From Tags to Testable Definitions: A verbatim-grounded framework for cross-disciplinary concept analysis

The live app: shir-openu.github.io/meta-tagging-platform-prototype — assemble a corpus, submit a definition of your own, and have it scored against that corpus in one run.

Every tagged paper: shir-openu.github.io/meta-tagging-showcase — 1,673 papers across 208 disciplines; 94.3% of their tags carry the exact passage that licenses them. Code and data: GitHub.

A field-neutral tag layer in which every non-trivial tag is grounded in verbatim source text, and a method that turns canonicalisation into a measurable question: concepts become records with attested senses, and candidate definitions are scored as binary decision rules against adjudicated cases.

Corpus update, 6 October 2026. Since this version of the paper, the live corpus has grown to 1,673 published records across 208 disciplines. Together they carry 123,065 anchored quotes: an anchored quote is one field of a record, or of one of its tags, whose quotation was found at an exact position in the paper’s text, and the 123,065 point at 63,387 distinct passages. 1,161 of the 1,673 records are redistributed as full text under open licences. The figures in the paper and in the paragraphs below are the ones measured for this version and are left unchanged.

Measured. 697 published records across 103 disciplines, 57,012 anchored quotes. The provenance audit re-checked 18,416 evidence strings and found three failures. Among 511 papers, 500 distinct free-text phenomenon labels produced two cross-disciplinary matches, whereas a controlled statistical vocabulary produced join keys spanning 31 disciplines. On 372 adjudicated case-definition observations from 34 papers the highest-ranked eligible definition reached MCC = 0.700; reliability on 140 blinded pairs was Cohen’s kappa = 0.547 at 78.6% observed agreement.

Rights. 111 papers are redistributed as full text under CC-BY, CC-BY-SA, CC0 or verified public domain. The other 586 are represented by metadata, tags and short quotations with a link to the source, and are never reproduced in full.

Status. A proof of concept, sized accordingly. The concept layer is declared incomplete in the paper itself: 2,103 concept records, 1,880 attestations carrying a quotation, 719 still requiring grounding. Automatic evaluation of a user-submitted definition is not yet public, and cross-corpus generalisation and a large human reliability study are development targets.

Files. The previewed PDF is the paper. The attached HTML is the extended technical documentation — the schema field by field, the provenance audit, and the confusion-matrix cells each linked to the case behind them. It is also live, and interactive, at shir-openu.github.io/meta-tagging-showcase/definitions-preprint-en_v24.html, where every section opens on click.

Earlier versions are not lost. All 30 remain under the concept DOI 10.5281/zenodo.21432188, which always resolves to the newest. The version immediately before this one is 10.5281/zenodo.22116896 (v13.13, 26 August 2026).