Dataset FZJ-2026-05619

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png
The Bundesrat corpus: A dataset of political discourse in the German Federal Council

 ;  ;  ;  ;

2026
Zenodo

Zenodo () [10.5281/ZENODO.22086439]

This record in other databases:  

Please use a persistent id in citations: doi:

Abstract: Structured, speaker-attributed corpus of all German Bundesrat (Federal Council) plenary session transcripts (Plenarprotokolle) from the chamber's first sitting to the present (sessions 1–1067, September 1949 – July 2026). Every attributed contribution is extracted from the official PDFs published by the Bundesrat and linked to the speaker's name, resolved person identity, official role, gender, and Bundesland (Federal State) the speaker represents. Scale: 1,067 sessions · 53,362 attributed contributions (47,290 spoken in the chamber, 6,072 submitted in writing) · 2,473 identified persons · 76 years of proceedings Why this chamber is different The Bundesrat is not a second elected chamber. Its members are delegates of the sixteen Land governments, appointed and recallable by them. Article 51(3) of the Basic Law requires each Land to cast its votes einheitlich as a bloc and the chamber has no Fraktionen (no party affiliations). There is therefore no party field, and its absence is a design decision rather than a gap: the operative political unit is territorial. Every contribution carries the Bundesland it was made for, which makes territorial representation and federal–state conflict directly measurable from text. The chamber is also permanent, never dissolved, no elections, continuous session numbering since 1949 so the corpus is a single unbroken series rather than one segmented by legislative period. Dataset files speeches.parquet, one row per attributed contribution by a named member: speaker name, resolved person_id, role, Bundesland, full German text, and contribution_type marking whether it was spoken in the chamber or submitted to the record in writing (102 MB). Gender is not stored here; join persons.parquet on person_id. persons.parquet, one row per resolved individual: canonical name, all name variants merged, career span, modal role and Bundesland, gender. Use this, not speakers.parquet, for person-level analysis. sessions.parquet, one row per sitting: date, meeting city, extraction engine, source PDF hash, presiding_land (the Land holding the rotating annual presidency), and separate counts of spoken and written contributions. speakers.parquet, name-string index: one row per unique name string as extracted. Useful for assessing extraction quality; not suitable for person-level aggregation. laender.csv, reference table resolving the 19 values the land field can take, with observed first and last year and a flag for the three historical Länder. CSV twins, every table is also published as CSV with identical contents, for Stata, SPSS and Excel users. datapackage.json, Frictionless Data schema with column descriptions and foreign-key constraints. Spoken and written contributions A member who cannot attend may submit a statement to the record; it is printed in the appendix and marked in the protocol's index with an asterisked page reference. Both kinds are held in speeches.parquet and separated by contribution_type: spoken for the 47,290 delivered in the chamber, written for the 6,072 given to the record but never said aloud. Written statements are negligible before 1970, with 44 in the whole of the 1960s and three quarters of those in 1968 and 1969, and routine afterwards. Any time-series analysis should therefore filter to contribution_type == 'spoken': combining the two introduces a step change around 1970 that reflects the practice taking hold, not a change in the chamber. Every completeness and recall figure quoted below is measured on the spoken subset alone. Territorial attribution land carries 19 values: the 16 current Bundesländer plus Baden, Württemberg-Baden and Württemberg-Hohenzollern, the three founding members that merged into Baden-Württemberg in April 1952. Historical Länder keep their own names, folding them into the successor would rewrite the 1949–52 delegations. land is null for 58% of spoken speeches by design: the presiding officer and federal officials represent no Land, and the chair alone accounts for roughly half of all speeches. Joining sessions.presiding_land raises territorial attribution from 41.7% to 99.3%. Cross-session identity Unlike the Bundestag, the Bundesrat has no official biographical register, its members sit for Land governments and are absent from the Bundestag Stammdaten. Person identity is therefore derived from the transcripts themselves: name strings are grouped by folded surname and merged where Bundesland, first name, career era and gender are all compatible, with the mapping published as a reviewable CSV rather than baked into code. 2,907 name strings resolve to 2,473 persons, covering 99.7% of contributions. The remainder are four components in which a single name string spans two different people, a distinction no name-to-person mapping can express, and those rows carry no identity rather than a wrong one. Coverage and sources Source PDFs are the official protocols from the Bundesrat's archive (bundesrat.de). Two document generations are covered: 747 sessions (1949–1999) survive only as page images and were recovered by OCR (Tesseract, German), roughly 27,000 pages released in machine-readable form here for the first time; 320 sessions (2000–2026) carry a PDF text layer read with pdftotext. The engine column in sessions.parquet flags which route was used and is the primary data-quality indicator for NLP use. Speaker attribution Contributions are attributed using four patterns: the presiding officer (Präsident/in, Vizepräsident/in), Land delegates (name + Bundesland), federal officials (name + Bundesminister/in etc.), and procedural roles (Berichterstatter/in, Schriftführer/in). Each appears in two typographic variants, surnames are set in capitals before about 1960 and in title case afterwards. role is null for 34% of spoken speeches, which is expected: a Land delegate speaking for their Land holds no procedural office. Gender annotation German inflects office nouns, so the protocol states gender directly, Präsidentin, Bundesministerin, Berichterstatterin alongside the honorific Frau. The field records forms of address, not biography, and is null rather than imputed where the record is silent. gender_source distinguishes the evidence: address (the speaker's own attribution line, 96.2%), firstname (given-name inference, 1.4%) and mention (how other speakers refer to them, 0.5%). Distribution across contributions: Male 84.9% · Female 13.2% · Unknown 1.9%. Restricted to speeches delivered in the chamber the shares are 86.9%, 12.2% and 0.7%, and the share of speeches delivered by women rises from 0.1% in 1949–59 to 40.4% in the 2020s. Data quality All contributions pass automated validation: zero null speaker names, zero page headers or (cid:N) artefacts inside speech text, zero column markers splitting a word, zero land values outside the 19 permitted, and zero Länder speaking outside their own existence. A standing audit of fifteen checks ships with the code and gates each at a documented baseline that may be lowered as defects are fixed but never raised. Licence CC BY 4.0. The underlying Plenarprotokolle are official documents of the Bundesrat.

Keyword(s): Bundesrat ; German Politics ; Political Discourse ; Parliamentary debate


Contributing Institute(s):
  1. Jülicher Systemanalyse (ICE-2)
Research Program(s):
  1. 1111 - Effective System Transformation Pathways (POF4-111) (POF4-111)
  2. 1112 - Societally Feasible Transformation Pathways (POF4-111) (POF4-111)

Appears in the scientific report 2026
Click to display QR Code for this record

The record appears in these collections:
Document types > Other Resources > Datasets
Institute Collections > ICE > ICE-2
Workflow collections > Public records
Publications database

 Record created 2026-11-06, last modified 2026-11-06


External link:
Download fulltext
Fulltext
Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)