WORDNET_FORMAT.md

# WordNet 3.0 Research Guide

This guide explains the structure of the WordNet files used in this dataset project,
how this adapter parses them, and how to interpret the resulting records for
research workflows.

## Scope

This project packages WordNet 3.0 lexical database files for four parts of speech:

- noun
- verb
- adjective
- adverb

The adapter supports three variants:

- `data` - synset records (`data.noun`, `data.verb`, `data.adj`, `data.adv`)
- `index` - lemma index records (`index.noun`, `index.verb`, `index.adj`, `index.adv`)
- `exceptions` - morphology exceptions (`noun.exc`, `verb.exc`, `adj.exc`, `adv.exc`)

## Source Files In This Project

Under `datasets/wordnet/data`:

- `data.noun`, `data.verb`, `data.adj`, `data.adv`
- `index.noun`, `index.verb`, `index.adj`, `index.adv`
- `noun.exc`, `verb.exc`, `adj.exc`, `adv.exc`

Important: WordNet files begin with license/header text before structured records.
Those rows are valid file content but not structured lexical records.

## Canonical Format References

Primary references for WordNet 3.0 file formats:

- Princeton `wndb(5WN)`: data/index/exception file grammar
- Princeton `wninput(5WN)`: pointer symbols and lexical conventions

See:

- [WordNet `wndb(5WN)`](https://wordnet.princeton.edu/documentation/wndb5wn)
- [WordNet `wninput(5WN)`](https://wordnet.princeton.edu/documentation/wninput5wn)

## Core Concepts

- Synset: a set of synonymous word senses.
- `synset_offset`: byte offset used as an address into `data.*` files.
- `ss_type`: part-of-speech/synset type marker (`n`, `v`, `a`, `s`, `r`).
- Pointer: typed relation from one synset (or specific word sense) to another.
- `source/target`: a four-hex-digit field that distinguishes semantic and lexical pointers.

### Synset Identity

`synset_offset` alone is not globally unique across all POS files.
Use `(synset_offset, pos)` as the unique synset key.

## Raw File Grammar

## `data.*` Files

Per `wndb(5WN)`, data lines are:

```text
synset_offset lex_filenum ss_type w_cnt word lex_id [word lex_id ...] p_cnt [ptr ...] [frames ...] | gloss
```

Where:

- `synset_offset`: 8-digit decimal byte offset.
- `lex_filenum`: 2-digit decimal lexical file id.
- `ss_type`: one of `n`, `v`, `a`, `s`, `r`.
- `w_cnt`: 2-digit hexadecimal count of words in the synset.
- `word`: token, underscores for spaces (adjectives may include markers like `(p)`).
- `lex_id`: 1-digit hexadecimal sense discriminator within lexicographer files.
- `p_cnt`: 3-digit decimal pointer count.
- `ptr`:

```text
pointer_symbol target_synset_offset pos source/target
```

- `source/target`: four hexadecimal digits, split as `ss tt`:
  - `0000` means semantic pointer (synset-to-synset)
  - non-zero means lexical pointer from source word number `ss` to target word number `tt`
- `frames` (verbs only):

```text
f_cnt + f_num w_num [ + f_num w_num ... ]
```

- `gloss`: free text after `|`.

## `index.*` Files

Per `wndb(5WN)`, index lines are:

```text
lemma pos synset_cnt p_cnt [ptr_symbol ...] sense_cnt tagsense_cnt synset_offset [synset_offset ...]
```

Where:

- `lemma`: lower-case lemma/collocation (`_` for spaces).
- `pos`: `n`, `v`, `a`, `r`.
- `synset_cnt`: number of senses/synsets for this lemma in POS.
- `p_cnt`: number of pointer symbols listed.
- `ptr_symbol`: pointer symbol types seen for lemma senses.
- `sense_cnt`: redundant historical count field.
- `tagsense_cnt`: number of tagged (frequency-ranked) senses.
- trailing `synset_offset` list length should equal `synset_cnt`.

## `*.exc` Exception Files

Per `wndb(5WN)`, exception lines are:

```text
inflected_form base_form [base_form ...]
```

These map irregular inflections to one or more base forms.

## Pointer Symbols You Will See

This project normalizes common pointer symbols to labels when parsing `data.*`.

| Symbol | Label |
| --- | --- |
| `!` | antonym |
| `@` | hypernym |
| `@i` | instance_hypernym |
| `~` | hyponym |
| `~i` | instance_hyponym |
| `#m` | member_holonym |
| `#s` | substance_holonym |
| `#p` | part_holonym |
| `%m` | member_meronym |
| `%s` | substance_meronym |
| `%p` | part_meronym |
| `=` | attribute |
| `+` | derivationally_related_form |
| `;c` | domain_of_synset_topic |
| `-c` | member_of_domain_topic |
| `;r` | domain_of_synset_region |
| `-r` | member_of_domain_region |
| `;u` | domain_of_synset_usage |
| `-u` | member_of_domain_usage |
| `*` | entailment |
| `>` | cause |
| `^` | also_see |
| `$` | verb_group |
| `&` | similar_to |
| `<` | participle_of_verb |
| `\\` | pertainym_or_derived_from_adjective |

Meaning and allowed POS combinations are defined by WordNet documentation.

## Adapter Parsing Behavior

## Variant `data` with `parse_records=true`

Adds structured fields:

- `is_record`
- `offset`, `synset_offset`
- `lex_filenum`
- `ss_type`
- `w_cnt`, `word_count`
- `words` (list of objects)
- `lemmas` (normalized with spaces)
- `p_cnt`, `pointer_count`
- `pointers` (list of objects)
- `frames` (list of verb frame objects; usually empty for non-verbs)
- `gloss`
- `parse_error` (flag when line partially matches but is malformed)

`words` entries include:

- `word`
- `lemma`
- `marker` (for adjective markers like `a`, `p`, `ip`)
- `lex_id`
- `lex_id_int`
- `word_number`

`pointers` entries include:

- `symbol`
- `label`
- `target_offset`
- `pos`
- `source_target`
- `source_word_number`
- `target_word_number`
- `is_semantic`

`frames` entries include:

- `frame_number`
- `word_number`
- `applies_to_all_words`

## Variant `index` with `parse_records=true`

Adds:

- `is_record`
- `lemma`, `lemma_text`
- `pos`
- `synset_cnt`
- `p_cnt`
- `ptr_symbols`
- `sense_cnt`
- `tagsense_cnt`
- `synset_offsets`
- `parse_error`

## Variant `exceptions` with `parse_records=true`

Adds:

- `is_record`
- `inflected_form`, `inflected_form_text`
- `base_forms`, `base_forms_text`

## Header/License Row Handling

Non-record lines are expected at the top of files.

- Parsed output marks these as `is_record=false`.
- WordNet dataset config enables automatic filtering of these rows when parsing.

Project default in `datasets/wordnet/dataset.json`:

```json
{
  "adapter_defaults": {
    "records_only": true
  }
}
```

Practical effect:

- If you run with `--parse-records`, non-record rows are removed by default.
- To keep raw non-record rows, set `records_only=false`.

## Research Workflows

## Quick CLI Usage

Parse data records and keep only true records (default behavior):

```bash
python -m tether.datasets run-dataset wordnet \
  --base-dir datasets/wordnet \
  --variant data \
  --parse-records \
  --json
```

Disable filtering for diagnostics/audit:

```bash
python -m tether.datasets run-dataset wordnet \
  --base-dir datasets/wordnet \
  --variant data \
  --parse-records \
  --option records_only=false \
  --json
```

Build parquet artifact bundle for analysis:

```bash
python -m tether.datasets prepare-dataset wordnet \
  --base-dir datasets/wordnet \
  --variant data \
  --parse-records \
  --output-dir artifacts/wordnet-research \
  --format parquet \
  --json
```

## Example Analysis Ideas

- Build graph edges from `pointers` using `(synset_offset, ss_type)` keys.
- Compare lexical vs semantic relation rates via `is_semantic`.
- Analyze gloss lengths, lemma counts (`word_count`), and pointer density (`pointer_count`).
- Filter by specific pointer symbols (for example `@` hypernym, `~` hyponym).

## Recommended Quality Checks

- Assert no remaining `is_record=false` rows in final analysis datasets when using defaults.
- Track `parse_error=true` counts as ingestion quality telemetry.
- Validate that index `len(synset_offsets) == synset_cnt` for parsed records.

## Known Caveats

- Some WordNet tokens contain punctuation and abbreviations (`A.D.`, `B.C.`).
- Not all pointer symbols apply to all parts of speech.
- Exception files may include forms not present in your selected downstream subset.
- Sense ranking in index files depends on WordNet tag counts and ordering conventions.

## Citation And License

- Citation: Miller (1995), WordNet: A Lexical Database for English.
- License: WordNet 3.0 license (Princeton University), included in source files.

Use source-level attribution and license text in downstream publications and releases as required.