---
language:
- en
license: cc-by-nc-nd-4.0
pretty_name: PoliticalDebatesCorpus-EN — English Political Debates Corpus
size_categories:
- 10B<n<100B
task_categories:
- text-generation
- fill-mask
- text-classification
tags:
- political-discourse
- parliamentary-debates
- argument-mining
- stance-detection
- domain-specific
- pretraining
---

# RooseDebates: English Political Debates Corpus

PoliticalDebatesCorpus-EN is a large-scale English corpus of political debate and speech transcripts,
assembled for domain-specific language model pre-training. It is the training corpus behind
**RooseBERT**, a language model tailored to the dialogic and argumentative nature of political
debates. The corpus brings together televised presidential debates and parliamentary debates
from multiple countries and international bodies, totalling **~11 GB** of text and spanning the
period **1946–2025**.

## Dataset Summary

- **Total size:** ~11 GB of cleaned English text
- **Time span:** 1946–2025
- **Domain:** Political debates, parliamentary sessions, presidential debates, UN debates
- **Languages:** English
- **Source types:** Televised exchanges, parliamentary debates, primary/general election debates
- **Intended use:** Pre-training and fine-tuning of language models for political discourse analysis

All sources originate from authoritative speakers and official political settings, where
political decisions are discussed, contested, and communicated to the public.

## Supported Tasks

The corpus was designed to support domain-adaptive pre-training for downstream tasks including:

- **Stance detection** (support / oppose toward a target)
- **Sentiment analysis** of debate speeches
- **Argument component detection and classification** (claims vs. premises)
- **Argument relation prediction and classification** (support / attack / no-relation)
- **Motion policy classification**
- **Named entity recognition (NER)**

## Data Sources

The corpus combines existing curated collections with transcripts scraped from official debate
websites. Each source was pre-processed to remove hyperlinks and markup tags and to collapse
multiple spaces.

| Source | Coverage | Period | Size |
|---|---|---|---|
| African Parliamentary Debates (HOME Project) | Ghana & South Africa | 1999–2024 | 573 MB |
| Australian Parliamentary Debates | Australia | 1998–2025 | 1 GB |
| Canadian Parliamentary Debates (OpenParliament) | Canada | 1994–2025 | 1.1 GB |
| European Parliamentary Debates (EUSpeech) | EU leaders & institutions | 2007–2015 | 110 MB |
| Irish Parliamentary Debates (ParlEE + Dáil Éireann DB) | Ireland | 1919–2019 | ~3.4 GB |
| New Zealand Parliamentary Debates (ParlSpeech) | New Zealand | 1987–2019 | 791 MB |
| Scottish Parliamentary Debates (ParlScot) | Scotland | until 2021 | 443 MB |
| UK Parliamentary Debates (House of Commons) | United Kingdom | 1979–2019 | 2.6 GB |
| UN General Debate Corpus (UNGDC) | UN General Assembly | 1946–2023 | 186 MB |
| UN Security Council Debates (UNSC) | UN Security Council | 1992–2023 | 387 MB |
| United States Debates (American Presidency Project) | United States | 1960–2024 | 16 MB |

## Pre-processing

Each dataset was cleaned with a uniform pipeline:

1. Removal of hyperlinks and markup tags
2. Collapsing of multiple consecutive whitespace characters
3. Merging of per-source splits into a unified corpus with balanced representation across sources

## Intended Uses & Limitations

**Intended uses.** Pre-training or continued pre-training of encoder-based language models for
political discourse analysis; fine-tuning on stance, sentiment, argument mining, and policy
classification tasks.

**Limitations.**

- The corpus is **English-only** and skewed toward Commonwealth and U.S./UN sources, so
  geographic and linguistic coverage is uneven.
- Transcripts reflect the political opinions and rhetoric of the speakers; the corpus is **not**
  a neutral source and may contain biased, adversarial, or offensive language.
- Time coverage varies by source, which may introduce period-specific framing effects.
- Models trained on this data should not be treated as authoritative on factual political claims.

## Licensing

This corpus aggregates multiple underlying datasets, each released under its own terms by its
original authors. **Users must review and comply with the license of each constituent source**
before redistribution or commercial use. The aggregation and accompanying card are released
under CC-BY-NC-ND-4.0; this does not override the licenses of the individual sources.

## Citation

If you use this dataset, please cite us:

```bibtex
@article{dore2025roosebert,
  title={RooseBERT: A New Deal For Political Language Modelling},
  author={Dore, Deborah and Cabrio, Elena and Villata, Serena},
  journal={arXiv preprint arXiv:2508.03250},
  year={2025}
}
```

Please also cite the original sources of the constituent datasets where appropriate (see the
**Data Sources** table and the references in the paper).

## Contact

Deborah Dore - deborah.dore@cnrs.fr