parliament-data/hansard-embeddings

Hansard Embeddings & Derived Analytics

Text embeddings and derived analytics cubes covering the parliamentary Hansard of four Australian jurisdictions: the Commonwealth (au), New South Wales (nsw), South Australia (sa), and Western Australia (wa).

No Hansard prose is in this dataset. It ships vectors, join keys, and derived metadata (subject headings, member names, parties, counts). The official record remains with each parliament; to hydrate the text, run the open-source harvester against the official sources yourself — hansard-researcher.

Over the last 20 years, some Parliaments have published their Hansard in a specific Hansard XML format. The complication at the moment is that you have to go to each Parliament to access their records. This dataset is derived from those XML sources, but the pipeline is designed to be agnostic to the source format. The harvester extracts the prose and metadata, the normalizer converts it into a common schema, and the enrichment step produces embeddings and derived analytics cubes. The cubes are designed to support research and analysis without ever needing to ship the prose itself.

What's in the repo

embeddings/   one vector per text unit, hive-partitioned:
              model_slug=<slug>/jurisdiction=<j>/year=<yyyy>/part-*.parquet
gold/         derived analytics cubes (*.parquet), the join target for
              subject_id / member ids
qdrant/       (optional) Qdrant collection snapshot — restore and search
              without re-indexing

Vector-space contract

Query vectors are only comparable if produced identically:

Model	nomic-embed-text (137M) served via Ollama
Dimensions	768
Distance	cosine
Task prefix	none — embed queries with no search_query: prefix

The model_slug hive partition versions the vector space. If the corpus is ever re-embedded with a different model or serving stack, the new vectors publish alongside under a new model_slug — different implementations are not the same vector space, so never mix partitions in one index.

embeddings/ schema

One row per embedded text unit (e.g. "a contiguous spoken passage within a subject").

column	type	notes
text_id	VARCHAR (UUID)	primary key of the text unit
fragment_id	VARCHAR (UUID)	the silver fragment the text came from
talker_id	VARCHAR (UUID)	speaker turn; joins gold/qa_pairs.question_talker_id / .answer_talker_id. ~84% populated — NULL is structural (procedural/chair text carries no speaker attribution)
subject_id	VARCHAR (UUID)	debate subject; joins the subject-grain gold cubes. ~99% populated — NULL is procedural preamble before the first subject
model	VARCHAR	raw model id (nomic-embed-text)
provider	VARCHAR	serving stack that produced the vector (ollama)
dim	INTEGER	768
embedding	FLOAT[]	the vector
date	DATE	sitting date
house	VARCHAR	chamber (or committee volume)
jurisdiction	VARCHAR	au / nsw / sa / wa (also a hive partition)
model_slug	VARCHAR	filesystem-safe model id (also a hive partition)
year	INTEGER	hive partition, derived from date

Join contract

The vectors are deliberately prose-free; the gold cubes make them usable:

• subject_id → gold/subject_occurrence, gold/contributions, gold/qa_pairs, gold/division_summary, gold/division_votes_detail — subject headings, who spoke, Q&A pairings, votes.
• talker_id → gold/qa_pairs question/answer talker ids.
• Member identity across cubes is member_source_id (per-jurisdiction official id) + member_name.
• text_id / fragment_id → resolve to prose only via a local run of the harvester + normalizer (silver layer). Prose never ships here.

Example — label a nearest-neighbour hit without any prose:

-- duckdb
SELECT e.text_id, s.subject_name, s.date, s.house, s.jurisdiction
FROM read_parquet('hf://datasets/parliament-data/hansard-embeddings/embeddings/**/*.parquet', hive_partitioning=1) e
JOIN read_parquet('hf://datasets/parliament-data/hansard-embeddings/gold/subject_occurrence.parquet') s
  USING (subject_id)
LIMIT 10;

gold/ cubes

These are as of 2026-07-07

cube	rows	grain
contributions	629,054	member × subject × house-day: turns/speeches/questions/answers/interjections/words
subject_occurrence	353,285	subject × house-day: participation counts, first-spoken time, linked bills
qa_pairs	135,677	paired question & answer with members, parties, portfolios, word counts
division_votes_detail	891,335	individual member vote in each division
division_summary	14,055	division outcomes with ayes/noes/margin
bills	7,339	bill: houses, sitting span, latest stage, debate volume
bill_journey	31,484	bill × house-day: stages reached, debate/divisions that day
bill_theme_link	36,888	bill × theme association
member_activity	1,255	member career totals
member_activity_by_week	98,417	member × ISO week activity
member_theme_rank	25,788	member ranking within each theme
member_vote_by_theme	19,042	member vote tallies by theme
theme_by_week	42,970	theme prevalence per week
theme_cooccurrence	1,288	theme pair co-occurrence
theme_share_by_jurisdiction	104	theme share per jurisdiction
theme_subject_names	1,040	top subject names per theme
theme_coverage	4	classification coverage per jurisdiction
theme_candidates	0	low-confidence classifications flagged for curation (empty when the classifier places everything above the cutoff)
sitting_days	5,243	house-day: session, times, duration, volume
pipeline_coverage	5,243	per house-day provenance: harvested/normalized/embedded/themed

Theme cubes carry engine/model columns — theme labels are themselves model-derived; treat them as annotations, not ground truth.

Every cube is self-describing — read the column list straight off the parquet:

-- duckdb
DESCRIBE SELECT * FROM read_parquet(
  'hf://datasets/parliament-data/hansard-embeddings/gold/contributions.parquet');

Optional extra: Qdrant snapshot

Everything above is fully usable without this — the parquet is the dataset. For anyone who wants batteries-included semantic search, qdrant/ holds a point-in-time collection snapshot: restore it into a Qdrant instance and query — no re-indexing of 6.3M points. (~25 GB; re-cut occasionally, so it may trail the parquet shards between cuts.) Point payloads carry jurisdiction, date, house, subject_id, talker_id, plus citation metadata so search results cite the official record without any prose: subject/debate context (subject_name, subject_uid, proceeding_name, subproceeding_name, committee_name, bill_names, extract_index), speaker (speaker, party, party_abbreviation, electorate, role, talker_kind), citation position (text_kind, page_no, time_anchor — ISO-8601 UTC), sitting formalities (parliament_num, session_num, review_stage) and provenance (source_url). Nulls are dropped per point, not stored; speaker fields are structurally absent on procedural/chair text.

curl -X POST 'http://localhost:6333/collections/<name>/snapshots/upload' \
  -H 'Content-Type: multipart/form-data' \
  -F 'snapshot=@hansard-embeddings.snapshot'

Provenance & licensing

• Derived from the official Hansard published by each parliament. The parliamentary record itself remains subject to each parliament's terms:

  Jurisdiction	Official source	Terms on the record
  Australia (Federal)	parlinfo.aph.gov.au	CC BY-NC-ND 4.0
  New South Wales	parliament.nsw.gov.au/hansard (API: api.parliament.nsw.gov.au)	Parliamentary copyright
  South Australia	hansardsearch.parliament.sa.gov.au	Parliamentary copyright
  Western Australia	parliament.wa.gov.au/hansard (public API)	CC BY-ND 4.0

  Those terms bind the prose, which is why none ships here; the full per-jurisdiction analysis is in the pipeline repo's LICENSES-DATA.md.

• Vectors and derived cubes in this dataset: CC-BY-4.0

• Pipeline: harvest → normalize → enrich (embed/themes) → aggregate, all open source at parliament-io/hansard-researcher. gold/pipeline_coverage records per-house-day provenance including harvested_at timestamps.

• Updated weekly by an incremental upload from the maintainer's pipeline; the current year's shards churn, historical shards are stable.

Citation

@misc{novaworks2026hansard,
  author    = {{NovaWorks Group Pty Ltd}},
  title     = {Parliamentary Hansard --- Embeddings \& Derived Analytics},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/parliament-data/hansard-embeddings},
  note      = {Project: \url{https://parliament.io}. Updated continuously; cite the access date.}
}