Ali-0111/applied-nlp-ner-stage2-expanded

Applied NLP / NER — Stage 2: Dataset Expansion

Stage 1's 98 corrected records (from Ali-0111/applied-nlp-ner-stage1-cleaned) plus 661 new records, all labeled under the full policy below, split into train/test.

Sources for the new records

Text only was reused from each source below -- every source's own gold NER labels (if any) were discarded, and every record was relabeled from scratch under this project's own 8-label policy. Original tags were used only as a filter to find entity-bearing text, and for Few-NERD's fine-grained art-*/product-* tags, to route text toward the WORKOFART/PRODUCT backfill.

• tomaarsen/conll2003 (Reuters news, mirrors the classic CoNLL-2003 NER benchmark) -- research/redistribution-permissive, widely reused in NLP research.
• wikiann "en" (Wikipedia) -- CC-BY-SA.
• DFKI-SLT/few-nerd "supervised" (Wikipedia) -- CC-BY-SA. Fine-grained art-*/product-* tags used specifically to backfill the WORKOFART/PRODUCT labels, which were thin after Stage 1.
• wikimedia/wikipedia "20231101.en" (official Wikimedia dump) -- CC-BY-SA + GFDL. Used via a regex heuristic over biographical-article lead paragraphs to backfill the JOB label (Wikipedia bio ledes reliably state the subject's occupation).
• sedthh/gutenberg_english (Project Gutenberg) -- public domain. A small sample for literary-register diversity.

WNUT-17 (Twitter) was originally planned for social-media register diversity but is no longer loadable from the Hub (its dataset script was deprecated with no maintained parquet mirror found) -- dropped; noted here for transparency.

Stage 2 summary

• Stage 1 corrected records carried forward: 98
• New records added: 661
• Total: 759 (644 train / 115 test, stratified 85/15 by source and by whether a record contains a thin label)
• Total entities: 5733

Label counts (combined Stage 1 + Stage 2):

Label	Count
AMOUNT	1237
TIMEDATE	970
ORGANIZATION	828
PERSON	805
LOCATION	694
JOB	650
WORKOFART	337
PRODUCT	212

Records by source:

Source	Count
tomaarsen/conll2003	220
DFKI-SLT/few-nerd	179
wikiann/en	150
wikimedia/wikipedia	100
stage1_starter_corrected	98
sedthh/gutenberg_english	12

Balance criteria (decision D4)

WORKOFART and JOB were the thinnest labels after Stage 1 (52 and 91 respectively out of ~80-160 for the others). Rather than uniformly resampling everything, new sourcing deliberately targeted these two: Few-NERD's art-*/product-* fine-grained tags for WORKOFART/PRODUCT, and a Wikipedia biographical-lede heuristic for JOB. Everything else was left close to its natural distribution rather than artificially rebalanced, on the reasoning that real-world text really is TIMEDATE/PERSON-heavy and an artificially uniform distribution would be less representative of what the model sees at inference time.

Labeling method (decision D3)

Records were labeled by LLM agents working from the full written policy below (not from any source dataset's original gold labels), each computing character offsets programmatically (text.find + exact-match assertion) rather than by hand-counting, and cross-checking precedent against already-labeled batches for consistency. 100% of records were run through an automated validator (span-matches-text, valid-label-only, no-overlap checks) with zero errors. A stratified ~11% manual sample (75 of 661 new records, spread across every source and backfill category) was read in full by the project owner; the sample surfaced only minor, low-frequency issues (a couple of tennis-score labeling inconsistencies, occasional record-boundary truncation from the source passage-assembly step, one Rule-A1 edge case) documented in the project report rather than requiring a full re-label -- consistent with the brief's own expectation that full coverage isn't required.

Deduplication and split

Exact-text dedup was run within the new pool and against Stage 1's 98 records before splitting (0 duplicates found -- prevented earlier at the sourcing stage). The train/test split is stratified by source dataset and by whether a record contains a WORKOFART or JOB entity, at roughly 85/15, so both splits see every register and the thin labels rather than risking them landing mostly in one split by chance.

Full annotation policy in force for this dataset

NER Labeling Policy

Version: v1 (Stage 1 — Rulecraft and Cleanup) Owner: Ali Agabalayev

This is the full annotation policy in force for the Stage 1 corrected dataset. It consists of the fixed Labels and Baseline labeling rules (verbatim from the task brief, not editable) plus the extra rules added below to close gaps the baseline does not settle. Every later stage that edits this policy must append a new dated section rather than silently changing an earlier one.

Labels

• PERSON: a named person, including given names and surnames.
• ORGANIZATION: a named company or institution.
• LOCATION: a named place such as a city, country, region, street, landmark, or geographic area.
• TIMEDATE: an expression that places something on a timeline, including dates, clock times, and durations used as time.
• PRODUCT: a named commercial product, device, or branded good.
• WORKOFART: a named creative or published work such as a book, film, song, or titled publication.
• JOB: an occupational title or formal work role when it functions as such in the sentence.
• AMOUNT: a measurable or countable quantity, not a time or date expression.

Baseline labeling rules (fixed, from the task brief)

1. A labeled mention should cover the entity itself, not the surrounding grammar.
2. Articles, prepositions, conjunctions, and other function words should stay outside the span unless they are truly part of the proper name (e.g. The New York Times keeps The).
3. Bare type-words and category descriptors ("person", "company", "city", "product", etc.) should not be labeled on their own.
4. A multi-word name should be one span when it forms one real named entity; separate entities should be labeled separately.
5. Coordinated names should be separate spans, unless the conjunction is part of one established name (Johnson & Johnson stays one span).
6. Ordinary punctuation and stray whitespace stay outside spans, unless the punctuation belongs to the name or abbreviation (AT&T keeps &).
7. Possessive markers stay outside the span unless the full possessive form is the name.
8. Labels follow context, not surface form alone (Cambridge University vs Cambridge the place).
9. Quantities used as time are TIMEDATE, not AMOUNT.

Added policy rules (v1 — Stage 1)

Rule A1 — Hedge/quantifier modifiers stay inside TIMEDATE and AMOUNT spans; plain articles do not

Gap closed: the baseline's article rule ("articles stay outside unless part of the proper name") is written for proper names (ORGANIZATION, PERSON, LOCATION). It does not say what to do with common-noun TIMEDATE/AMOUNT phrases like a few days, an hour or so, about 50 tickets — these aren't proper names at all, so the baseline rule doesn't directly resolve the boundary.

Rule: a small set of hedge/quantifier words — about, roughly, nearly, approximately, around, a few, several — stay inside the span when they modify a TIMEDATE or AMOUNT value, because they change what the value actually denotes (about 50 and 50 are different quantities). Plain grammatical articles (a, an, the) with no hedging function stay outside the span, same as the baseline treats them elsewhere.

Examples:

• about 50 tickets → <AMOUNT>about 50</AMOUNT> tickets (kept: hedge changes the value)
• an hour or so → <TIMEDATE>hour or so</TIMEDATE> (dropped: an is a plain article, or so is part of the vague-duration phrase itself, not a leading article)
• a few days → <TIMEDATE>few days</TIMEDATE> (dropped: a is a plain article)

Consistency check: does not contradict baseline rule 2 (still keeps function words outside by default) or rule 9 (still separates time-quantities from AMOUNT); only adds the hedge-word carve-out the baseline leaves open.

Rule A2 — Reference/identifier numbers are not AMOUNT

Gap closed: the baseline defines AMOUNT as "a measurable or countable quantity." It does not say whether numbers that function as identifiers rather than quantities (legal citation volume/page numbers, ID numbers, serial/model numbers, case numbers) count. A citation page number like 358 in 362 U.S. 352, 358 is not something measured or counted — it's a reference locator.

Rule: numeric tokens that function as reference identifiers (citation volume/page numbers, case numbers, ID/serial numbers) are not labeled AMOUNT (or any other label) — they are left unlabeled. This is judged by function in context: if the number answers "how many / how much," it's AMOUNT; if it answers "which one" (a locator/identifier), it is left unlabeled.

Examples:

• M62 U.S. 352, 358 → 352 and 358 left unlabeled (citation locators, not quantities)
• She bought 50 tickets → <AMOUNT>50</AMOUNT> (a true countable quantity, labeled as before)

Consistency check: narrows AMOUNT without touching its core definition or any other rule; does not conflict with Rule A1 (A1 is about modifier inclusion for values that are already legitimately AMOUNT/TIMEDATE; A2 decides whether a number is AMOUNT-eligible in the first place).

Data-quality exclusion criterion (not a labeling rule — a sample-removal criterion)

Two starter records were removed rather than corrected, per the brief's allowance to remove samples that would worsen model behavior, with justification:

• idx 61: retained originally as a candidate "noisy but real" legal-document record; on review, the density of OCR-style corruption (Silegation, nenorendum, Gctober ZL2002, recorá, etc.) makes several spans unverifiable against a confident human reading of the intended words. Removed.
• idx 80: not coherent English prose — a keyword/search-log style block of concatenated fragments. Its existing annotations are actively wrong in ways a model would learn from (military labeled JOB, php/washingtonpost labeled WORKOFART). Removed.

This criterion — source text whose word-level integrity is too degraded to annotate with confidence, or text that is not coherent natural-language prose — will be applied the same way during Stage 2 sourcing.

Record-level corrections requiring judgment, not a new rule (enforcing existing rules/definitions)

These are individual mislabeled entities fixed by applying the existing baseline definitions — no new rule was needed, the baseline already settles them once the concrete error is spotted by a human/QA pass:

• COMPANY is not one of the 8 labels; each of its 8 occurrences was re-examined individually rather than blind-remapped, since most turned out not to be companies at all:
  • idx 16 Guided Advanced Tactical Rocket → PRODUCT (a device, not a company)
  • idx 39 Remembrance Day → TIMEDATE (a named observance/date; closest fit of the 8 labels)
  • idx 47 Schamberger - O'Connell → ORGANIZATION (an actual firm name — the one case the default remap fit)
  • idx 79 2012 → TIMEDATE (a year, not a company)
  • idx 48 Dan → merged with the adjacent, separately-mislabeled Riley and Sanford spans into one PERSON span Dan Riley Sanford (baseline rule 4: one real named entity, one span)
  • idx 69 Ronald → merged with adjacent McDonald into one PERSON span Ronald McDonald (same rule, appears twice in the record)
  • idx 86 Kim + Jong-Un (two separate COMPANY spans) → merged into one PERSON span Kim Jong-Un (same rule)
• idx 69 also had the actual McDonald's-the-company mentions inconsistently labeled across the same document: two as ORGANIZATION, two as LOCATION, with no context shift justifying the split (every mention refers to the same company/brand, not a place) — normalized to ORGANIZATION throughout. The possessive 's was correctly kept in all four spans: per baseline rule 7, McDonald's is a case where "the full possessive form is the name," not a case where grammatical possession should be stripped. The same applies to idx 79 Andre's (a restaurant name), which was also relabeled LOCATION → ORGANIZATION (it denotes the business, not a place) with the possessive kept.
• idx 28, idx 49: zero-entity records were missed-entity cases, not true negatives — nintendo (idx 28) and WebMD (idx 49) are named organizations that were left unlabeled. Added as ORGANIZATION.
• idx 2: WORKOFART span of Bandits: From Greed to Generosity was missing its leading word — widened to the full title Vindication of Bandits: From Greed to Generosity.
• idx 27: The Executioner Bug Zapper Racket relabeled WORKOFART → PRODUCT (a physical device, not a creative/published work); The kept as it's part of the branded product name, same exception baseline rule 2 already carves out for established names.
• idx 43: AMOUNT span of both parental verification and consent removed outright — not a quantity under any reading, and no other label fits either; this was a bad span, not a boundary or label variant of a real entity.
• Bare type-words removed (baseline rule 3), beyond the ones already caught in the initial scan: idx 48 school, idx 69 country., idx 88 movie. — all generic category nouns, not named entities.
• Numeric mislabels found via a cross-record consistency check (same surface number labeled both AMOUNT and TIMEDATE in different records) — three were genuine errors, not legitimate context-dependence, and were relabeled TIMEDATE → AMOUNT per baseline rule 9 (a quantity only becomes TIMEDATE when it's used as time): idx 37 five ("the five pillars of Islam" — a count), idx 63 3 ("hearing about it 3 times" — a frequency count, not a time marker), idx 89 150 ("150 period-authentic costumes" — a count).
• Rule A1 applied: idx 9, 10 (×2), 85 (×2) TIMEDATE spans had their leading plain article dropped (a few days → few days, an hour or so → hour or so, etc.); idx 60 AMOUNT span 150 was widened to about $150 since about is a hedge word that changes the value under Rule A1.