ANNOTATION_GUIDELINES.md

# Annotation Guidelines

Version 1.0. Defines what counts as an entity span in this dataset. This is the
specification the extraction was held to, the rubric the human review judged
against, and the answer to "what scheme is this?"

---

## 1. Scope

This dataset annotates **surface spans of named entities in reference
transcripts of English speech**, for measuring whether an ASR system
transcribes those entities correctly.

It approximates **OntoNotes 5.0** name conventions, with the deviations in §5.
It is not a full OntoNotes annotation: only the types in §2 are covered, spans
are flat, and no coreference is recorded.

This is not a text-NER training corpus. Spans are targets for transcription
accuracy. Where the two purposes conflict, transcription accuracy wins — this
is why titles are excluded (§4.3) and why the minimum-span rule (§4.1) applies.

## 2. Entity types

| type | covers | spans |
|---|---|---|
| `PERSON` | named individuals | 1,193 |
| `GPE` | countries, states, cities, towns — polities | 857 |
| `ORG` | companies, agencies, institutions, teams, committees | 511 |
| `NORP` | nationality, religious, ethnic and political groups, including adjectives | 246 |
| `LOC` | physical places that are not polities: regions, landmarks, water, celestial bodies | 223 |
| `OTHER_NAMED` | named events, works, products, laws and facilities | 215 |

Distinctions that decide edge cases:

- A company named after a place is `ORG`, not `GPE`. Northern Rock is a bank.
- A street is `LOC`, not `GPE`.
- An ethnic or national group is `NORP`, not `PERSON`.
- A stadium or building is `OTHER_NAMED`, not `LOC`.

### Where these came from

The source delivery annotates three levels: `Noun Type`, `Entity Type`,
`Entity subtype`. This release carries a mapping of the middle level:

| delivery label | released as | method |
|---|---|---|
| Names | `PERSON` | deterministic rename |
| State based location | `GPE` | deterministic rename |
| Geographic location | `LOC` | deterministic rename |
| Institutions | `ORG` | deterministic rename |
| Ethnicity/nationality demonym | `NORP` | retyped, see below |
| Concepts, Objects, Living things, Titles | classified individually | see below |

2,813 spans were renamed by rule with no model involvement.

**`NORP`.** The delivery files demonyms under Noun Type `Person` with Entity
subtype `Ethnicity/nationality demonym`. `American` is not a person. These were
retyped rather than deleted, since the annotation itself is intentional
upstream; only the label was wrong.

**`OTHER_NAMED` and the removals.** 522 spans carried delivery labels that are
not entity types — a class holding real entities (A World's Fair, Operation
Dragoon) alongside things that are not names at all (planned lots phase three,
coffee shop or diner). Each was classified by two independent model passes over
a closed inventory. Raw agreement 82.5%, Cohen's kappa 0.785. After collapsing
the fine-grained classes the two passes could not reliably separate — EVENT,
WORK_OF_ART, PRODUCT, LAW, FAC — into `OTHER_NAMED`, 351 were auto-decided and
69 were adjudicated by hand. 90 spans were judged not to be entities and were
removed from the release.

## 3. The exactness contract

Every span is an **exact, contiguous substring of the reference transcript**,
character for character. Verified on every build; the release process fails if
it is ever violated.

- Case follows the reference. Belebele/FLEURS references are cased; MLCommons
  references are verbatim lowercase. A `PERSON` span in MLCommons is
  `sami baghdady`, not `Sami Baghdady`. Lowercase is correct, not a defect.
- No normalization, expansion or correction is applied to spans. A misspelling
  in the reference appears in the span.
- Numerals appear as the reference has them, spelled out or not.
- 98.8% of spans additionally carry character offsets (`start`, `end`). The
  remainder are cases where one string was annotated more times than it occurs.

The canonical scorer normalizes case and punctuation, so a system that
capitalizes correctly is not penalized on the lowercase subset.

## 4. Boundary rules

### 4.1 Minimum span
The shortest contiguous string that names the entity. No context, no
disambiguation, no trailing generic nouns. `adsb system` -> `adsb`.

### 4.2 Possessives
The possessive clitic is excluded: `Vatican City's` -> `Vatican City`.

An internal possessive that is part of the name is **kept**:
`Addenbrooke's Hospital`, `St. Peter's Square`, `Children's Hospital of
Pittsburgh` are each one span. Splitting on the apostrophe is wrong for the
common case and was reverted during preparation.

### 4.3 Titles and roles
Excluded from `PERSON` spans, before or after the name.
`vice chair sami baghdady` -> `sami baghdady`. `selectman jim williams` ->
`jim williams`. Rationale: the title is a common noun any recognizer gets right,
and including it dilutes the measurement. Exception: a title inside an
organizational name stays (`Office of the Town Administrator`).

### 4.4 Nationality, religious and political adjectives
Annotated as `NORP`, not folded into an adjacent span.
`Vichy French` -> `Vichy` as `LOC`; the adjective is separate.
`the German company Siemens` -> `Siemens` as `ORG`.
Compound demonyms are single units: `South African` is one `NORP` span, never
truncated to `South`.

### 4.5 Adjacent entities
Two entities side by side are two spans. `Boston Massachusetts` is two `GPE`
spans. Apposition and juxtaposition both fall under this.

### 4.6 Determiners
A leading `the` is excluded unless part of the official name.
`the Pentagon` -> `Pentagon`. `The Hague` keeps it.

### 4.7 Organization suffixes
Legal and corporate suffixes that are part of the name are included
(`Acme Corp`, `Northeastern University`). Trailing descriptors are not
(`the Acme Corp facility` -> `Acme Corp`).

### 4.8 Disfluencies
Spontaneous speech contains restarts. Annotate the completed mention only:
`sami— sami baghdady` -> `sami baghdady`.

### 4.9 Repeated mentions
Every mention is annotated and scored independently.

### 4.10 Nesting
Spans are flat and non-overlapping. Where a name contains another
(`Bank of Boston`), only the outer entity is annotated.

## 5. Deviations from OntoNotes 5.0

1. Reduced inventory. No DATE, MONEY, CARDINAL, PERCENT, TIME, QUANTITY.
2. EVENT, WORK_OF_ART, PRODUCT, LAW and FAC are merged into `OTHER_NAMED`,
   because two independent annotation passes could not separate them reliably.
   Publishing distinctions that cannot be reproduced would be misleading.
3. Flat spans; OntoNotes permits nesting.
4. Case follows the source transcript, including all-lowercase references.

## 6. Difficulty tiers

| tier | definition | example |
|---|---|---|
| `A` | Discriminative. Multi-token names, rare surnames, uncommon place and organization names. | `jong seork park`, `wellesley town hall` |
| `B` | Trivial. Bare common given names, very high-frequency place names, and all demonyms. | `david`, `jim`, `Europe`, `american` |

**Report tier A only for the headline metric.** Tier B spans are transcribed
correctly by essentially any English recognizer and do not discriminate between
systems. They are retained so users can re-tier or report both; a tier A+B
figure is inflated and not comparable.

2,334 spans are tier A and 911 are tier B (28.1%). Tier B is heavier in the
MLCommons subset (33.5%) than in Belebele/FLEURS (22.6%), because meeting
speech is full of bare first names.

Tier assignment maps §2 types onto coarse classes
(`PERSON`->person, `GPE`/`LOC`->location, `ORG`->organization, `NORP`->demonym,
`OTHER_NAMED`->other), then applies an embedded frequency list. The list is in
`assign_tiers.py` so the assignment is reproducible without a network fetch;
`--audit-types` prints the mapping and reports anything unmapped.

## 7. Annotation provenance

Both the type labels and the surface spans originate from automated processes,
not from human annotators writing from scratch.

- Spans were recovered by LLM extraction after a capitalization heuristic
  returned zero spans on the lowercase MLCommons references.
- Types were derived from the source delivery as described in §2.
- A deterministic boundary sweep then applied §4, re-validating every edit
  against §3 and reverting anything that broke the contract.
- A stratified sample was reviewed by hand; the result is in the dataset card.

Substring validation confirms a span **exists** in its reference. It does not
confirm the span is **correct**. The card reports those separately.

## 8. Review procedure

Each sampled span is judged against this document:

| verdict | meaning |
|---|---|
| `correct` | right type, right boundaries |
| `boundary` | right entity, wrong extent (§4 violation) |
| `type` | right extent, wrong type (§2) |
| `spurious` | not an entity of any covered type |
| `tier` | correct, but tiered wrong (§6) |

Sampling is stratified across subset and alignment status. The reported rate is
`correct / n` with a Wilson score interval. Reviewers read the full transcript,
not only the span — several defect classes are visible only in context.

## 9. Known limitations

- Boundary artifacts remain in an unmeasured portion of unreviewed spans.
- Three types (`OTHER_NAMED`, and the `NORP` retypings) involved model
  judgement rather than a rule; the agreement figure is in §2.
- English only.
- Two domains: read encyclopedic prose, and US municipal and legislative
  meetings. Do not generalize to conversational, telephony or broadcast speech.
- MLCommons references are verbatim transcripts and carry disfluencies and
  transcription noise inherited from the source corpus.
- 47 clips carry zero annotations after non-entity removal.

## 10. Changelog

| version | change |
|---|---|
| 1.0 | initial release |