agoulah/ontario-hansard-official

Ontario Hansard (Official) — speaker-identity-resolved

~4.3 million paragraph-level records from the official Hansard of the Legislative Assembly of Ontario, parliaments 29–44 (1974–2025), with each paragraph's speaker resolved to a stable member identity and tagged with the confidence tier of that resolution.

Overview

• 4,316,254 records, one per Hansard paragraph.
• Parliaments 29–44, sessions spanning 1974–2025.
• 51 years of floor proceedings: debates, question period, statements, petitions, procedural business.
• Each record carries the verbatim speaker string, a resolved member_slug where identifiable, and a resolution_tier recording how the identity was established — so consumers can filter by confidence.

The dataset ships at paragraph grain. Turn-level records (one row per continuous run of a speaker) are one deterministic groupby away — see Turn reconstruction — while the reverse is lossy, so the finer grain is shipped.

Provenance & method

Source. Records were scraped from the ola.org Solr API, the Assembly's canonical public Hansard search backend. The scraper is retained in the project repository as the reproducibility path; this dataset ships the resulting static parquet directly (no fetch at load time).

Speaker-identity resolution. Each paragraph's verbatim speaker string (speaker_raw, e.g. Hon. Mr. Norton, The Acting Speaker (Mr. Cousens)) is resolved to a member identity through a tiered pipeline. The tier is preserved per record in resolution_tier — this provenance is the dataset's distinguishing feature: a consumer can restrict to deterministic-only identities, or include adjudicated ones, by filtering a column.

tier	meaning	member_slug	records
T1	Deterministic alias match (canonical name/alias table)	set	1,940,393
T2	Deterministic surname-unique-in-parliament match	set	1,663,375
T3A	AI-adjudicated residual, human-confirmed	set	23,398
T3J	Human-judged residual	set	123
T0	Procedural / chair / clerk / collective / non-member (registry)	usually null	686,935
T3	Residual unresolved (surname-only, no discriminator)	null	760
T4	Unresolved — OCR-mangled / non-member speaker string	null	1,270
	total		4,316,254

• T1/T2 are deterministic (no model in the loop). T2 is the identity layer for the scan-era parliaments, where it has no external ground truth (OLA's structured MemberID labels only begin in 2018) — its confidence is established by internal evidence instruments (see Data quality).
• T3A is an AI first pass over the residual, confirmed by a human-reviewed veto stage; T3J is the small set decided directly by a human.
• T0 is the procedural/non-member layer (Speaker, Clerk, "Hon. members", the Lieutenant-Governor, etc.), identified by registry and role.

Resolution rate. Excluding the procedural/non-member T0 layer (non-procedural denominator = 3,629,319):

• Deterministic (T1+T2): 99.30%
• Deterministic + adjudicated (T1+T2+T3A+T3J): 99.94%
• Residual unresolved (T3+T4): 0.056% (2,030 records).

Note on an earlier 99.88% figure. An internal report (recon/RESOLVER_STATS.md) cites 99.88% from an older resolver run. That run mislabelled 2,346 Lieutenant-Governor and other non-member records as unresolved failures (T4). Correctly classifying them as non-members (T0) is a correctness improvement, not a denominator trick: the same underlying data, better classified. The count of resolved members is identical either way (3,627,289); fixing the misclassification simply removes non-members from the "unresolved" bucket, and the rate is 99.94%. All figures here are computed from the data in this release.

Data quality / known exceptions

This corpus preserves text that the Assembly's own structured exports corrupt. A parity audit against OLA's published structured-Hansard CSV/XLSX exports found the exports are lossier than this scrape:

• The exports are Windows-1252, and mis-encode accented French as cp850 mojibake (e.g. d‚veloppement for développement); a subset is unrecoverable.
• Indigenous-language passages are destroyed: Oji-Cree syllabics appear in the export as ??? ????? ???. This scrape preserves the original Unicode.
• At the exact-character level the export text matches only 23.6% of the time; 99.96% under a corruption fold. The structured exports are useful for their MemberID labels only, never for text.

This dataset preserves the original Unicode text as scraped.

T2 confidence (internal-evidence instruments). Because T2 has no external ground truth in the scan era, its principal failure mode — an OCR mutation turning one surname into a different member's unique surname — was measured directly:

• Norton ↔ Worton (parls 30–32), measured and cleared. The one OCR-class (N↔W) collision between two high-volume co-sitting members. After splitting on honorific (Norton was a cabinet minister → Hon. Mr. Norton; Hon. Mr. Worton never occurs), the residual bare strings were content- and riding-scanned: the split is perfectly directional with zero cross-contamination — no Mr. Norton record references Worton's riding (Wellington South/Guelph) and no Mr. Worton record references Norton's (Kingston). The collision is theoretical, not realized.
• Distance-1 surname pairs reviewed and cleared. The remaining 40 edit-distance-1 near-miss pairs (61,036 records; e.g. Wilson/Tilson, Rae/Ray, Wells/Kells) were reviewed and found to be genuinely distinct co-sitting members, not OCR variants.

Deferred to v2.

• Date-partitioned identity. Mr. Villeneuve in parliament 32 (173 records) maps to more than one member across the parliament's span and requires a date-aware split; it is left at tier T3 (unresolved) in v1 rather than guessed.
• Committee transcripts. v1 is the official (House) Hansard config only.

Residual unresolved. ~64 distinct OCR-mangled speaker strings (and the Lieutenant-Governor / non-member set) remain at T3/T4 with member_slug null — no fuzzy salvage was attempted, by design.

Paragraph chain. Each record carries previous_id/next_id provenance links and a derived seq (within-sitting order). seq is the ordering contract and is clean for 100% of sittings; the raw prev/next links are auxiliary and are intact for 95.4% of sittings (4,234 / 4,436) — the remainder have id-sequence gaps from revisions or corrupt forward links, which do not affect ordering. Every deviation is enumerated in recon/CHAIN_VALIDATION.md. One paragraph (sitting 1987-01-29) has sentence text leaked into its id field — a single-record source wart.

Turn reconstruction

A turn is a maximal run of consecutive paragraphs by the same speaker within a sitting, in seq order. Ordering uses seq (so it is robust to the raw-chain defects above), but turn boundaries additionally break when the paragraph chain is broken — a revision-deleted gap — so two same-speaker paragraphs straddling deleted content are never merged into a fabricated-contiguous turn. seq is a sound total ordering, but contiguity detection for turn-merging consults previous_id — a revision-deleted gap is invisible in seq (which is dense by construction), so the reconstruction script breaks turns on the id chain, not on seq-adjacency; a naive groupby("seq") merge would silently fuse across such gaps. This snippet is identical to examples/reconstruct_turns.py:

import pandas as pd

def reconstruct_turns(df: pd.DataFrame) -> pd.DataFrame:
    df = df.sort_values(["parliament", "session", "date", "seq"])
    speaker = df["member_slug"].fillna(df["speaker_raw"])
    sitting = df["parliament"].astype(str) + "|" + df["session"] + "|" + df["date"].astype(str)
    # New turn when speaker changes, sitting changes, OR the paragraph chain is
    # broken — i.e. this row's previous_id is not the row physically before it.
    # A revision-deleted gap breaks the chain, so two same-speaker paragraphs
    # straddling deleted content are NOT merged into a fabricated-contiguous turn.
    # (seq is gapless by construction, so it cannot itself carry the gap; the
    # paragraph-id chain does. See SCHEMA.md.)
    chained = df["previous_id"] == df["id"].shift()
    new_turn = (
        (speaker != speaker.shift())
        | (sitting != sitting.shift())
        | ~chained
    )
    df = df.assign(_speaker=speaker, _turn=new_turn.cumsum())
    g = df.groupby("_turn", sort=False)
    return pd.DataFrame(
        {
            "parliament": g["parliament"].first(),
            "session": g["session"].first(),
            "date": g["date"].first(),
            "seq_start": g["seq"].first(),
            "speaker": g["_speaker"].first(),
            "speaker_raw": g["speaker_raw"].first(),
            "member_slug": g["member_slug"].first(),
            "resolution_tier": g["resolution_tier"].first(),
            "n_paragraphs": g.size(),
            "id_first": g["id"].first(),
            "id_last": g["id"].last(),
            "text": g["text"].apply(lambda s: "\n\n".join(s.astype(str))),
        }
    ).reset_index(drop=True)

Worked example — Question-Period day 2025-04-16 (parliament 44), regenerated from the shipped data by running examples/reconstruct_turns.py:

1173 paragraphs -> 341 turns on 2025-04-16
  [T0] Procedure (1 para): The House met at 0900.…
  [T1] donna-skelly (1 para): Good morning, everyone.…
  [T0] Prayers (1 para): Prayers.…
  [T1] donna-skelly (11 para): I beg to inform the House that the Clerk has laid upon the table the r…
  [T1] steve-clark (3 para): I move, seconded by the member for Essex, that an humble address be pr…
  [T1] donna-skelly (4 para): Mr. Clark moves, seconded by Mr. Leardi, that an humble address be pre…

The full runnable script is examples/reconstruct_turns.py in the project repository.

Schema

The frozen 17-column contract is defined in SCHEMA.md. Summary:

column	type	null	notes
id	string	no	OLA paragraph ID, primary stable key
previous_id / next_id	string	yes	para-chain provenance links
seq	int32	no	within-sitting order (ordering contract)
date	date32	no	sitting date
parliament	int64	no	29–44
session	string	no	e.g. "44-1"
speaker_raw	string	no	verbatim speaker string
member_slug	string	yes	resolved identity; null if unresolved/non-member
resolution_tier	string	no	T0|T1|T2|T3A|T3J|T3|T4
tier_reason	string	yes	T0 role / T4 reason subtag
topic	string	yes	source topic
type_of_business	string	yes	source business type
hansard_title	string	yes	source title
text	string	no	the paragraph content
url / para_url	string	yes	provenance URLs

from datasets import load_dataset
ds = load_dataset("<repo>", "official", split="train")
# deterministic-identity subset only:
ds_det = ds.filter(lambda r: r["resolution_tier"] in ("T1", "T2"))

Licensing & attribution

Ontario Hansard — the Official Report of Debates — is a publication of the Legislative Assembly of Ontario and the underlying proceedings are subject to Crown copyright. The Assembly publishes structured Hansard in bulk, machine-readable form for public download via its Data Resources page. This dataset is a derived, speaker-resolved reformatting of that same public record.

This dataset is not an official version of the proceedings of the Legislative Assembly of Ontario, and is not affiliated with, produced by, or endorsed by the Assembly. For official records, consult ola.org.

Attribution: Legislative Assembly of Ontario, Official Report of Debates (Hansard).

Use of the underlying Hansard text is governed by Crown copyright and any applicable terms of the Legislative Assembly of Ontario; users are responsible for determining the terms that apply to their use. This notice describes the status of the source material and does not grant, extend, or restrict any rights in it. The dataset's structure, speaker-resolution, and accompanying code are the work of the dataset author; see the project repository for their terms.