ResPolitica

A unified dataset for political information, including from Voting Advice Applications (VAAs).

Dataset Summary

ResPolitica currently covers party-position mediation: given a party and a policy statement, the party's stance and rationale. It unifies three sources — Wahl-O-Mat (German VAA), StemWijzer (Dutch VAA), and EUANDI 2024 (EU Parliament candidate positions) — into a single schema of 12,584 party × statement observations across 19 elections in 7 countries.

This is one task type within a broader political information mediation family the project is building toward: navigation ("which parties support X?"), comparison ("how do parties A and B differ?"), and issue mapping ("what are the main conflicts around X?") are planned but not yet covered.

Powers the Polistemics benchmark, which evaluates LLMs as mediators of political information in elections.

Supported Tasks

• Party-position retrieval: given (party_id, statement_id), retrieve stance_label and rationale_text.
• Suitable as a source corpus for RAG-style evaluation of LLMs mediating political information, or for direct classification/stance-detection tasks.

Languages

Statement and rationale text appear in the original source language (de, nl, or the source language for EUANDI rows) plus an English machine/human translation (statement_text_en, rationale_text_en).

Dataset Structure

Configs / Files

Config	File	Rows	Description
default	respolitica_unified.parquet	12,584	All observations, all sources, all fields
federal	respolitica_unified_federal.parquet	4,444	Federal-level elections only
states	respolitica_unified_states.parquet	7,030	State-level elections only
normalized_elections	respolitica_normalized/elections.parquet	19	Election registry
normalized_parties	respolitica_normalized/parties_canonical.parquet	143	Global party registry
normalized_parties_election	respolitica_normalized/parties_election.parquet	344	Party × election appearances
normalized_statements	respolitica_normalized/statements.parquet	666	Unique statements per election
normalized_observations	respolitica_normalized/observations.parquet	12,584	Fact table (FKs + stance + rationale)

The respolitica_unified* configs are flat, denormalized exports — every field on one row. respolitica_normalized/* is the relational decomposition of the same corpus (elections / parties / statements / observations as separate tables joined by foreign keys). Use whichever fits your workflow; both derive from the same source pipeline.

Data Fields (respolitica_unified)

Field	Type	Description
observation_id	string	Unique row identifier
source_dataset	string	qual-o-mat, stemwijzer, or euandi — see Licensing below
source_version	string	Upstream source snapshot/version identifier
election_id	string	Slug, e.g. bundestagswahl2025, nl_tk2025
election_date	date	ISO date
election_type	string	Election type
election_level	string	federal, state, or european
country_iso	string	ISO 3166-1 alpha-2 (or eu for EU Parliament)
region	string	Sub-national region, where applicable
election_year	int	Election year
party_id	string	Canonical party slug
party_name_short	string	Short party name
party_name_full	string	Full party name
party_euro_affiliation	string	European party family affiliation, where known
statement_id	string	Unique statement identifier
statement_number	int	Statement ordinal within its election
statement_text	string	Statement in source language
statement_text_en	string	English translation
statement_language	string	Language code of statement_text
statement_category	string	Topical category, where classified
stance_raw	string	Raw stance value from the source
stance_label	string	Agree, Disagree, or Neutral
stance_numeric	float	1.0 / 0.0 / -1.0
rationale_text	string	Party's rationale, source language
rationale_text_en	string	English translation
rationale_language	string	Language code of rationale_text
source_url	string	Link to the original source, where available
source_file	string	Originating raw source file
extraction_date	datetime	When this row was extracted from its source
quality_flag	string	e.g. missing_rationale, clean
rationale_length_chars	int	Character length of rationale_text
rationale_length_words	int	Word length of rationale_text
has_rationale	bool	False when rationale is blank or a placeholder
has_source_url	bool	Whether source_url is populated

respolitica_normalized/* tables carry a subset of these fields, decomposed relationally with foreign keys (election_id, party_id, statement_id) instead of denormalized on every row.

Source Composition

Source	Rows	Coverage
Wahl-O-Mat (qual-o-mat)	10,754	German federal + state elections
StemWijzer (stemwijzer)	720	Dutch TK2025 only (TK2017/21/23 excluded, see below)
EUANDI 2024 (euandi)	1,110	EU Parliament 2024 (DE, ES, FR, GR, IT + EU-level)

Rationale coverage (has_rationale) is 94.5% overall.

Dataset Creation

Built by a pipeline that fetches each source, parses it into a common schema, and unifies the result: respolitica_data.py fetch → respolitica_data.py prepare. See the source repository for the full pipeline.

Considerations for Using the Data

• StemWijzer TK2017/21/23 raw source files are excluded from the pipeline for copyright reasons — only TK2025 is included.
• EUANDI licensing (see below) restricts commercial use of that subset specifically.
• Rationale coverage is not 100% (94.5%) — filter on has_rationale if your use case requires it.
• Party positions and rationales are the parties' own public statements, standardized by this project's pipeline; they are not independently fact-checked.

Licensing Information

This dataset (the compiled/processed corpus) is released under CC BY 4.0 — share and adapt, including commercially, with attribution.

Exception: rows where source_dataset == "euandi" originate from the EUANDI 2024 dataset (European University Institute, Robert Schuman Centre), licensed CC BY-NC-SA 4.0 (non-commercial, share-alike). That restriction carries over to those specific rows.

For CC BY-only use (e.g. commercial), filter out the restricted subset:

df = df[df["source_dataset"] != "euandi"]

Citation

@misc{polistemics2026,
  title   = {Polistemics: Evaluating LLMs as Information Mediators in Politics \& Elections},
  author  = {Peters, Baran},
  year    = {2026},
  note    = {Preprint}
}