Hasanblanka/polygraf-ner-stage_2-dataset

Polygraf NER Stage 2 Expanded

This repository contains the final result of Stage 2 — Dataset Expansion for the Polygraf Applied NLP / NER technical task.

Dataset structure

Each row follows the original three-field character-span schema:

{
  "unique_index": 100,
  "source_text": "Example English text.",
  "privacy_mask": [
    {"start": 0, "end": 7, "label": "PERSON", "value": "Example"}
  ]
}

source_text[start:end] exactly equals value for every span. Spans are flat, sorted, non-overlapping, and use exclusive end offsets.

Files:

• data/train.jsonl: 506-record training split.
• data/test.jsonl: 127-record test split.
• data/stage2_expanded_all.jsonl: all 633 records in one convenience file.

The default Hugging Face configuration loads only train.jsonl and test.jsonl; the combined file is provided for direct download and must not be loaded alongside the splits as an additional training split.

Sources and licensing

The dataset combines:

1. 90 Stage 1 seed records retained from polygraf-ai/applied-nlp-ner-candidate-starter-100. These remain subject to the evaluation-use-only terms in LICENSE.
2. 543 new Stage 2 records selected from the Open Newswire English feed restricted to cc-by and cc-by-sa sources: https://feed.opennewswire.org/?licenses=cc-by,cc-by-sa&languages=en.

The applicable publisher, article URL, license identifier, license statement, and source hash for every Stage 2 row are in source_attribution.jsonl. See DATASET_LICENSES.md. Because this is a mixed-license dataset, the repository uses license: other; users must follow the terms attached to each row.

Selection and quality control

• 829 Open Newswire records were collected.
• 719 passed structural, language, and text-quality checks.
• 18 exact or near duplicates were removed.
• Publisher balancing capped overrepresented sources.
• 550 candidates were retained for annotation.
• Initial spans combined a small CoNLL-2003 DistilBERT candidate generator, deterministic policy rules, and a Stage 1 gazetteer.
• All 550 candidates received manual semantic QA.
• 7 low-quality or duplicate records were removed.
• 543 new records remained and were merged with the 90 Stage 1 seed records.

The main removal reasons before annotation were publisher cap, excessive length, duplicate URL, very short text, weak English signal, and near duplication. DistilBERT was used only for PERSON, ORGANIZATION, and LOCATION candidate generation; it was not treated as final ground truth.

Split

The final 633 records were split deterministically with seed=42 and a 20% test ratio:

Split	Records	Stage 1	Stage 2	Spans
train	506	72	434	2,978
test	127	18	109	786
total	633	90	543	3,764

Normalized exact-text duplicates were absent. Six high-similarity pairs, forming one four-record template group, were kept in the same split. Therefore no detected exact or near duplicate crosses train and test.

Label statistics

Label	Train	Test	Total
AMOUNT	285	72	357
JOB	300	71	371
LOCATION	589	124	713
ORGANIZATION	389	121	510
PERSON	336	87	423
PRODUCT	90	41	131
TIMEDATE	461	143	604
WORKOFART	528	127	655

Labels

• PERSON: a named person, including given names and surnames.
• ORGANIZATION: a named company or institution.
• LOCATION: a named city, country, region, street, landmark, building, or geographic area.
• TIMEDATE: dates, clock times, ages, frequencies, and durations used as time.
• PRODUCT: a named commercial product, device, software item, or branded good.
• WORKOFART: a named creative or published work, including books, films, songs, articles, reports, and titled publications.
• JOB: an occupational title or formal work role in context.
• AMOUNT: a measurable or countable quantity that is not time or date.

Fixed baseline labeling rules

1. A span covers the entity itself, not surrounding grammar.
2. Articles, prepositions, conjunctions, and function words stay outside unless they are genuinely part of the proper name.
3. Bare type words and category descriptors are not labeled by themselves.
4. A multi-word real name is one span; separate entities are separate spans.
5. Coordinated names are separate unless the conjunction belongs to one established name.
6. Ordinary punctuation and stray whitespace stay outside unless they belong to the name or abbreviation.
7. Possessive markers stay outside unless the possessive form is the name.
8. Labels follow context, not surface form alone.
9. Quantities used as time are TIMEDATE, not AMOUNT.

Additional annotation policy

A1. Temporal specificity, ages, and vague adverbs

Calendar dates, clock times, seasons, durations, frequencies, ages, and relative expressions with a concrete timeline reference are TIMEDATE. General adverbs such as ever, never, always, again, yet, just, initially, and already are excluded when they do not identify a time point or interval. Question placeholders such as when and how long are excluded.

A2. Quantity core and unit boundaries

AMOUNT includes the numeric/count core, ranges, and scale words such as 3.5, 3-4, 1.2 million, half, both, and thousands. Currency symbols/codes, measurement units, counted nouns, and approximators such as about, over, and at least stay outside. Numeric identifiers are not amounts. A time unit is included in a TIMEDATE duration or age because it defines the temporal class.

A3. Person names, nicknames, and usernames

Adjacent given and family names referring to one person form one PERSON span. Titles and possessive markers stay outside. A nickname inside a full name may stay in the same span with its quotation marks. Usernames and standalone nicknames are PERSON only when context clearly identifies a real person. Structural dialogue placeholders are not labeled.

A4. Product, organization, and published work

Named software, games, platforms, medicines, fonts, devices, and branded goods are PRODUCT when the artifact or service is meant. The company operating or producing it is ORGANIZATION. Films, songs, books, articles, reports, and other publication titles are WORKOFART.

A5. Jobs versus departments and activities

JOB applies only to an occupation or formal work role in context. Departments, teams, sectors, fields, activities, and generic organizational functions are not jobs. A specifically named department may be ORGANIZATION.

A6. Locations, establishments, and companies

Companies, sports teams, restaurants, and institution names are ORGANIZATION even when someone visits them. Geographic areas, streets, addresses, landmarks, and named physical objects are LOCATION when physical place is emphasized. The same surface form may receive different labels in different contexts.

A7. No nested labels inside work titles

The dataset uses flat, non-overlapping spans. When another entity appears inside a complete titled work, the whole title is one WORKOFART span and no nested span is added.

A8. Proposed names

A phrase explicitly proposed as a name is labeled according to the intended entity type even if the object does not yet exist.

A9. Record-removal quality threshold

A record is removed only when severe OCR/encoding corruption, unrelated generated noise, metadata-only content, harmful template duplication, or truncation makes reliable annotation impossible. Ordinary spelling and grammar variation is retained as natural language diversity.

The standalone annotation_policy.md provides the same policy with examples and detailed justifications.

Limitations

• The dataset is small and source distribution is not uniform.
• Open Newswire records are news-heavy and can contain publisher-specific writing templates.
• Annotation follows a project-specific flat span policy and may differ from common NER benchmarks.
• Initial automation had limited native label coverage; manual QA corrected the final spans but cannot guarantee the absence of all semantic judgment errors.