Elamur Research⌕
☰
home / papers / 2026 / How Good is AI on Swiss Voting Booklets? A Multilingual OCR and Alignment Benchmark

How Good is AI on Swiss Voting Booklets? A Multilingual OCR and Alignment Benchmark

Paper · 2026 · Elina Stüssi and Jannis Vamvas

Summary

The paper tackles turning Switzerland's federal voting booklets into a clean, machine-readable parallel corpus. These official referendum guides carry the same carefully edited content in German, French, Italian and the low-resource Romansh Grischun. Two linked tasks are involved: OCR of scanned pages and cross-lingual alignment, where OCR errors can carry over into alignment. The authors built VotingBooklets-Diamond, a hand-corrected, multi-annotator-checked test set. It contains eleven booklets from votes in 1977, 1985 and 2007, mixing old scans with born-digital PDFs, and includes paragraph-level gold alignments. On this set they compared Tesseract-based pipelines, LLM post-correction, multilingual embedding models and Gemini variants. Direct OCR with Gemini 2.5 Flash Lite performed best across all conditions. Alignment worked best with a hybrid method: Sentence-SwissBERT proposes matches, and Gemini repairs only the low-confidence pairs. Running this best pipeline over the entire booklet archive produced VotingBooklets, a four-language corpus of several million tokens. It is useful for Romansh NLP, multilingual embeddings and longitudinal research on political discourse.

Key points

  • Two resources: VotingBooklets-Diamond is the gold test set, with about 7.6k DE, 9.6k FR, 8.6k IT and 4.8k RM tokens (no Romansh in 1977). VotingBooklets is the full corpus, with roughly 1.12M DE, 1.42M FR, 1.07M IT and 0.86M RM tokens over about 4–5k pages per language. Both are on Hugging Face under CC BY 4.0, with code on GitHub.
  • OCR comparison: Systems were scored by WER/CER. Gemini 2.5 Flash Lite beat both Pytesseract and Pytesseract plus Gemini post-correction almost everywhere. On the 2007 booklets, Pytesseract exceeded 0.20 WER, while Gemini stayed below 0.08.
  • Romansh OCR: Romansh did not lag behind the other languages. Gemini reached a CER of 0.0024 for 1985 and 0.0095 for 2007.
  • Alignment systems compared: paraphrase-multilingual-MiniLM-L12-v2, Sentence-SwissBERT, Gemini-only alignment, and two hybrids (SSB+Gemini3 and SSB+Gemini2.5). The SSB-plus-Gemini correction approach performed best, with notable gains for Romansh.
  • Annotation method: Gold transcriptions follow the physical page order rather than reading order. Each transcription and each alignment was verified by two additional annotators, and all disagreements were fully resolved.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksEngineering · 133 worksMedia Technology · 34 works

🧭 Topics 2Research subjects and languages the work is about.Evaluation & benchmarks · 1,220 worksMultilingual & cross-lingual NLP · 828 works

Cite

Cite as Elina Stüssi and Jannis Vamvas. 2026. How Good is AI on Swiss Voting Booklets? A Multilingual OCR and Alignment Benchmark. Proceedings of the 11th Edition of the Swiss Text Analytics Conference.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{stussi2026how,
  title = {{How Good is AI on Swiss Voting Booklets? A Multilingual OCR and Alignment Benchmark}},
  author = {Elina Stüssi and Jannis Vamvas},
  booktitle = {Proceedings of the 11th Edition of the Swiss Text Analytics Conference},
  year = {2026},
  doi = {10.5167/uzh-434762},
  url = {https://doi.org/10.5167/uzh-434762}
}

References

24 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pierre Andrews, Mikel Artetxe, Mariano Coria Meglioli, Marta R. Costa-jussà, Joe Chuang, David Dale, Mark Duppenthaler, Nathanial Paul Ekberg, Cynthia Gao, Daniel Edward Licht, Jean Maillard, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Eduardo Sánchez, Ioannis Tsiamas, Arina Turkatenko,… find
  2. Mikel Artetxe and Holger Schwenk. 2019. Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond. Transactions of the Association for Computational Linguistics, 7:597– 610.
  3. Daniel Deutsch, Eleftheria Briakou, Isaac Rayburn Caswell, Mara Finkelstein, Rebecca Galor, Juraj Juraska, Geza Kovacs, Alison Lui, Ricardo Rei, Jason Riesa, Shruti Rijhwani, Parker Riley, Elizabeth Salesky, Firas Trabelsi, Stephanie Winkler, Biao Zhang, and Markus Freitag. 2025. WMT24++: Expanding…
  4. Christa Dürscheid and Elisabeth Stark. 2011. sms4science: An International Corpus-Based Texting Project and the Specific Challenges for Multilingual Switzerland. In Crispin Thurlow and Kristine Mroczek, editors, Digital Discourse. Language in the New Media, Oxford Studies in Sociolinguistics, page 2… find
  5. William A. Gale and Kenneth W. Church. 1993. A Program for Aligning Sentences in Bilingual Corpora. Computational Linguistics, 19(1):75–102.
  6. Anne Göhring and Martin Volk. 2011. Le corpus Text+Berg Une ressource parallèle alpin français-allemand (The Text+Berg Corpus An Alpine French-German Parallel Resource). In Actes de la 18e conférence sur le Traitement Automatique des Langues Naturelles. Articles courts, pages 97–102, Montpellier, Fr… find
  7. Johannes Graën, Igor Mustac, Nikolina Rajovic, Jonathan Schaber, Gerold Schneider, and Noah Bubenhofer. 2023. Swissdox@LiRI. A large database of media articles made accessible to researchers. In Krister Linden, Jyrki Niemi, and Tha-lassiaEditors Kontino, editors, CLARIN Annual Conference Proceedings… find
  8. Gavin Greif, Niclas Griesshaber, and Robin Greif. 2025. Multimodal LLMs for OCR, OCR Post-Correction, and Named Entity Recognition in Historical Documents. Preprint, arXiv:2504.00414. — [arXiv:2504.00414](https://arxiv.org/abs/2504.00414) find
  9. Juri Grosjean and Jannis Vamvas. 2024. Fine-tuning the SwissBERT Encoder Model for Embedding Sentences and Documents. In Proceedings of the 9th edition of the Swiss Text Analytics Conference, pages 41–49, Chur, Switzerland. Association for Computational Linguistics. find
  10. Vishwambhara Hebbalalu. 2026. Dual-Stage OCR Correction for Classical Languages using LLMs: A Comparative Evaluation. Preprint, ResearchGate. find
  11. Oana Ignat, Jean Maillard, Vishrav Chaudhary, and Francisco Guzmán. 2022. OCR Improves Machine Translation for Low-Resource Languages. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1164–1174, Dublin, Ireland. Association for Computational Linguistics. find
  12. Jenna Kanerva, Cassandra Ledins, Siiri Käpyaho, and Filip Ginter. 2025. OCR Error Post-Correction with LLMs in Historical Documents: No Free Lunches. In Proceedings of the Third Workshop on Resources and Representations for Under-Resourced Languages and Domains (RESOURCEFUL-2025), pages 38–47, Talli… find
  13. Julia Krasselt, Philipp Dressen, Matthias Fluor, Cerstin Mahlow, Klaus Rothenhäusler, and Maren Runte. 2020. Swiss-AL: A Multilingual Swiss Web Corpus for Applied Linguistics. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 4145–4151, Marseille, France. European Lan… find
  14. Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. 2023. Trocr: transformer-based optical character recognition with pre-trained models. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Confe… find
  15. Francesco Maria Molfese, Andrei Stefan Bejgu, Simone Tedeschi, Simone Conia, and Roberto Navigli. 2024. CroCoAlign: A Cross-Lingual, Context-Aware and Fully-Neural Sentence Alignment System for Long Texts. In Proceedings of the 18th Conference of the European Chapter of the Association for Computati… find
  16. Joel Niklaus, Jakob Merane, Luka Nenadic, Sina Ahmadi, Yingqiang Gao, Cyrill A. H. Chevalley, Claude Humbel, Christophe Gösken, Lorenzo Tanzi, Thomas Lüthi, Stefan Palombo, Spencer Poff, Boling Yang, Nan Wu, Matthew Guillod, Robin Mamié, Daniel Brunner, Julio Pereyra, and Niko Grupen. 2025. SwiLTra-… find
  17. Guilherme Penedo, Hynek Kydlíček, Vinko Sabolčec, Bettina Messmer, Negar Foroutan, Amir Hossein Kargaran, Colin Raffel, Martin Jaggi, Leandro Von Werra, and Thomas Wolf. 2025. FineWeb2: One Pipeline to Scale Them All — Adapting Pre-Training Data Processing to Every Language. In Second Conference o…
  18. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,…
  19. Yves Scherrer and Bruno Cartoni. 2012. The Trilingual ALLEGRA Corpus: Presentation and Possible Use for Lexicon Induction. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2890–2896, Istanbul, Turkey. European Language Resources Association… find
  20. Rico Sennrich and Martin Volk. 2010. MT-based Sentence Alignment for OCR-generated Parallel Texts. find
  21. In Proceedings of the 9th Conference of the Association for Machine Translation in the Americas: Research Papers, Denver, Colorado, USA. Association for Machine Translation in the Americas. find
  22. Rico Sennrich and Martin Volk. 2011. Iterative, MT-based Sentence Alignment of Parallel Texts. In Proceedings of the 18th Nordic Conference of Computational Linguistics (NODALIDA 2011), pages 175– 182, Riga, Latvia. Northern European Association for Language Technology (NEALT). find
  23. Raymond W. Smith. 2007. An Overview of the Tesseract OCR Engine. Ninth International Conference on Document Analysis and Recognition (ICDAR 2007), 2:629–633. find
  24. Brian Thompson and Philipp Koehn. 2019. Vecalign: Improved Sentence Alignment in Linear Time and Space. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1342–1348,…

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/stussi-2026-how-good-is-ai-swiss/