Datasets:

jbellis
/

lexonomy

license: cc0-1.0
language:
  - en
pretty_name: Lexonomy
size_categories:
  - 1K<n<10K
tags:
  - lexicography
  - ontology
  - semantics
  - taxonomy
  - age-of-acquisition
  - vocabulary
  - lexical-resource
  - human-ai-collaboration

Lexonomy

A hierarchical ontology of English vocabulary in which every placement is an adjudicated argument — and the placement order has been tested against empirical age-of-acquisition norms.

Lexonomy is a corpus of ~6,000 parent–child–qualifier triples ("slips") organizing English words into a single derivation tree rooted in thirteen primes. Its distinguishing property is not size but adjudication density: every edge was individually argued under a written constitution, every mutation carries its own provenance and undo, and the corpus's central structural claim was measured against the Kuperman age-of-acquisition norms (31,105 words).

The empirical result

The corpus's ordering law — Temporal Coherence — holds that a parent concept should not be demonstrably younger than its child: the hierarchy should recapitulate the chronological growth of language. This is a testable claim, and it was tested:

Every placement was swept against the Kuperman AoA norms (Kuperman, Stadthagen-Gonzalez & Brysbaert, 2012; 31,105 words) with a 2.0-year margin.
The sweep produced 492 corrections, each individually adjudicated by the human curator with the engine's full argument on record.
Measured TC debt fell from 808 to 429 suspect placements; the remaining floor is nonzero by design (human-cleared dismissals stay counted as data-suspect forever — the metric gauges the data condition, not the adjudication).

Correction notes survive in the notes column, so the reasoning behind each AoA-driven change is inspectable row by row.

The framework

The generative rule ("Bellis's Law"): every non-prime concept equals a parent word plus a differentiating qualifier —

child = parent⟨qualifier⟩

The qualifier is a subtraction, not an addition: it names what distinguishes this child from its siblings under the same parent. A placement succeeds when the fully qualified breadcrumb from child to prime reads as a stepwise definition.

Thirteen primes form the floor (stored with parent = [prime]): the five physical senses (smell, taste, hear, see, touch); six intellectual prime concepts (you, thing, here, is, now, do), corresponding to who/what/where/why/when/how; and yes, no. Primes are generated, not carved; everything above them is subsetting.

The framework deliberately underdetermines the tree: two competent curators can disagree about a placement. Lexonomy's aim is to make such disagreements atomic and arguable — one edge, one qualifier, one verdict — rather than to claim oracle answers.

Data structure

One CSV, one row per slip.

column	meaning
`child`	the word being placed
`parent`	its parent word, or the sentinel `[prime]` (root by design, 13 rows). A second sentinel, `[orphan]` (placed, parent not yet ruled), is defined by the schema but currently empty — every non-prime row has a parent
`qualifier`	the differentiating word (bare canonical value — no bracket notation in the data)
`qual_kind`	the qualifier's class: `operator` (ordinary lexical subtraction), `variant` (mechanical/grammatical derivation, e.g. maker = make⟨noun⟩), `negation` (total subtraction: the child is the parent's direct negation), or empty (unqualified — a standing work item, not a claim)
`category`	curator's category tag, where assigned
`session_id`	provenance tag for the pass that last ruled the row
`notes`	full curation trail: prior placements (`was parent⟨qualifier⟩`), law citations, AoA arguments, recorded dissents (`[alt: …]`). Append-only; loose text by design
`status`, `gm_`, `wl_`	internal worklist/game flags; safe to ignore for analysis
`created_at`	row creation timestamp

Notation appearing in the data. Word values are bare strings with two conventions: square-bracket sentinels occur only in parent ([prime], [orphan], documented above), and a parenthesized sense tag disambiguates polysemes in child/parent — e.g. scale(weight) vs. scale, feel(perceive), marked(adjective) (37 rows). The notes column additionally uses the shorthand parent✖qualifier to cite placements compactly — music✖lyrical means "music qualified by lyrical"; a (was parent✖qualifier) clause records the placement a correction replaced.

Kind census at export: operator 4,326 · variant 1,412 · negation 156 · unqualified 78.

A note on synonymy: rows with qualifier synonym (kind variant) store a peer relationship by database contrivance — the less-common term is filed under the more-common one for storage convenience, but no subordination is claimed. True synonymy proved rare in practice.

Method: constitutional human–AI curation

The corpus was curated by a single human curator working with an LLM engine under a written constitution — a fixed document of laws (temporal coherence; the least-claiming word upstream; qualifier economy; differential duty within a family; synonym skepticism; no cycles; authorized insertion of missing generations) sent with every engine call. The process:

The family (one parent and all its children), not the slip, is the unit of adjudication — a qualifier's job is defined relative to its siblings.
Batch passes ran two independent draws per family; only consensus committed. Keep-vs-change disagreement resolved to incumbency with the dissent recorded ([alt: …] in notes); change-vs-change routed to the human.
Certificates, not attendance: a family counts as done only when every slip was adjudicated; certificates are recorded claims that can be challenged and re-run.
The human curator is the highest court; the engine drafts and argues, the curator rules.

The full constitution is included in this repository (CURATION_CONSTITUTION.md) both as documentation and as a reusable methodology artifact for AI-assisted corpus curation with auditable human authority.

Intended uses

Computational lexical resources: a small, densely adjudicated complement to WordNet-scale resources; every edge carries its argument.
Language acquisition / psycholinguistics: the TC-vs-AoA sweep is a worked example of testing a hand-built hierarchy against acquisition norms; the corrections are labeled training/evaluation material for placement-ordering models.
Lexical semantics: a corpus of ~6,000 minimal-differentia judgments (which single word distinguishes fracture from crack?) with recorded alternatives and dissents.
Human–AI methodology research: a complete, documented instance of constitutional LLM curation with provenance.

Limitations

Single curator. All final rulings are one person's; the constitution and per-row dissent trail are the check, not a second judge. Inter-annotator agreement is an open experiment (the repository owner welcomes it).
English only, and the AoA validation inherits the coverage limits of the Kuperman norms.
The core framework is axiomatic. The prime roster and the qualifier-as-subtraction rule are stipulated, not derived; the empirically tested component is the placement ordering (Temporal Coherence).
78 rows are unqualified (placement filed, differentia not yet ruled) — visible work state, not noise. Every non-prime row has a parent (no orphans at publication).

Citation

@misc{bellis2026lexonomy,
  author = {Bellis, Jack},
  title  = {Lexonomy: a hierarchical vocabulary ontology with age-of-acquisition-validated placement order},
  year   = {2026},
  url    = {https://lexonomy.org}
}

License

CC0 1.0 (public domain dedication) — free to use, share, and adapt for any purpose, no attribution required. Citation is appreciated but not a condition.