README.md

---
license: cc-by-nc-sa-4.0
source_datasets:
- coastalcph/eu_debates
language_creators:
- found
multilinguality:
- multilingual
language:
- bg
- cs
- da
- de
- el
- en
- es
- et
- fi
- fr
- hr
- hu
- it
- lt
- lv
- mt
- nl
- pl
- pt
- ro
- sk
- sl
- sv
tags:
- politics
size_categories:
- 10K<n<100K
pretty_name: EU Debates (JSONL Conversion)
---

# Dataset Description

This dataset is a **conversion of the original [`coastalcph/eu_debates`](https://huggingface.co/datasets/coastalcph/eu_debates)** dataset released by [Chalkidis and Brandl (2024)](https://arxiv.org/abs/2403.13592).

The goal of this repository is to provide the same underlying data **without a Python loading script**, in a standard format (JSON Lines / Parquet) compatible with the current Hugging Face `datasets` library and automated data loading.

The original EU Debates corpus consists of approx. 87k individual speeches in the period 2009–2023.
The data was exhaustively scraped from the official European Parliament Plenary website ([link](https://www.europarl.europa.eu/)). All speeches are time-stamped, thematically organized in debates, and include metadata about:

- the speaker's identity (full name, euro-party affiliation, speaker role),
- the debate (date and title),
- language information, and (where available) machine-translated versions in English.

Older debate speeches are originally in English, while newer ones are linguistically diverse across the 23 official EU languages. Machine-translated English versions are provided using the EasyNMT framework with the [M2M-100 (418M)](https://huggingface.co/facebook/m2m100_418M) model (Fan et al., 2020).

This repository only changes the **storage format** (to `train.jsonl` / Parquet) and **removes the Python loading script**.
The data contents and fields are preserved from the original dataset.

# Data Fields

Each row / JSONL line is a single speech with the following fields:

- `speaker_name`: `string`, full name of the speaker.
- `speaker_party`: `string`, name of the euro-party (group) that the MEP is affiliated with.
- `speaker_role`: `string`, role of the speaker (e.g., Member of the European Parliament (MEP), EUROPARL President).
- `debate_title`: `string`, title of the debate in the European Parliament.
- `date`: `string`, full date of the speech in `YYYY-MM-DD` format.
- `year`: `string`, year of the speech in `YYYY` format.
- `intervention_language`: `string`, language code of the original intervention.
- `original_language`: `string`, language code of the original text.
- `text`: `string`, full original speech of the speaker.
- `translated_text`: `string` or `null`, machine translation of the speech into English if the original is not English, otherwise `null`.

# Data Instances

Example of a data instance:

```json
{
  "speaker_name": "Michèle Striffler",
  "speaker_party": "PPE",
  "speaker_role": "MEP",
  "debate_title": "Famine in East Africa (debate)",
  "date": "2011-09-15",
  "year": "2011",
  "intervention_language": "fr",
  "original_language": "fr",
  "text": "Monsieur le Président, Madame le Commissaire, chers collègues, la situation humanitaire sans précédent que connaît la Corne de l'Afrique continue [...]",
  "translated_text": "Mr. President, Mr. Commissioner, dear colleagues, the unprecedented humanitarian situation of the Horn of Africa continues [...]"
}
```

# How to Use

### From the Hugging Face Hub

If the dataset is hosted under `RJuro/eu_debates`:

```python
from datasets import load_dataset

eu_debates = load_dataset("RJuro/eu_debates", split="train")
```

### From Local Files

If you downloaded the `train.jsonl` file locally:

```python
from datasets import load_dataset

eu_debates = load_dataset(
    "json",
    data_files={"train": "train.jsonl"},
    split="train",
)
```

If you use Parquet instead:

```python
from datasets import load_dataset

eu_debates = load_dataset(
    "parquet",
    data_files={"train": "train.parquet"},
    split="train",
)
```

# Dataset Statistics

The statistics below are inherited from the original `coastalcph/eu_debates` dataset.

### Distribution of speeches across euro-parties:

| Euro-party  | No. of Speeches |
|-------------|-----------------|
| EPP         | 25,455 (29%)    |
| S&D         | 20,042 (23%)    |
| ALDE        | 8,946 (10%)     |
| ECR         | 7,493 (9%)      |
| ID          | 6,970 (8%)      |
| GUE/NGL     | 6,780 (8%)      |
| Greens/EFA  | 6,398 (7%)      |
| NI          | 5,127 (6%)      |
| **Total**   | **87,221**      |

### Distribution of speeches across years and euro-parties:

| Year | EPP | S&D | ALDE | ECR | ID | GUE/NGL | Greens/EFA | NI | Total |
|---|---|---|---|---|---|---|---|---|---|
| 2009 | 748 | 456 | 180 | 138 | 72 | 174 | 113 | 163 | **2044** |
| 2010 | 3205 | 1623 | 616 | 340 | 341 | 529 | 427 | 546 | **7627** |
| 2011 | 4479 | 2509 | 817 | 418 | 761 | 792 | 490 | 614 | **10880** |
| 2012 | 3366 | 1892 | 583 | 419 | 560 | 486 | 351 | 347 | **8004** |
| 2013 | 724 | 636 | 240 | 175 | 152 | 155 | 170 | 154 | **2406** |
| 2014 | 578 | 555 | 184 | 180 | 131 | 160 | 144 | 180 | **2112** |
| 2015 | 978 | 1029 | 337 | 405 | 398 | 325 | 246 | 240 | **3958** |
| 2016 | 919 | 972 | 309 | 387 | 457 | 317 | 225 | 151 | **3737** |
| 2017 | 649 | 766 | 181 | 288 | 321 | 229 | 162 | 135 | **2731** |
| 2018 | 554 | 611 | 161 | 242 | 248 | 175 | 160 | 133 | **2284** |
| 2019 | 1296 | 1339 | 719 | 556 | 513 | 463 | 490 | 353 | **5729** |
| 2020 | 1660 | 1564 | 823 | 828 | 661 | 526 | 604 | 346 | **7012** |
| 2021 | 2147 | 2189 | 1290 | 1062 | 909 | 708 | 990 | 625 | **9920** |
| 2022 | 2436 | 2273 | 1466 | 1177 | 827 | 962 | 1031 | 641 | **10813** |
| 2023 | 1716 | 1628 | 1040 | 878 | 619 | 779 | 795 | 499 | **7954** |

### Distribution of speeches across the 23 EU official languages:

| Language | No. of Speeches |
|----------|-----------------|
| en       | 40,736 (46.7%)  |
| de       | 6,497 (7.5%)    |
| fr       | 6,024 (6.9%)    |
| es       | 5,172 (5.9%)    |
| it       | 4,506 (5.2%)    |
| pl       | 3,792 (4.4%)    |
| pt       | 2,713 (3.1%)    |
| ro       | 2,308 (2.7%)    |
| el       | 2,290 (2.6%)    |
| nl       | 2,286 (2.6%)    |
| hu       | 1,661 (1.9%)    |
| hr       | 1,509 (1.7%)    |
| cs       | 1,428 (1.6%)    |
| sv       | 1,210 (1.4%)    |
| bg       | 928 (1.1%)      |
| sk       | 916 (1.1%)      |
| sl       | 753 (0.9%)      |
| fi       | 693 (0.8%)      |
| lt       | 618 (0.7%)      |
| da       | 578 (0.7%)      |
| et       | 342 (0.4%)      |
| lv       | 184 (0.2%)      |
| mt       | 0 (0.0%)        |

# Citation Information

If you use this dataset, please cite the original work:

> Llama meets EU: Investigating the European political spectrum through the lens of LLMs.
> Ilias Chalkidis and Stephanie Brandl.
> In the Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL),
> Mexico City, Mexico, June 16–21, 2024.

```bibtex
@inproceedings{chalkidis-and-brandl-eu-llama-2024,
    title = "Llama meets EU: Investigating the European political spectrum through the lens of LLMs",
    author = "Chalkidis, Ilias  and Brandl, Stephanie",
    booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
}
```

This repository only provides a format-converted, script-free version of the original dataset; all credit for data collection and annotation goes to the original authors.