Datasetpublic

biglam/presto-rumour-claims

PRESTO Corpus — Rumour Claims from 19th-Century British Newspapers 7,460 rumour claims automatically extracted from 19th-century British newspapers, plus a 200-row hand-adjudicated evaluation sample. The source material is the British Library's Heritage Made Digital newspaper collection — the same corpus published as biglam/hmd_newspapers. PRESTO (Pattern-based Rumour Extraction with Semantic Tracking) applies dependency-pattern matching to find and structure rumoured… See the full description on the dataset page: https://huggingface.co/datasets/biglam/presto-rumour-claims.

sourceHugging Facecc0-1.0updated 2mo agoView on Hugging Face

1likes31downloads

Dataset Card

PRESTO Corpus — Rumour Claims from 19th-Century British Newspapers

7,460 rumour claims automatically extracted from 19th-century British newspapers, plus a 200-row hand-adjudicated evaluation sample.

The source material is the British Library's Heritage Made Digital newspaper collection — the same corpus published as `biglam/hmd_newspapers`. PRESTO (Pattern-based Rumour Extraction with Semantic Tracking) applies dependency-pattern matching to find and structure rumoured propositions in OCR'd text.

Configs

Config	Rows	What it is
`default`	7,460	Silver-standard: machine-extracted claims.
`gold`	200	Hand-adjudicated sample with `extraction_correct` + `extraction_reason`.

Fields (`default`)

—matched_sentence — the source sentence containing the rumour cue
—rumour_content — the extracted rumoured proposition
—pattern_type — dependency pattern used (standalone, that, of)
—ner_person / ner_gpe / ner_org — entity lists (parsed to real lists; the source CSV stored these as stringified Python lists)
—title, location, date — publication metadata (dates normalised to YYYY-MM-DD; range 1801–1888)
—matched_variant, impresso_ocr_score, extraction_success, rumour_content_entity_count

Caveats

Read these before using the default config as ground truth.

—It is silver-standard. The 7,460 rows are machine-extracted. Only the 200 gold rows are human-checked. Validation accuracy is reported in the source deposit, not here.
—`impresso_ocr_score` is constant — every row is 1.0. It carries no information in this release.
—`extraction_success` is constant — every row is True (failed extractions were not retained).
—`pattern_type` has 3 values, not 5. The source description mentions it-is-rumoured and to-the-effect patterns, but no rows in this release use them.

Load

python

from datasets import load_dataset

silver = load_dataset("biglam/presto-rumour-claims", "default", split="train")  # 7,460 machine-extracted
gold   = load_dataset("biglam/presto-rumour-claims", "gold", split="train")     # 200 hand-adjudicated

# only the gold config can tell you whether an extraction was right
correct = gold.filter(lambda r: r["extraction_correct"])
print(f"{len(correct)}/{len(gold)} adjudicated correct")

Source & credit

Wanshu Zhang. PRESTO Corpus: Structured Rumour Claims from 19th-Century British Newspapers. Zenodo, 2026-05-08. <https://zenodo.org/records/20077470> — CC0-1.0.

Underlying newspapers: British Library Heritage Made Digital, public domain.

This repository is a format conversion (CSV → Parquet) for easier loading. All content is the original authors'. Please cite the Zenodo deposit.

bibtex

@dataset{zhang_2026_presto,
  author    = {Zhang, Wanshu},
  title     = {{PRESTO Corpus: Structured Rumour Claims from 19th-Century British Newspapers
               (Heritage Made Digital)}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20077470}
}