edithatogo/nz-hansard-corpus

NZ Hansard Corpus Dataset Card

Dataset Summary

This workspace prepares a document-level corpus from New Zealand Hansard CSV extracts. The current pipeline inventories the source ZIP, discovers schema, normalizes document-level records to Parquet, validates a machine-readable record schema, and builds local DuckDB/search artifacts.

This card describes the published canonical document-level dataset. Publication does not imply endorsement by New Zealand Parliament.

Published Locations

• Hugging Face dataset: https://huggingface.co/datasets/edithatogo/nz-hansard-corpus
• Zenodo record: https://zenodo.org/records/20595194
• DOI: https://doi.org/10.5281/zenodo.20595194
• GitHub repository: https://github.com/edithatogo/corpus-nz-hansard
• GitHub release: https://github.com/edithatogo/corpus-nz-hansard/releases/tag/v0.1.0
• Corpus-family sibling: corpus-nz-legislation, published today as https://huggingface.co/datasets/edithatogo/nz-legislation-corpus

Source

• Source archive: 2024-09-06 Hansard Extract from DocumentsDB.zip
• Source archive SHA-256: 2ac02c0042a4fb291fd8e401db5f469de2539e42c9e07c4c72eca16be9a17299
• Source files: Hansard-47.csv through Hansard-54.csv
• Source scope: New Zealand Parliament Hansard records for Parliament numbers 47 through 54, as present in the supplied DocumentsDB extract.
• Coverage audit: manifests/historical_coverage_audit.json and docs/historical-coverage-audit.md.

The current publication verifies the supplied DocumentsDB extract and normalized row preservation. It does not claim full historical NZ Hansard completeness. Parliament 47-54 source-file presence is partial coverage evidence, not proof that every sitting, proceeding, vote, question, debate, speech, or document type for those parliaments is complete.

Legal and Provenance Notes

New Zealand Parliament's Parliamentary Practice material states that Hansard is the official report of debates in the House, and that no copyright exists in New Zealand Parliamentary Debates/Hansard. See docs/licensing-and-provenance.md for source links and caveats.

Original repository code, documentation, manifests, and release tooling are distributed under the MIT License. See NOTICE.md for the scope boundary between original project materials and the underlying parliamentary source text.

This dataset card does not provide legal advice.

Dataset Structure

The normalized Parquet dataset is document-level, not speech-turn-level.

Hugging Face viewer configuration:

• Config: default
• Split: train
• Data file: data/hansard.parquet
• Downloadable documentation, manifests, and schemas are intentionally excluded from viewer split detection.

This is the corpus-family Hugging Face layout pattern: place only viewer data under explicit configs[].data_files, keep manifests/schemas/docs as downloadable assets, and verify datasets-server splits, first-rows, and parquet endpoints after publication.

Columns:

• stable_id
• jurisdiction
• country
• source
• source_archive
• source_file
• source_row_number
• parliament_number
• parliament_document_id
• document_type
• title
• abbreviated_title
• status
• content
• member_of_parliament_raw
• member_of_parliament_count
• portfolio_raw
• last_modified
• document_content_date
• language
• text_sha256
• source_hash
• pipeline_version

Current Counts

• Source files: 8
• Source/schema-discovered rows: 193,922
• Normalized Parquet rows: 193,922
• Record schema validation errors: 0
• DuckDB rows: 193,922
• Normalization warnings: 0

Rows by Parliament:

Parliament	Rows
47	24,378
48	19,709
49	23,877
50	39,803
51	34,808
52	21,171
53	23,402
54	6,774

Intended Uses

• Local research over document-level Hansard text.
• Corpus-quality checks and exploratory analysis.
• Reproducible use of the public document-level dataset.
• Preparation for reporting, semantic model, or search/RAG tracks.

Out-of-Scope Uses

• Treating member_of_parliament_raw as a resolved member identity.
• Treating this MVP as a structured speech-turn corpus.
• Inferring party membership directly from this dataset.
• Claiming the dataset has been officially endorsed by Parliament.
• Claiming this dataset is a full historical NZ Hansard corpus.

Known Limitations

• No explicit source Party column exists.
• MemberOfParliament is raw and may contain multiple semicolon-separated names.
• Content is document-level text and may contain embedded separators.
• Coverage is limited to the supplied DocumentsDB extract. Historical completeness outside or within Parliament 47-54 remains unproven until official sitting and proceeding sources are reconciled.
• Licensing/provenance has been checked against official public guidance, but users should review the linked provenance notes for their own use case.
• Generated DuckDB output may include a .wal file on OneDrive-backed workspaces.

Distribution Policy

Public distribution policy:

• The source ZIP is not redistributed by default. Publication workflows require SOURCE_ARCHIVE_URL and verify the source archive SHA-256 before rebuilding.
• Hugging Face hosts the normalized document-level Parquet dataset.
• Zenodo hosts the citable archive and DOI record.
• DuckDB and SQLite search outputs are treated as regenerated/local convenience artifacts unless a reviewer explicitly asks for prebuilt database files.
• Non-authoritative speech-turn candidates are not part of the canonical document-level dataset.

Regeneration

See README.md and docs/pipeline-handoff.md for reproducible commands.