--- pretty_name: Applied NER Stage 4 - Diagnosed and Improved language: - en task_categories: - token-classification tags: - named-entity-recognition - ner license: other --- # Applied NER Stage 4 — Diagnosed and Improved Stage 2 after targeted review and augmentation against ten error patterns observed on fresh Stage 3 probes. The 159-record test split is unchanged. ## Dataset structure - Repository ID: THemidli/applied-ner-stage4-improved - Splits: train (841), test (159) - Total records: 1,000 - Total spans: 5,541 - Test rows changed: 0 - Wild-probe sentence overlap: 0 | Label | Spans | |---|---:| | AMOUNT | 539 | | JOB | 451 | | LOCATION | 749 | | ORGANIZATION | 735 | | PERSON | 1,061 | | PRODUCT | 636 | | TIMEDATE | 830 | | WORKOFART | 540 | ## Changes from Stage 2 - Reviewed 80 existing training records with a Hermes proposal plus independent Qwen exact-consensus gate. - Manually adjudicated all 46 proposed changes: 15 accepted, 20 reverted, and 11 replaced by custom corrections. In total, 26 existing records changed. - Added 200 exact-offset-validated training records: 20 for each of the ten Stage 3 error families. - Removed 0 records and changed 0 test records. The policy itself did not change in Stage 4. The intervention targets coverage and consistency under the same policy. ## Source and license notes The 800 inherited records retain Stage 2 terms. The 200 candidate-authored template records are CC BY-SA 4.0. Metadata remains license: other because the inherited starter portion has no stated general-purpose license. A 1,000-row ledger records provenance. ## Labels - **PERSON** — A named person, including given names and surnames. - **ORGANIZATION** — A named company or institution. - **LOCATION** — A named place, including 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 context. - **AMOUNT** — A measurable or countable quantity that is not a time or date expression. ## Baseline labeling rules (fixed) - A labeled mention covers the entity itself, not surrounding grammar. - Function words stay outside unless they are truly part of the proper name. - Bare type words and category descriptors are not labeled on their own. - A multi-word name is one span; separate entities are labeled separately. - Coordinated names are separate spans unless the conjunction belongs to an established name. - Ordinary punctuation and stray whitespace stay outside unless they belong to the name. - Possessive markers stay outside unless the full possessive form is the name. - Labels follow context, not surface form alone. - Quantities used as time are TIMEDATE, not AMOUNT. ## Added policy in force ### P1. Namedness over capitalization Label only a name, title, or contextually specific referent. Capitalization alone is not evidence of an entity, and lower-case occupational titles can still be JOB mentions. Example: Research sector is not an organization; entrepreneur is JOB when it names the role. Why: Prevents capitalized common nouns from becoming false positives while preserving contextual roles. ### P2. Aliases, abbreviations, and honorifics Label a conventional alias, initials, or acronym when it independently identifies the same entity. Exclude honorifics and courtesy titles from PERSON unless they are inseparable from the established name. Example: Manny Pacquiao and Pac-Man; Dr. stays outside Sarah Chen. Why: Makes short-form mentions reusable without absorbing grammatical titles into names. ### P3. Numeric core and range boundaries For AMOUNT, include the numeric core, an attached sign or currency symbol, an attached magnitude suffix, and a currency or measurement unit when that unit is integral to the scalar value. Include a hyphenated numeric range as one span. Exclude approximation words and counted category nouns. A word such as meters is inside for physical length but outside when it means a count of utility devices. Example: about $300m; 3 feet 1 inches; 20,000 meters were installed; 50 tickets. Why: Retains value-changing units while keeping approximation and counted categories out. ### P4. Time expressions and vague temporal language TIMEDATE includes explicit dates, clock times, anchored relative dates, ages, recurrences, and durations with their time unit. A continuous interval may include its internal range connector when the phrase denotes one interval. Do not label vague words such as moment, eventually, recent, or long term unless the phrase provides a concrete duration, age, recurrence, or calendar/clock anchor. Example: May 2, 18 months, and 36-year-old, but moment is left unlabelled. Why: Separates timeline-bearing expressions from general temporal discourse. ### P5. Job-title extent A JOB span contains the complete role noun phrase, including seniority or specialization when these change the role, but excludes determiners, employer names, person names, and incidental adjectives. Example: an International Solutions Analyst at Harber Inc. Why: Avoids both bare-head under-labeling and sentence-fragment over-labeling. ### P6. Works, reports, and candidate names Label the full presented title of a creative or published work, including a titled report or article. A title and subtitle joined by a colon form one span, including the colon. Label a proposed name when the text explicitly presents it as the name of a work or commercial product. Generic references such as the report, the song, or the product remain outside. Example: Project Performance Audit Report and proposed PawTrak. Why: Handles titles and naming lists consistently without promoting generic document/product nouns. ### P7. Institution versus physical place Use ORGANIZATION when a named institution or facility acts, employs, publishes, treats, or otherwise behaves as an institution. Use LOCATION when the same name denotes the physical site or destination. Example: She works at Cambridge University; the event is in Cambridge. Why: Makes the baseline context rule operational for schools, hospitals, banks, and venues. ### P8. No nested or overlapping spans Do not emit nested or overlapping entities. Select the longest contextually correct entity; annotate a short form separately only when it appears as a separate surface mention. Example: Use one New York University span, not LOCATION New York inside it. Why: The flat token-classification target cannot represent nested spans without contradictory labels. ### P9. Demonyms and adjectival place references Do not label nationalities, demonyms, or adjectival place modifiers as LOCATION unless the words denote an actual place in context. Example: American researchers is unlabelled; researchers in America is labelled. Why: Prevents attributes of people or products from being mistaken for locations. ### P10. Every surface mention is evaluated Evaluate each textual occurrence independently. Repeated named mentions are all labelled, while a later pronoun or generic description is not copied into the entity set. Example: Google ... Google produces two ORGANIZATION spans; the company is not a third span. Why: Prevents both missing repeated mentions and coreference-driven false spans. ### P11. Named software and technical artifacts Treat named software, games, programming languages, platforms, and technical systems as PRODUCT even when they are open source. Generic technology categories, file types, and programming terms remain outside. Example: Minecraft and Java, but plugin and programming language stay outside. Why: The label inventory has no software class; PRODUCT is the closest stable target for named technical artifacts. ### P12. Handles and anonymized person identifiers A username, stage name, explicit nickname, or anonymized placeholder is PERSON when it identifies a particular person in context. Generic family terms, occupations, and descriptive epithets are not PERSON. Example: Person1 and Pac-Man, but mum and the guitarist are outside. Why: Preserves person identity in conversational and privacy-oriented text without labeling generic descriptions. ## Limitations The 200 additions are controlled templates and less diverse than natural text. They do not overlap the fixed test or wild sets, but can still encourage pattern-specific overfitting.