unreliable, most acutely for common names, name changes, and multiple valid name transliterations.
The tradeoff is that ORCID’s accuracy depends on adoption and correct use, not an algorithm. An ORCID record is only as complete and correct as what the researcher (or an authorized integration) has actually asserted onto it. A researcher who registers an iD but never connects it to a publisher, funder, or institutional system, or who has pre-ORCID publications that were never manually or automatically linked back, will still have gaps. See CASRAI’s guides on registering and finding ORCID iDs and optimizing an ORCID profile for the practical mechanics of closing those gaps.
Algorithmic and probabilistic name matching
Before ORCID existed, and still today for the large body of legacy publication data that predates iD adoption, citation databases solve disambiguation computationally: a matching algorithm groups publication records that it infers belong to the same person, without requiring that person to have claimed anything. The two major commercial citation databases each run their own version of this:
- Scopus Author ID is auto-generated by Elsevier’s matching algorithm once a researcher has two or more Scopus-indexed documents—it is not requested or self-registered. The exact algorithm is proprietary and undocumented in detail, but published third-party evaluations have reported precision and recall in the high-90s and mid-90s percent range respectively for Scopus author identifiers, which in practice still means a meaningful minority of researchers end up with duplicate or incorrectly split profiles, particularly for common names, affiliation changes, or subject-area shifts. Elsevier provides a correction path (the Author Feedback Wizard, plus a ‘request to merge authors’ workflow) precisely because the algorithm doesn’t self-correct. See CASRAI’s guide on creating, finding, and fixing a Scopus Author ID.
- Web of Science takes a related but distinct approach: its Author Search (surfaced in the interface as Researcher Search) groups publication records into algorithmically-inferred ‘Author Records,’ which a searcher can then refine, merge, or split manually before treating the result as a complete list, and separately supports a claimed ResearcherID/Web of Science Researcher Profile that a researcher can register directly, similar in spirit to ORCID. See CASRAI’s guide on Web of Science Author Search and disambiguation.
The published research literature on author name disambiguation (a substantial body of work spanning machine-learning classification, clustering, and network-based methods) generally follows one of two strategies: author grouping, which clusters publication records using similarity functions across features like coauthor names, subject classification, affiliation, and funding/grant information, or author assignment, which treats disambiguation as a classification problem trained on labeled examples. Both approaches are estimating a probability that two records belong to the same person from indirect signals, which is structurally different from ORCID’s model of the researcher directly asserting their own identity. That structural difference is exactly why algorithmic matching, however well-tuned, has a residual error rate that a claimed persistent identifier does not share in the same way—and why the two approaches function best as complements rather than substitutes.
How CRIS systems reconcile both approaches
Institutional Current Research Information Systems (CRIS) sit downstream of both ORCID and the citation databases, and they face the disambiguation problem at a different scale: reconciling a single person’s identity across every internal and external data source the institution pulls from—HR records, grant systems, the institutional repository, and imported publication metadata from Scopus, Web of Science, Crossref, and elsewhere. The organizing entity for this is the Person record, the join point for everything the CRIS knows about a given researcher. Under the CERIF data model maintained by euroCRIS, a Person record commonly carries an ORCID iD as a standard field (now treated as the default authoritative external signal for identity), alongside historical name variants—maiden names, transliterations, prior name forms—to support matching of legacy outputs that predate ORCID adoption or that were published under a different name. Time-bound relationships (which organizational unit a person belonged to when, which projects they were PI on and for how long) are represented through CERIF’s linking-entity structure rather than static fields, which matters directly for disambiguation: a name-and-affiliation match that would fail a snapshot comparison can still resolve correctly once the system accounts for the fact that the affiliation was only current for a bounded period.
In practice, most CRIS deployments use a layered strategy rather than picking one method: ORCID iD as the primary, highest-confidence signal where present; algorithmic/probabilistic matching (often via imported Scopus or Web of Science author identifiers, or the platform’s own internal matching) as a fallback for records without a claimed iD; and manual review or merge/split tooling as the last resort for cases both automated layers get wrong. This mirrors the same crosswalk logic that connects other identifier types across systems—see CASRAI’s identifier crosswalk guide for how ORCID, DOI, ROR, and institutional IDs interconnect more broadly, and the Researcher Identifier and ISNI dictionary entries for adjacent identifier schemes that some CRIS platforms also draw on.
Practical guidance for researchers
- Register an ORCID iD once, early, and keep it for the rest of your career—even a name change or institutional move later on will not require a new identifier.
- Connect the iD to every system that will assert works onto it: publisher submission systems, funder application portals, and your institution’s CRIS or repository, so that new publications link automatically rather than depending on a later manual match.
- Periodically search your own name in Scopus and Web of Science to check for duplicate or split author profiles, especially after a name change, institutional move, or shift into a new subfield—these are exactly the conditions that break algorithmic matching. Use the correction tools each platform provides rather than assuming the record will self-correct.
- If you have published under more than one name (including transliteration variants), add those as alternate names on your ORCID record so downstream systems have the information needed to link legacy work.
Practical guidance for research offices and CRIS administrators
- Treat a verified ORCID iD as the primary matching key wherever it exists on a Person record, and reserve algorithmic/probabilistic matching for the residual cases where it doesn’t—not the other way around.
- Capture historical name variants on the Person record explicitly, rather than relying on the current name alone, to correctly attribute legacy publications.
- Build in a manual review or merge/split workflow for ingested publication metadata; no automated matching layer, including ORCID-based matching with incomplete researcher-side data, will resolve every case correctly on its own.
- When importing from multiple sources (Scopus, Web of Science, Crossref, an institutional repository), expect and plan for genuine disagreement between sources on how many distinct ‘authors’ a given name resolves to—this is a normal consequence of each source running its own matching logic, not a data-quality bug to chase down to zero.
Frequently asked questions
Does having an ORCID iD guarantee my publications will never be confused with another researcher’s?
Not automatically. An ORCID iD eliminates the ambiguity for any work that has actually been asserted onto your record by you or an authorized system. Work that hasn’t been connected—older publications, or anything from a system that doesn’t yet integrate with ORCID—still depends on algorithmic or manual matching until it’s linked.
Why do I have two different Scopus Author IDs for what should be one profile?
Scopus Author IDs are generated automatically by an algorithm, not assigned by request, and that algorithm can split one person’s output across profiles when it encounters a name variant, affiliation change, or subject-area shift it doesn’t confidently match. Elsevier’s merge and feedback tools exist specifically to correct this after the fact.
How is a CRIS Person record different from an ORCID record?
An ORCID record is a researcher-controlled, cross-publisher, cross-funder identifier that the researcher owns for life. A CRIS Person record is institution-controlled and typically incorporates the researcher’s ORCID iD as one field among many (alongside HR data, grant history, and organizational affiliations), scoped to that institution’s own reporting and administrative needs rather than portable across institutions.
Related CASRAI resources
For the broader identifier landscape this problem sits within, see the Persistent Identifiers & Research Information Systems pillar, the ROR ID dictionary entry for organization-level identity resolution, and CASRAI’s guide on compiling one complete publication list across PubMed, Scopus, Web of Science, and Google Scholar, which runs directly into this same disambiguation problem from a workflow angle.
