Skip to content

Commit 0e8384b

Browse files
erinepshovel-codeerinepshovel-code
committed
research(english-gonol): full OEWN definition re-affixiation run
Run the complete pinned OEWN 2025 corpus through English Gonol Construction without sampling, hash placement, carrier buckets, or lexical rebuilds: 122520 word gonols closed once and reused, 53060 multiword lemma compositions, 185155 definition gonols from ordered constituent word gonols, and 185155 ordered affixiation records. Exact UCNS orthogonal-affixiation geometry is unresolved, so every affixiation record exposes geometry_state=hmmm with no invented substitute. Frozen manifest committed; 265MB records.jsonl is local persisted state (gitignored) bound by records_sha256. The 1-7 carrier sweep is unchanged as the control experiment.
1 parent 457b062 commit 0e8384b

6 files changed

Lines changed: 873 additions & 0 deletions

File tree

‎research/english-gonol/.gitignore‎

Lines changed: 9 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,9 @@
1+
__pycache__/
2+
*.py[cod]
3+
.pytest_cache/
4+
5+
# Full-corpus affixiation records are large local persisted state. The
6+
# deterministic manifest (which binds the records file via records_sha256)
7+
# is committed; the records file itself is reproduced with the runner.
8+
experiments/oewn-affixiation-v0/records.jsonl
9+
experiments/oewn-affixiation-v0/records.worker-*.jsonl

‎research/english-gonol/README.md‎

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -119,6 +119,19 @@ python -m english_gonol.orthogonal_carrier_sweep \
119119

120120
See [`docs/orthogonal-carrier-sweep-v0.md`](docs/orthogonal-carrier-sweep-v0.md).
121121

122+
The full-corpus definition re-affixiation run (no sampling, no hash
123+
placement, no carrier buckets; reuses each closed word gonol once):
124+
125+
```bash
126+
python -m english_gonol.definition_affixiation_run \
127+
--source-root /path/to/oewn-2025/src/yaml \
128+
--out-dir experiments/oewn-affixiation-v0 --workers 2
129+
```
130+
131+
See [`docs/oewn-orthogonal-affixiation-v0.md`](docs/oewn-orthogonal-affixiation-v0.md).
132+
The large `records.jsonl` is local persisted state and is not committed; the
133+
committed `manifest.json` binds it via `records_sha256`.
134+
122135
## Research status
123136

124137
Standing: **stack-local research, not canon**. The construction is an

‎research/english-gonol/docs/oewn-orthogonal-affixiation-v0.md‎

Lines changed: 124 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,124 @@
1+
# OEWN 2025 full-corpus orthogonal definition affixiation — v0
2+
3+
Status: `UNRESOLVED` experimental full-corpus run. Standing: implemented
4+
experimental candidate inside English Gonol Construction only. Nothing here is
5+
UCNS geometry canon, an English lexical truth claim, a word embedding, or a
6+
measurement of semantic quality.
7+
8+
## Question
9+
10+
Run the **complete** pinned OEWN 2025 corpus through English Gonol
11+
Construction, reusing every already-closed word gonol, and affixiate each
12+
word's ordered definitions orthogonally without imposing any semantic axis or
13+
carrier bucket beforehand.
14+
15+
Boundary contract applied:
16+
17+
- **Do not sample**: every lexeme, sense, and definition is processed.
18+
- **Do not hash-place**: gonol identity comes from exact surface identities.
19+
- **Do not bucket words onto carriers**: no carrier or angle cell exists here.
20+
- **Do not stop on collisions**: there is no collision concept in this run.
21+
- **Do not rebuild lexical identity**: each unique surface is closed once and
22+
reused everywhere it appears.
23+
- **Do not invent orthogonal geometry**: the exact UCNS orthogonal-affixiation
24+
geometry remains unresolved, so the run closes every definition gonol from
25+
existing authority and exposes the affixiation boundary as `hmmm` with no
26+
substitute.
27+
28+
The earlier `1..7` hash-carrier sweep is unchanged and remains a control
29+
experiment; it supplies no placement rules to this run.
30+
31+
## Method
32+
33+
1. Load the pinned OEWN 2025 YAML tree
34+
(`globalwordnet/english-wordnet@dc343f2683279ecbb13fab4e2fd778d7b162d287`,
35+
tag `2025-edition`, 73 files, verified counts).
36+
2. Collect every surface used by the corpus: lemmas, lemma token splits,
37+
forms, form token splits, and every whitespace token of every definition.
38+
Single-token surfaces become word gonols; multiword surfaces are
39+
compositions, not word gonols.
40+
3. Close each single-token surface exactly once:
41+
`construct_gonol(scale="word", source=surface, source_id="oewn:surface:<surface>")`.
42+
4. Reuse that one word gonol everywhere the surface appears. Multiword lemmas
43+
that own senses are composed from their ordered constituent word gonols
44+
through the existing recursive scale
45+
(`scale="recursive"`, `relation="oewn:lemma:<lemma>:<pos>"`).
46+
5. For every sense of every lexeme, for every definition of its synset, close
47+
a definition gonol from the definition's ordered constituent word gonols:
48+
`construct_gonol(scale="definition", participants=(G(w_1),...,G(w_m)),
49+
relation="oewn:sense:<sense_id>", source_id="oewn:def:<sense_id>:<index>")`.
50+
6. Order each word's definition gonols and record the requested affixiation
51+
sequence `G(w); affixiate; D_1 ⊥ D_2 ⊥ ... ⊥ D_n` with
52+
`geometry_state = hmmm` and the UCNS boundary reason. No substitute
53+
geometry is implemented.
54+
7. Persist word identity, sense identity, constituent identities, definition
55+
order, orthogonal affixiation order, provenance digests, corpus hashes,
56+
constructor and UCNS authority identities, and a deterministic replay
57+
digest.
58+
59+
## Frozen run
60+
61+
- Module: `english_gonol.definition_affixiation_run`
62+
- Manifest: `experiments/oewn-affixiation-v0/manifest.json` (committed)
63+
- Records: `experiments/oewn-affixiation-v0/records.jsonl` (545,890 lines,
64+
~265 MB; local persisted state, not committed, bound by `records_sha256`)
65+
- Replay:
66+
`python -m english_gonol.definition_affixiation_run --source-root /path/to/oewn-2025/src/yaml --out-dir experiments/oewn-affixiation-v0 --workers 2`
67+
68+
## Results
69+
70+
| quantity | value |
71+
|---|---|
72+
| OEWN lexeme entries | 135,969 |
73+
| OEWN synsets | 107,519 |
74+
| OEWN senses | 185,129 |
75+
| OEWN relations | 244,727 |
76+
| word gonols closed | 122,520 |
77+
| multiword lemma compositions | 53,060 |
78+
| definition gonols closed | 185,155 |
79+
| affixiation records | 185,155 |
80+
| source tree sha256 | `3a46546a1ffbb4aed98990535ad5155c69be12ad09fdf093701b257d2a3e468f` |
81+
| records sha256 | `0eb37abb8975a1896bb80dfa02cb11b736cc352a798ee147263560113eb9ee9a` |
82+
| replay digest | `a67ed0a1af86710fa880a6b5b6eae0ba3861a22bcd7b12076996dde6cf3bd729` |
83+
84+
- `word_gonols` = unique single-token surfaces (lemmas + form tokens +
85+
definition tokens). Each was closed once and reused; no lexical identity was
86+
rebuilt.
87+
- `definition_gonols` = sense-definition pairs over the whole corpus, i.e.,
88+
every definition belonging to every sense of every word. No sense or
89+
definition was sampled out.
90+
- `affixiation_records` = one per definition gonol, carrying its per-word
91+
order, the previous orthogonal target, and the exposed `hmmm` geometry
92+
boundary.
93+
94+
## What is constructed vs what remains boundary-exposed
95+
96+
Constructed and persisted (existing authority):
97+
98+
- the complete closed word gonol registry for the corpus;
99+
- recursive compositions for every multiword lemma that owns senses;
100+
- a definition gonol for every sense-definition, composed from the exact
101+
ordered constituent word gonols;
102+
- per-word ordered affixiation sequences.
103+
104+
Boundary-exposed, not implemented:
105+
106+
- the geometric realization of `D_1 ⊥ D_2 ⊥ ... ⊥ D_n`. The exact UCNS
107+
orthogonal-affixiation geometry is unresolved, so every affixiation record
108+
carries `geometry_state = hmmm` and the reason string. No orthogonal axis
109+
was invented and no placement was substituted.
110+
111+
The resulting persisted structure is therefore the full recursive
112+
word -> definition -> constituent word gonol graph, with the orthogonal
113+
affixiation geometry explicitly pending UCNS authority.
114+
115+
## hmmm
116+
117+
- Emergent-relationship analysis is deferred until after construction, as
118+
instructed; it has not been performed here.
119+
- The exact UCNS orthogonal-affixiation geometry remains unresolved; until it
120+
is bound, `D_1 ⊥ D_2 ⊥ ... ⊥ D_n` cannot be geometrically realized.
121+
- Whether definition gonols should bind the raw definition text as a source
122+
unit (rather than as a receipt field only) is an open construction decision.
123+
- The word surface registry is surface-keyed; exact homograph/part-of-speech
124+
identity is preserved in the sense records, not by splitting word identity.

0 commit comments

Comments
 (0)