bil-y/polidata-de-stance-prediction

Party Stance Prediction (Wahl-O-Mat → EU and I)

Stance-prediction task: given a party and a German political statement, predict the party's three-point position (stimme nicht zu / neutral / stimme zu).

• train / validation / test ← the Wahl-O-Mat product (3,780 / 504 / 504 rows = 630 / 84 / 84 theses; party positions on federal, state and European election theses). Every thesis is answered by all six major blocs (afd, cdu_csu, fdp, gruene, linke, spd), so the splits are rectangular and align exactly with ood_test. validation and test are thesis-cluster holdouts stratified per election — Wahl-O-Mat reuses the same thesis across elections (reworded by state/election type), so near-duplicate theses are grouped by cluster_id and whole clusters are held out. Each of the 21 elections contributes exactly 4 theses to validation and 4 to test, so no Bundesland or year is over-represented. The model is thus evaluated on genuinely unseen theses / seen parties.
• ood_test ← the EU and I 2024 product (177 rows; the six German parties on the European election VAA), an out-of-distribution test set from a different instrument.

Two deliberate consequences. Holdouts are drawn only from clusters confined to a single election, so theses Wahl-O-Mat reused across elections are always in train — evergreen national topics are never tested. And train therefore carries the near-duplicates: 630 theses but 356 distinct questions (1.8× redundancy), while both holdouts are 1.0×. Group by cluster_id to de-duplicate or down-weight them.

from datasets import load_dataset
ds = load_dataset("bil-y/polidata-de-stance-prediction")  # splits: train (3,780), validation (504), test (504), ood_test (177)

Harmonisation (so the Wahl-O-Mat splits and the OOD test align)

• Language. Wahl-O-Mat theses are German; the EU and I statements (English) are replaced by curated German translations.
• Label. EU and I's five-point scale is collapsed by sign to the Wahl-O-Mat three-point scale (-1/0/1); "No opinion" rows are dropped.
• Party. Aligned at the bloc level (party_bloc): Wahl-O-Mat CDU / CSU (cdu_csu) and EU and I CDU (cdu) both roll up to cdu_csu. Join the Wahl-O-Mat splits and ood_test on party_bloc.

The party's justification (justification, German; justification_en English where available) is included for reasoning use-cases — but do not feed it when predicting answer: it states the rationale and would leak the label.

Fields

field	type
id	Value('string')
source	Value('string')
source_ref	Value('string')
cluster_id	Value('int64')
party_id	Value('string')
party_bloc	Value('string')
party	Value('string')
scope	Value('string')
year	Value('int16')
topic	Value('string')
statement	Value('string')
answer	ClassLabel(names=['stimme nicht zu', 'neutral', 'stimme zu'])
stance	Value('int8')
justification	Value('string')
justification_en	Value('string')

License & attribution

• train / validation / test — derived from the Wahl-O-Mat datasets published by the Bundeszentrale für politische Bildung (bpb), the sole copyright holder (datasets). We are not the bpb and claim no ownership; the bpb is not the author of this analysis. Use is permitted for scientific/journalistic analysis per the bpb terms.
• ood_test — EU and I data, licensed CC-BY-NC-SA-4.0.

Provenance

Built from the wahl_o_mat and euandi_2024 products. Field descriptions and label codes live in datapackage.json next to the processed splits.