Elamur Research⌕
☰
home / papers / 2026 / From Concordance to Inference: ParlaCAP Helps ParlaMint Escape the Linguistics Lab

From Concordance to Inference: ParlaCAP Helps ParlaMint Escape the Linguistics Lab

Paper · 2026 · Nikola Ljubešić

Summary

ParlaMint offers comparable, metadata-rich transcripts of European parliamentary debates. However, it has mostly served corpus linguists rather than political scientists or sociologists. This talk paper presents ParlaCAP, a project funded by an OSCARS Open Science cascading grant. It turns ParlaMint into an analysis-ready resource for the wider SSH community. The collection has over 8 million speeches from 28 parliaments in more than 20 languages, which is too large for manual coding. The project therefore labels every speech automatically with two XLM-RoBERTa-based classifiers. XLM-R-ParlaSent handles sentiment, and XLM-R-ParlaCAP assigns policy topics using the Comparative Agendas Project (CAP) scheme. The topic model relies on teacher–student training: GPT-4o generated the labels, and a lighter XLM-R model learned from them so it can run efficiently at scale. Both classifiers perform at roughly the level of human annotators. The resulting FAIR dataset is distributed through the CROSSDA repository and an API with flexible filtering. Example use cases show that cross-country comparisons of policy attention and sentiment take only a few lines of Python.

Key points

  • Scale: Sentiment and CAP policy-topic labels are added to more than 8 million ParlaMint speeches, which span 28 European parliaments and upwards of 20 languages.
  • Models: Two models built on XLM-RoBERTa do the enrichment, XLM-R-ParlaSent for sentiment and XLM-R-ParlaCAP for agenda/topic classification. Both are reported to match human annotators.
  • Training approach: XLM-R-ParlaCAP was trained with a teacher–student setup. GPT-4o supplied the training labels, and the XLM-R student offers scalable deployment.
  • Access: The data is available through CROSSDA, the Croatian CESSDA repository (DOI 10.23669/1ZTELP), and an API at parlacap.ipipan.waw.pl. Users can filter by country, time, speaker attributes, agenda category and sentiment.
  • Purpose: The project links linguistic resources with political-science methodology. It supports cross-national and longitudinal research on how policy, sentiment and political identity interact.

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

 Institutions

🔬Jožef Stefan InstituteLjubljana · Slovenia17 works · 2,165 citations🎓University of LjubljanaLjubljana · Slovenia13 works · 38,693 citations🔬Institute of Contemporary HistoryLjubljana · Slovenia4 works · 32 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksSocial Sciences · 1,772 worksGeneral Social Sciences · 227 works

Cite

Cite as Nikola Ljubešić. 2026. From Concordance to Inference: ParlaCAP Helps ParlaMint Escape the Linguistics Lab. Proceedings of the ParlaCLARIN V Workshop on Interoperability, Multilinguality, and Multimodality in Parliamentary Corpora.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4adx5jcvvjei

@inproceedings{ljubesic2026concordance,
  title = {{From Concordance to Inference: ParlaCAP Helps ParlaMint Escape the Linguistics Lab}},
  author = {Nikola Ljubešić},
  booktitle = {Proceedings of the ParlaCLARIN V Workshop on Interoperability, Multilinguality, and Multimodality in Parliamentary Corpora},
  year = {2026},
  doi = {10.63317/4adx5jcvvjei},
  url = {https://doi.org/10.63317/4adx5jcvvjei}
}

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ljubesic-2026-concordance-inference-parlacap-helps-parlamint/