A Parliamentary Discourse Dataset from the German Bundestag

Structured, speaker-attributed corpus of all German Bundestag plenary session transcripts (Plenarprotokolle) from the first legislative period to the present (WP01–WP21, September 1949 – April 2026). Every attributed speech is extracted from the official PDFs published by the Deutscher Bundestag under open data policy and linked to the speaker's name, parliamentary role, party affiliation, and gender. Scale: 4,611 sessions · 1,033,723 speeches · 4,205 identified MdBs · 76 years of parliamentary debate Dataset files speeches.parquet — one row per attributed speech: speaker name, role, party, gender, stammdaten_id, full German text (~1 GB) persons.parquet — one row per MdB: cross-session identity linking all name variants via stammdaten_id; canonical name, birth date, career span, total speeches. Use this — not speakers.parquet — for person-level analysis sessions.parquet — one row per plenary session: date, city, Wahlperiode, source PDF hash, extraction engine speakers.parquet — name-string index: one row per unique name string as extracted from the transcripts. Useful for understanding extraction quality; not suitable for person-level aggregation (the same politician often appears under several name variants across sessions) parties.csv — reference table of 31 German parliamentary parties, 1949–present speeches.csv.gz — CSV fallback for Stata and Excel users (same columns as speeches.parquet) datapackage.json — Frictionless Data schema with column descriptions and foreign key constraints Cross-session identity The same politician often appears under different name strings across sessions (e.g. “Schmidt”, “Dr. Schmidt”, “Frau Dr. Schmidt”). Cross-session person linkage is provided via stammdaten_id, matched against the official Bundestag Stammdaten biographical XML. The persons.parquet table aggregates all name variants for the same MdB into one row with correctly summed speech counts, career span, and birth date. Coverage: ~98.5% of speeches are linked to a stammdaten_id; the remaining ~1.5% are ambiguous surname-only attributions or speakers not in the Stammdaten. Coverage and sources Source PDFs are the official Stenografische Berichte downloaded from the Bundestag open-data portal (bundestag.de). Two PDF generations are covered: scanned and OCR'd documents (WP01–WP09, Bonn era, 1949–1987) and born-digital documents (WP10–WP21, 1987–present). The engine column in sessions.parquet flags whether pdftotext (born-digital) or pdfminer (OCR fallback) was used; this is the primary data-quality indicator for NLP use. Speaker attribution Each speech is attributed using four patterns extracted from the transcript format: presiding officers (Präsident/in, Vizepräsident/in), regular members (name + party), government officials (name + Bundeskanzler/in, Bundesminister/in, etc.), and procedural roles (Berichterstatter/in, etc.). The party field is null for ~60% of speeches — this is expected, as presiding officers and ministers are not identified by party in the transcript. Gender annotation & distribution Gender is derived by matching speaker names against the official Bundestag Stammdaten biographical XML (all MdBs since 1949), with fallbacks for role title, honorific prefix, manually researched overrides, and a gender_guesser first-name heuristic. The gender_source column distinguishes stammdaten (authoritative, 83%), role_title (gendered job title in attribution, 6.4%), title_prefix (Frau/Herr honorific, 0.5%), manual (historically researched, 2.9%), and inferred (name-based heuristic, 4.7%). Gender distribution: Female 26.8% · Male 73.2% · Unknown 0.0%. Data quality All speeches pass automated validation: zero null speaker names, zero sequence gaps, zero CID artefacts, zero party-misclassified-as-Bundesland errors. Eight sessions with conflicting source PDFs were deduplicated (first lexicographic occurrence retained). 252 non-person names incorrectly accepted by the parser (table headers, legislative terms, agenda fragments) are excluded at build time via a curated exclusion list. OCR sessions (WP01–WP09) may contain Unicode replacement characters (U+FFFD); the engine field identifies these sessions. Licence CC BY 4.0. The underlying Plenarprotokolle are official government documents of the Deutscher Bundestag and are in the public domain.