A new preprint presents OpenSubAffil, a dataset that maps raw affiliation strings in OpenAlex to disambiguated sub-institutional entities and inferred parent-child relationships. The results suggest that identifying and grouping institutional names is more reliable than reconstructing every deeper layer of an organisation’s internal structure.
The final release covers 18,635 educational institutions. It contains 39,932,792 raw affiliation strings, 638,843 disambiguated sub-institutional entities, 54,116,686 mapping records and 638,843 parent-child relationships.
From raw names to a structured dataset
The researchers used the publicly accessible August 2025 CWTS version of OpenAlex. That snapshot contained 270 million scholarly works and 142 million raw affiliation strings. Starting with records already linked by OpenAlex to at least one top-level institution, they retained English-language records: 77,681,216 of 93,970,964 input records, or 82.67%.
The primary release is limited to educational institutions, and the processing covered English-language affiliation strings. The dataset therefore represents a defined part of the OpenAlex universe rather than a complete map of every organisation appearing in scholarly affiliations.
The construction pipeline had four broad stages: extracting and filtering OpenAlex data, identifying sub-institutional name spans with two-stage named-entity recognition, clustering similar names, and reconstructing the hierarchy. Named-entity recognition here means software that identifies the parts of an affiliation likely to refer to a subunit.
During preprocessing, Gemini 3 Flash was used through its batch API to expand abbreviated subunit names into formal English forms. Fuzzy matching then assigned those entities to likely parent institutions, and fewer than 4% of records were excluded because that assignment was considered unreliable.
The researchers embedded normalized names with Qwen3-Embedding-4B and clustered them within each institution using average-linkage agglomerative clustering, with a cosine-distance threshold of 0.15. An intermediate 95% cumulative-coverage filter retained 1,556,977 canonical names across 66,332 institutions, covering 53,262,990 raw affiliation strings.
To infer hierarchy, the pipeline built networks from names that appeared together in affiliations and used several signals to score the likely direction of parent-child links. It then selected a tree rooted at each top-level institution.
Name matching performed strongly in expert checks
The name-disambiguation test used institutions and entities grouped into size or variant-count tiers. The researchers selected 10 institutions from each tier and sampled 10 entities per institution, producing a validation set of 300 entities across 30 institutions. Two experts judged 2,817 original members and 527 candidate members; their raw agreement was 90.22%, with a Cohen’s kappa of 0.67.
On that expert sample, macro precision was 0.884, macro recall was 0.941 and macro F1 was 0.888. The corresponding micro scores were 0.853 for precision, 0.934 for recall and 0.892 for F1. In plain terms, recall—the share of judged matches recovered—was higher than precision, the share of proposed matches judged correct.
The scores come with an important qualification: recall was measured only inside a pool of candidates reachable through textual similarity. That means the test does not establish how well the system would find every legitimate name variant outside that pool, and the analysis reported no confidence intervals.
The deeper the hierarchy, the harder the test
Hierarchy validation compared the inferred links with two external benchmarks. The Wikidata-derived benchmark contained 6,790 parent-child edges across 705 institutions. The supplementary GERiT benchmark contained 15,302 edges across 296 institutions.
Against Wikidata, the all-edge evaluation produced macro precision of 0.940, macro recall of 0.793 and macro F1 of 0.818. The micro figures were 0.830, 0.777 and 0.803. Results weakened when the test focused on nested links—deeper relationships within the hierarchy: across 60 institutions, nested macro F1 was 0.266 and macro recall was 0.234.
The GERiT comparison was substantially lower. For all edges, micro precision was 0.329, recall was 0.117 and F1 was 0.173, while the macro F1 was 0.304. For nested edges, micro F1 fell to 0.085 and macro F1 to 0.045. The analysis notes that GERiT contains deeper organisational levels that are often omitted from author affiliations.
The pattern points to a clear boundary for the resource. Grouping different strings that refer to the same sub-institution can be tested directly against expert judgments, but a full organisational tree must infer links from how names appear together. Intermediate units that authors rarely mention may therefore be absent from the reconstructed hierarchy.
A useful foundation, with a defined scope
OpenSubAffil is presented as a foundation for higher-resolution analyses within institutions, but the release does not establish complete organisational trees. Its coverage is restricted to educational institutions and English-language affiliations, and complete hierarchies may require institutional or national-directory data. Benchmark results also depend on whether names can be aligned between datasets.
The dataset is publicly available on Zenodo under a CC BY 4.0 licence. Expert-annotation files, the processed Wikidata benchmark and the OpenAlex education-institution ROR mapping are available in the accompanying code repository; GERiT source data are not redistributed. The source code is available on GitHub.
The manuscript is an Article in Press and an unedited version. It was received on 7 May 2026 and accepted on 12 August 2026, and the publication notice warns that errors may affect the content before final publication.
Paper data and sources
Original title: A large-scale dataset of sub-institution name disambiguation and hierarchical structures from OpenAlex
Authors: Zhentao Liang, Jin Mao, Gang Li
Journal/Repository: Liang, Z., Mao, J. & Li, G. A large-scale dataset of sub-institution name disambiguation and hierarchical structures from OpenAlex. Sci Data (2026)
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: 10.1038/s41597-026-08136-9
Original paper · Full text