ramiz0/ner-stage2-dataset-expansion

NER Stage 2 — Dataset Expansion

Published at: https://huggingface.co/datasets/ramiz0/ner-stage2-dataset-expansion

This is the Stage 2 output of a 4-stage applied NER project: Stage 1's 98 corrected records (ramiz0/ner-stage1-rulecraft-cleanup) expanded with 650 new records to a combined 748, all under the same annotation policy, then deduplicated and split into train/test.

Fields: source_text, privacy_mask (list of {start, end, label, value}). New records also carry source (wikiann | wikipedia | wikinews).

Stats: 748 total records (98 Stage 1 + 650 new). Split: 598 train / 150 test (79.9% / 20.1%, stratified by each record's dominant label).

Label	Stage 1	Final (combined)
PERSON	121	600
ORGANIZATION	99	600
LOCATION	76	600
TIMEDATE	129	600
AMOUNT	80	262
JOB	82	449
WORKOFART	56	278
PRODUCT	83	88

Labels

• PERSON: a named person, including given names and surnames. Example: <PERSON>Barack Obama</PERSON> / <PERSON>Sarah Chen</PERSON>.
• ORGANIZATION: a named company or institution. Example: <ORGANIZATION>Google</ORGANIZATION> / <ORGANIZATION>Mayo Clinic</ORGANIZATION>.
• LOCATION: a named place such as a city, country, region, street, landmark, or geographic area. Example: <LOCATION>Berlin</LOCATION> / <LOCATION>New York City</LOCATION>.
• TIMEDATE: an expression that places something on a timeline, including dates, clock times, and durations used as time. Example: <TIMEDATE>March 15, 2024</TIMEDATE> / <TIMEDATE>50 minutes</TIMEDATE>.
• PRODUCT: a named commercial product, device, or branded good. Example: <PRODUCT>iPhone 15</PRODUCT> / <PRODUCT>MacBook Pro</PRODUCT>.
• WORKOFART: a named creative or published work such as a book, film, song, or titled publication. Example: <WORKOFART>Oppenheimer</WORKOFART> / <WORKOFART>Spider-Man: Brand New Day</WORKOFART>.
• JOB: an occupational title or formal work role when it functions as such in the sentence. Example: <JOB>software engineer</JOB> / <JOB>CEO</JOB>.
• AMOUNT: a measurable or countable quantity, not a time or date expression. Example: <AMOUNT>50</AMOUNT> tickets / <AMOUNT>120,000</AMOUNT>.

Baseline labeling rules

These are the fixed main rules (unedited from the source task spec).

• A labeled mention should cover the entity itself, not the surrounding grammar. Correct: She visited <LOCATION>Paris</LOCATION> yesterday. Incorrect: labeling visited Paris yesterday as one span.
• Articles, prepositions, conjunctions, and other function words should stay outside the span unless they are truly part of the proper name. Correct: He works at <ORGANIZATION>Google</ORGANIZATION>. Incorrect: <ORGANIZATION>at Google</ORGANIZATION>. <ORGANIZATION>The New York Times</ORGANIZATION> should keep The when it belongs to the established name.
• Bare type-words and category descriptors (words that name a category rather than a specific entity) should not be labeled on their own. Examples include "person", "company", "organization", "team", "hospital", "city", "product", "book", and "quantity". The company hired 200 people. should leave company unlabeled. She joined <ORGANIZATION>Acme Corp</ORGANIZATION>. should still label the named organization. The person who called did not leave a name. should leave person unlabeled.
• A multi-word name should be one span when it forms one real named entity; separate entities should be labeled separately. <PERSON>Barack Obama</PERSON> should be one span. <PERSON>Barack</PERSON> and <PERSON>Michelle</PERSON> should be two spans.
• Coordinated names should be separate spans, unless the conjunction is part of one established name. <ORGANIZATION>Google</ORGANIZATION> and <ORGANIZATION>Microsoft</ORGANIZATION> should be two spans. <ORGANIZATION>Johnson & Johnson</ORGANIZATION> should stay one span because & belongs to the name.
• Ordinary punctuation and stray whitespace should stay outside spans, unless the punctuation belongs to the name or abbreviation. He moved to <LOCATION>Berlin</LOCATION>. should leave the final period outside the span. <ORGANIZATION>AT&T</ORGANIZATION> should keep & inside the span.
• Possessive markers should stay outside the span unless the full possessive form is the name. <PERSON>Maria</PERSON>'s laptop should leave 's outside the span.
• Labels should follow context, not surface form alone. The same words may be an organization in one sentence and a place in another. She works at <ORGANIZATION>Cambridge University</ORGANIZATION>. / The conference was held in <LOCATION>Cambridge</LOCATION>.
• Quantities used as time should be labeled as TIMEDATE, not AMOUNT. She bought <AMOUNT>50</AMOUNT> tickets. / The train arrives in <TIMEDATE>50 minutes</TIMEDATE>.

Full policy (baseline + Stage 1 additions, unchanged in Stage 2)

The baseline rules above do not cover every case. The following 14 rules were added in Stage 1 to close specific gaps found while reviewing the starter dataset, and remain in force unchanged for Stage 2.

1. TIMEDATE specificity. A span may only be labeled TIMEDATE if it denotes a specific, placeable point in time or a measurable duration — an explicit date, clock time, duration with a unit or count, or a named calendar period (holiday, quarter, season tied to context). Vague deictic or aspectual words that mark tense/recency without specifying an actual time value (e.g. now, moment, again, ever, never, before, always, yet, just, already, initially, recent, latest, long) must not be labeled, even though they are temporal in a loose grammatical sense. Test: does the span have a placeable value or measurable magnitude? (past couple of days passes — rough but real magnitude; some time fails — no magnitude at all.)

2. PERSON via referring nickname. A span may be labeled PERSON when it is a moniker, alias, or nickname used as a fixed referring label for one specific individual within the text — the same word used consistently to pick out that one person — even if it is not a conventional given or family name. Does not extend to generic role/category nouns referring to no particular individual (e.g. "waiter," "dealer"); those stay excluded under the bare-category-descriptor rule, or fall under JOB if functioning as an occupational title in context. Example: <PERSON>Dreads</PERSON>: the only guy that can dance — a nickname used repeatedly to refer to one specific performer.

3. JOB context-functional scope. JOB applies whenever an occupational noun functions as an occupational descriptor of who is doing/being something in the sentence — a named individual, a clear single referent, or a group acting collectively (e.g. "bouncers turned him away"). It does not apply when the same noun is used purely as a demographic/statistical count subject (e.g. "13.3% of the man-days... were worked by black bricklayers" — a percentage subject, not a role-in-action), nor to industry/sector/ department/team nouns that name an organizational unit rather than an occupation (e.g. "energy sector," "analytics team"). Example kept: <JOB>bouncers</JOB> at the door — group acting.

4. AMOUNT magnitude test. AMOUNT may be applied to approximate-but-real magnitude words/expressions that carry a genuine, even if imprecise, order-of-magnitude value (e.g. "hundred," "thousands," "dozens," "several hundred") in addition to exact numerals. It must not be applied to spans with zero resolvable magnitude (e.g. "many," "some," "a lot") or to words that are not quantities at all. Example kept: <AMOUNT>hundred</AMOUNT> bears — real order-of-magnitude value.

5. Honorifics excluded from all labels. Courtesy/professional honorifics attached to a name (Mr., Ms., Mrs., Dr., Prof., Cardinal, etc.) are never labeled PERSON or JOB — only the bare name is PERSON. A bare honorific does not "function as" a role description; it is a form of address, not a description of what the person does. Example: Ms. <PERSON>Charlie</PERSON> — honorific stays outside the span.

6. PRODUCT extends to named technologies without a commercial owner. PRODUCT includes any specifically named, identifiable technology, tool, language, or platform that a person can use, install, or reference by name, regardless of whether it has a single commercial/brand owner. Open standards and specs (SQL, HTML) count the same as clearly branded software. Example: <PRODUCT>SQL</PRODUCT> query engine — no single commercial owner, still PRODUCT.

7. Countries/kingdoms are LOCATION regardless of grammatical role or fictionality. A named country- or kingdom-type political entity is labeled LOCATION, whether it is the grammatical subject of an action or referred to as a place, and whether it is real or fictional. Example: <LOCATION>Spade Kingdom</LOCATION>'s commandor's poison attack — a fictional kingdom, treated the same as a real country acting as an agent (e.g. "North Korea says...").

8. Age expressions are AMOUNT. Age expressions ("N-year-old," "aged N years old") are labeled AMOUNT. They are a measured quantity/count attached to a person, not a duration doing scheduling/timing work in the sentence. Example: a <AMOUNT>19-year-old</AMOUNT>.

9. League/competition names are ORGANIZATION. Named sports leagues/competitions are labeled ORGANIZATION, consistent with treating them as a named institution. Example: <ORGANIZATION>La Liga</ORGANIZATION> side.

10. Named businesses/venues are ORGANIZATION, not LOCATION. Any named business or venue with a proper name is ORGANIZATION — a named institution you visit is still an institution, not merely a place. Applies to restaurants, venues, casinos, clinics, studios, and similar named businesses. Example: <ORGANIZATION>Joyride Taco House</ORGANIZATION> — was LOCATION in the source data, corrected.

11. Award/prize titles are left unlabeled. Award and prize titles do not fit any of the 8 labels — WORKOFART is restricted to creative/published works, and an award/distinction is not one. Since the label set is fixed, an award name that doesn't fit any label is left unlabeled rather than stretched into the closest-sounding one. Example: <ORGANIZATION>Costume Designers Guild</ORGANIZATION> Excellence Award — the granting body stays labeled; the award title itself does not.

12. Institutional documents/reports are WORKOFART. Named formal institutional documents (strategy papers, audit reports, and similar titled publications) are labeled WORKOFART, matching the "titled publication" wording directly. Example: <WORKOFART>Project Performance Audit Report</WORKOFART> (PPAR).

13. Legal citations require actual title text. A legal/regulatory citation is labeled WORKOFART only if actual title text is present (e.g. a spelled-out case name or written-out code name), not when it is a bare locator string of volume/section/page numbers with no title expressed (e.g. "462 U.S. 352, 358," "41 CFR § 60-11"). Example excluded: CFR in "41 CFR § 60-11 et seq." — no title text, just a locator.

14. Usernames/handles and transcript role-placeholders are PERSON. A username or handle that functions as a fixed referring label for one specific individual is PERSON (extends rule 2 to formatted identifiers), and dialogue-transcript speaker labels ("Person1"/"Person2") are also PERSON, since each consistently refers to one specific speaker across the dialogue. Example: <PERSON>otavio.telin</PERSON>; <PERSON>Person1</PERSON>: Which route are you interested in?

Stage 2 methodology

Sources: a mix of an existing annotated dataset remapped into our schema, and raw text freshly labeled under our policy — unimelb-nlp/wikiann (English split, CC-licensed) for PERSON/ORGANIZATION/LOCATION, and Wikipedia biography/company/creative-work articles plus Wikinews (CC BY-SA / CC BY) for JOB/TIMEDATE/PRODUCT/WORKOFART/AMOUNT-rich sentences the WikiANN pool doesn't naturally contain.

First-pass labeling: spaCy (en_core_web_sm) for PERSON/ORGANIZATION/ LOCATION/TIMEDATE/AMOUNT-adjacent categories and WORKOFART, plus curated gazetteer/trigger-word pattern matching for JOB. spaCy's native PRODUCT tag was found to have ~0% precision in QA (mislabeling people, countries, and titles as PRODUCT) and was replaced with a trigger-word pattern approach, which still only found a small number of genuine candidates — see the PRODUCT note below.

Policy correction: every candidate span passes through an automated correction pass applying the mechanically-expressible subset of the policy above (bare category words, vague TIMEDATE/AMOUNT words, honorifics, age expressions, venue/league LOCATION→ORGANIZATION correction, award-title exclusion, demographic-JOB exclusion), validated for value/offset and non-overlap invariants. Rules requiring sentence-level semantic judgment were checked via a manual stratified spot-check (prioritizing JOB and WORKOFART, the two labels most reliant on pattern-matching rather than a trained model) rather than blind trust in automated output.

Balance: target floor of 200 / ceiling of 600 total spans per label (Stage 1 + Stage 2 combined), to prevent easy-to-find labels (PERSON, ORGANIZATION, LOCATION, TIMEDATE) from dominating while forcing deliberate sourcing of rarer labels (JOB, WORKOFART). Record selection prioritized records containing the scarcest labels first, with a reserved quota so WikiANN's PERSON/ORGANIZATION/LOCATION coverage wasn't crowded out by richer-but-rarer Wikipedia/Wikinews records.

PRODUCT floor not met (documented limitation): three approaches to PRODUCT candidate generation were tried — spaCy's native tag (~0% precision), broader source categories (didn't fix a label-mapping problem), and trigger-word patterns (found only a handful of genuine hits, since biographical/brand-article prose rarely phrases things as "launched the X"). PRODUCT landed at 88 total spans, well under the 200 floor. This is a known weak point flagged for Stage 3/4 QA to watch for.

Dedupe: exact-match plus substring dedupe against Stage 1's 98 records and across the new candidate pool (catches truncated/split duplicate sentences from overlapping source text, without the false-positive risk of a fuzzy-similarity threshold).

Split: stratified 80/20 train/test by each record's dominant (most frequent) label, so both splits get proportional coverage of every label including rare ones — actual split landed at 598/150 (79.9%/20.1%).

Corrections summary (new records only; Stage 1 corrections unchanged)

• 650 new records added (98 Stage 1 + 650 = 748 total): 200 from WikiANN, 417 from Wikipedia, 33 from Wikinews.
• All new spans passed automated policy correction and structural validation (0 value-mismatch errors, 0 overlap errors, 0 duplicate texts, 0 train/test text overlap across all 748 records).
• Source datasets: unimelb-nlp/wikiann (English), Wikipedia, Wikinews.