We present Swiss4Arg, a multilingual dataset for argument mining based on official Swiss voting brochures, covering all four national languages: German, French, Italian, and Romansh (a Romance minority language exclusively spoken in Switzerland). The brochures contain explicitly formulated arguments systematically organized into pro and con sections, resulting in a cross-lingually aligned resource with balanced argumentative stance. We create a dataset that comprises 53 federal voting issues collected between 2019 and 2025, annotated for argument components (major claims, claims, premises, and non-argumentative units) and relations (support and attack) by three native speakers per language. Inter-annotator agreement (Krippendorff's alpha) reaches 0.548 for component classification and 0.623 for relation and target classification, computed jointly across all annotators on the parallel data. We release both adjudicated labels and individual annotator annotations, enabling analyses of disagreement and cross-linguistic variation, and provide baseline results using GPT-5.4 and SwissBERT. Swiss4Arg establishes a novel benchmark for multilingual argument mining, with a particular focus on cross-lingual comparability and low-resource settings such as Romansh. The dataset and code are publicly available at https://anonymous-hf.up.railway.app/a/gozw0nion72y/ (HuggingFace) and https://anonymous.4open.science/r/MultiVote-03BD/ (GitHub).
Figures & tables
Figure 2: Example of manual sentence alignment across languages: sentences are merged into segments so that each unit captures the same semantic content in all four languages (here, the German and Romansh sentences in row 2 are moved up to align with the Italian and French counterparts).
Figure 3: Illustration of the major claim generation process. The official German summary is condensed into a single synthetic major claim using GPT-5-nano. In this example, the generated German major claim (Same-sex marriage should be allowed) is translated into French, Italian, and Romansh using Supertext.
Component
Relation
Claim
Premise
Non-Arg.
Total
Percentage
Support
Attack
Total
Percentage
Unanimous
368
307
168
843
29.3%
1530
200
1730
67.0%
Majority
398
1435
129
1962
68.1%
645
177
822
31.8%
Tie
22
53
1
76
2.6%
21
10
31
1.2%
Adjudicated
793
1789
298
2881
2196
386
2583
Table 1: Annotator agreement per label. The Adjudicated row gives the final label counts after adjudication.
Figure 4: Pairwise Krippendorff’s α for component (top) and relation+target classification (bottom) across all 12 annotators.
GPT-5.4
SwissBERT
Lang.
Label
Prec.
Rec.
F1
Macro F1
Prec.
Rec.
F1
Macro F1
DE
Claim
0.395
0.447
0.419
0.543
0.871
0.704
0.778
0.866
Premise
0.807
0.614
0.698
0.873
0.942
0.906
Non-Arg.
0.379
0.792
0.512
0.936
0.889
0.912
FR
Claim
0.381
0.423
0.400
0.528
0.865
0.696
0.771
0.856
Premise
0.810
0.604
0.692
0.866
0.932
0.898
Table 2: Component classification results per language and overall for GPT-5.4 and SwissBERT. Precision, Recall, F1 = per-class F1, and Macro F1 reported per language.
GPT-5.4
SwissBERT
Lang.
N
EM
F1
N
EM
F1
DE
1455
0.40
0.30
2244
0.51
0.59
FR
1417
0.34
0.23
2220
0.48
0.47
IT
1538
0.47
0.36
2226
0.48
0.47
RM
1567
0.49
0.36
2232
0.49
0.48
Overall
5977
0.43
0.32
8922
0.489
0.477
Table 3: GPT 5.4 and SwissBERT relation and target prediction results evaluated only on units where the component label was correctly predicted. EM = exact match on the joint relation+target label. F1 = Macro F1 Score.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Example text from VoteInfo App. More examples and the texts in the other three languages can be found there or at https://www.bk.admin.ch/de/sammlung-der-abstimmungsbuechlein-seit-1978
Language
2-vs-1
3-way
% Majority
French
811
31
96.3%
German
769
52
93.7%
Italian
1387
131
91.4%
Romansh
943
44
95.5%
Overall
3910
258
93.8%
Appendix
Table 4: Within-language disagreement types per language team. 2-vs-1: two annotators agree, one disagrees. 3-way: all three annotators assign different labels. The % Majority column reflects the proportion of within-language disagreements resolvable by majority vote and serves as a proxy for annotation difficulty per language.
Language
Component
Rel.+Target
French
0.662
0.693
German
0.685
0.678
Romansh
0.624
0.581
Italian
0.463
0.371
Appendix
Table 5: Pooled Krippendorff’s α per language for component and relation+target annotation.
Figure 7: Pairwise confusion matrices for component labels within the French annotation team.
Figure 8: Pairwise confusion matrices for component labels within the German annotation team.
Figure 9: Pairwise confusion matrices for component labels within the Italian annotation team.
Figure 10: Pairwise confusion matrices for component labels within the Romansh annotation team.
Figure 11: Component label distribution per language across all individual annotators. Italian annotators assign Claim at a substantially higher rate (42.0%) and Premise at a lower rate (35.8%) compared to the other language teams (Claim: 25.8–29.4%, Premise: 54.6–59.9%), driven primarily by IT-1.
Figure 12: Component confusion matrices for GPT-5.4 predictions vs. majority-vote ground truth, per language. Rows = ground truth, columns = predicted. Major Claims are excluded.