Hist-Pile-NER-Type

A multilingual historical-newspaper silver annotation dataset by Emanuela Boros, prepared for future HistGliNER development. This dataset combines the passage-to-conversation construction approach of UniversalNER / Pile-NER-type with an operational NERC adaptation of the Impresso / HIPE-2020 annotation guidelines v2.2.0.

This is not human-reviewed gold data, an official HIPE release, or an exact reproduction of the original open-type Pile-NER-type dataset. No HistGliNER model is included.

Contents

Configuration	Rows	Contents
default	1667	Fine literal-type conversations, with id and conversations
coarse	1667	Coarse literal-type conversations
hipe_iob	1606	Token arrays with HIPE-style BIO labels and automatic sentence boundaries
source_metadata	2271	Recovered page/issue metadata, one row per annotated passage, joined by id
spans	2271	Original OCR passages, exact character spans, fine/coarse types, nesting flags, metonymic types, ambiguity/OCR flags and provenance

The canonical annotations contain 12002 entity occurrences, from 604 issues and 165 newspaper title strings. Reported source dates range from 1710-12-01 to 1942-03-05; dates are unverified metadata.

65 passages were flagged unreadable by the teacher; 534 contain type ambiguity. Such passages remain in spans for audit but are excluded from conversational training exports. 1831 passages have model-written review notes. A further 5 assistant-corrected records are marked review_required and excluded from training conversations pending human review. Human-reviewed records: 0. Automatic span validity does not establish annotation correctness or completeness.

Source and sampling

Source: biglam/europeana_newspapers, revision d27f7fa9092e255832ca1d496574967add680f7d.

The original 1,000-passage pilot uses seeded language-quota cluster sampling of page windows through Hugging Face's row service, verifying its x-revision header. For larger language subsets, 20 random windows of at most 20 rows were drawn (30 for Swedish); the small Yiddish and Ukrainian subsets were retrieved in full. Candidate pages were shuffled and one contiguous passage per selected page was chosen. Truncated responses, very short passages and exact normalized passage duplicates were excluded.

The budget-limited expansion requested 9,000 additional annotated passages. It prepared 9,000 candidates and currently contributes 1,271 successfully annotated passages; 7,729 candidates remain unannotated or failed validation and are not included as labeled data. The original pilot files are retained locally, and the pre-expansion published revision is 85414a284311d38c54ba30b305e0339ddf156c0f.

Expansion candidates use the same source revision and nine times the pilot's language quotas, with up to eight nonoverlapping passages per page including any pilot passage. Cached source pages are supplemented by additional seeded row windows as needed. Exact normalized duplicates, same-page overlaps, and shared contiguous 13-word sequences with the pilot, other expansion passages, or the pinned HIPE-2020 French/English/German test sets are excluded. The latter uses case-folded Unicode word sequences and is not a complete fuzzy OCR contamination check. Prepared candidates are deterministically shuffled before budget-capped annotation; the labeled subset is therefore not guaranteed to meet the full target quotas.

Passages have 64–256 cl100k_base tokens, preserving exact OCR character substrings. These are language quotas, not population weights or balanced historical-period coverage. This approach is not the original paper's uniform 50K-passage sampling. No OCR-quality threshold was applied. No text translation, OCR correction or image-based annotation was performed. Publication dates, language labels and OCR confidence are supplied by the source and are not independently verified. ji is retained as the source code for Yiddish; the Hub language tag is yi.

Source language code	Sampled passages
de	227
el	216
et	234
fi	228
fr	225
hr	137
ji	54
pl	217
ru	240
sr	252
sv	206
uk	35

Recovered source metadata

Load source_metadata and join its id to any annotated view:

from datasets import load_dataset
metadata = load_dataset("emanuelaboros/Hist-Pile-NER-Type", "source_metadata", split="train")
by_id = {row["id"]: row for row in metadata}

This view recovers newspaper title, source-supplied publication date, languages, issue/page IDs, page-index suffix, IIIF image URL, OCR confidence, illustration boxes, character offsets and source hashes. The original source fields also remain in spans and hipe_iob. Metadata and exact passage substrings are matched against the cached source rows used by the revision-checked sampler; no annotation model is called.

These are page/issue metadata, not segmented article descriptions. newspaper_title is not an article headline. Byline, article title/boundaries, publisher and publication location are unavailable in the source row schema and are not inferred. Dates remain historically unverified; page index is not a verified printed page number. Illustration boxes describe images, not article/entity regions. OCR scores describe full source pages. Preserved links have not been independently checked for availability, and original rights status is unchanged.

See source_metadata/README.md for field definitions and source_metadata/statistics.json for coverage. A JSONL copy is provided. Only annotated passages are published; the metadata manifest for all prepared candidates is retained locally.

Annotation

Teacher: gpt-5-mini-2025-08-07, low reasoning effort. Temperature was not set because this reasoning model does not support the original recipe's temperature-zero configuration. The prompt and label mapping are included as annotation_prompt.txt and labels.json; the prompt SHA-256 is ff5b624f9a2159777bb285ae5e6854ec09faa711261499a069a7164b5dab8383.

HIPE-style full mention boundaries include attached titles, functions and qualifying phrases. Fine literal types map to the five coarse families: person, organization, location, human production and absolute date. Nested entities, literal/metonymic distinctions and unresolved type alternatives are preserved in the span configuration. Separate component labeling, OCR correction hypotheses and Wikidata linking are outside this pilot. The components field is an empty reserved field, not an assertion that mentions have no components. Guideline application beyond HIPE's original languages is an experimental extension.

The model quotes original strings and identifies repeated occurrences. Code derives zero-based, end-exclusive Unicode character offsets and rejects invalid alignments. Inconsistent ambiguity metadata is reconciled conservatively: either an ambiguity flag or supplied alternative types keeps the mention ambiguous. Such records are excluded from the conversation views, and recovered records carry a review note. No uncertain type is silently resolved by this step. Five final copy failures received explicit Codex assistant corrections recorded in review_repairs.py in the private code repository; these remain provisional and excluded from training views. Initial requests supplied JSON-encoded text; final retries used delimited plain text to reduce quote-escaping errors, recorded in annotation_provenance.input_format. Failed records are retried with validation feedback; raw responses and a cumulative cost ledger are retained locally. Small preliminary GPT-4.1-mini and GPT-5-mini calibration batches were discarded and are excluded from released annotations.

Pile-NER-type format

Each conversation starts with the source passage and an acknowledgement, then asks What describes <type> in the text?. Answers are JSON lists of entity surface strings. Repeated identical strings are collapsed in conversational answers; all occurrences remain in spans. Questions use readable English type labels while the mentions retain their source language.

Up to three negative labels per passage are sampled without replacement, weighted by entity occurrence frequency in the eligible dataset, excluding labels annotated as present. Broad person/organization labels are also excluded as negatives when an annotated subtype in that family is present. Their answers are []. These labels are absent from machine annotations, not independently verified absent from the text. Missed entities can therefore introduce false negatives. Conversation exports use literal types; metonymic annotations remain in spans and are not mixed into the same query semantics.

At the user's request, HIPE supplies a fixed fine-type inventory. This differs from UniversalNER's unrestricted type discovery and should not be described as equivalent open-type supervision.

Loading

from datasets import load_dataset
conversations = load_dataset("emanuelaboros/Hist-Pile-NER-Type", split="train")
spans = load_dataset("emanuelaboros/Hist-Pile-NER-Type", "spans", split="train")

HIPE-2020 IOB export

The hipe_iob configuration contains 1606 passages, 183,825 tokens, 7,071 outer literal mentions, 291 nested mentions, and 17 metonymic interpretations, in 12 languages. Download one TSV per language, or load the token arrays:

hipe = load_dataset("emanuelaboros/Hist-Pile-NER-Type", "hipe_iob", split="train")

The TSV files follow the ten-column layout of CLEF-HIPE-2020 v1.4: TOKEN, NE-COARSE-LIT, NE-COARSE-METO, NE-FINE-LIT, NE-FINE-METO, NE-FINE-COMP, NE-NESTED, NEL-LIT, NEL-METO, MISC. Tagging is BIO/IOB2: every mention starts with B-, subsequent tokens use I-, and tokens outside an annotated mention use O. Like the original v1.4 files, NE-NESTED carries fine literal types. Outermost entities occupy the main literal columns; one non-overlapping level of inner entities occupies NE-NESTED.

Component annotation and entity linking were not performed: NE-FINE-COMP, NEL-LIT, and NEL-METO are _ throughout. _ means unannotated; it is not O or an asserted NIL link. Existing metonymic types populate their own BIO columns. MISC uses NoSpaceAfter for adjacent tokens and EndOfLine only for a newline present in the source passage. EndOfSentence marks automatic punctuation-based sentence boundaries and the last token of every passage (including a potentially truncated final fragment). The deterministic punctuation_entity_safe_v1 heuristic handles punctuation runs and adjacent closing quotes/brackets; guards common abbreviations, initials, numeric periods, decimals and lowercase continuations; treats Greek semicolons as question marks only for Greek; and suppresses boundaries inside annotated entity spans. Newlines alone do not end sentences. These boundaries are heuristic, not human-reviewed, and may undersegment abbreviations or miss punctuation damaged by OCR. The export contains 8,928 boundary markers, including 1606 passage-final markers. EndOfSentence follows the HIPE format MISC convention. Unavailable page-layout boundaries are not inferred. Each passage is a separate document, identified by comment metadata and separated by a blank line.

This is a HIPE-style silver derivative, not official HIPE data or a replication of its tokenizer. Tokens use the deterministic Unicode regex \w+(?:[-_]\w+)*|\S, also used by this project's GLiNER export. Token text is never corrected. The Parquet view retains original text, source metadata, text hash, page offsets, and zero-based, end-exclusive Unicode token_start/token_end arrays so exact original whitespace and character spans remain recoverable. Language identifiers retain source conventions, including ji for Yiddish.

Of 2,271 annotated source passages, 665 are excluded: 534 type ambiguous, 65 unreadable, 53 entity cuts token, 6 nesting deeper than two, 5 review required, 2 crossing entities. The complete rejected passage IDs and first rejection reasons are in hipe_iob/rejected.json; measured counts are in hipe_iob/statistics.json. Entire passages are excluded rather than silently dropping mentions or changing their boundaries. All accepted literal, nested, and metonymic tags were decoded and checked against the original character spans. The original spans configuration still retains all 2,271 annotated passages. This export requires no additional model calls and remains train-only, with zero human-reviewed passages.

Evaluation and limitations

All configurations are train-only. There is no gold validation or test set. Before HistGliNER evaluation, create a separate expert-reviewed set and group issues and near-duplicate/reprinted material before splitting. The original pilot removes exact normalized passage duplicates only. The expansion adds the overlap filters described above; neither stage rules out all near-duplicates or benchmark contamination. Do not treat its automated labels as ground truth for performance claims.

Historical typography, damaged OCR, source metadata errors, multilingual text and uneven language coverage can cause missed mentions, wrong boundaries and wrong types. In particular, low-resource language annotations require native-speaker review. Review notes and OCR flags are model judgments, not independent measurements. Historical text may contain discriminatory language.

The recorded before/after experiment at the pre-expansion revision found lower HIPE-2020 F1 after fine-tuning the historical baseline on the pilot. Scaling the silver annotations is not evidence that they improve downstream performance. The expansion has not yet been evaluated in that experiment.

Rights

The inspected BigLAM source card does not state a blanket text redistribution license. No blanket public-domain or unrestricted-use claim is made here, and no new license is asserted over source text. Source IDs, issue URIs and image links are preserved to support item-level provenance and rights checks. Consult the originating collections' terms before reuse. The Pile-NER-type or HIPE corpus licenses are not automatically inherited: neither annotated corpus is copied here.

Code availability

Implementation code is maintained separately from this dataset. The source is in EmanuelaBoros/hist-pile-ner-type, a private GitHub repository requiring access. The annotation prompt and label definitions remain available here as dataset metadata.

Construction record

The measured token-price estimate, including discarded calibration and completed retries, is US$7.8896 across 2530 request reservations. Cached-input discounts are ignored, so this is conservative for known completed requests, not an invoice. The pipeline imposed an US$8 cumulative API cap within the user's EUR 10 budget, counting unresolved requests at their reserved maximum.

See statistics.json for counts. Generation date: 2026-10-05.

References

• Zhou et al. UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition. Paper.
• Ehrmann, Watter, Romanello and Clematide. Impresso Named Entity Annotation Guidelines, v2.2.0. Guidelines.
• BigLAM. Europeana Newspapers.