README.md

---
# Dataset Card for EuroParlVote

## Dataset Details

### Dataset Description

EuroParlVote links **European Parliament debate speeches** to **roll-call votes** and **Member of European Parliament (MEP) demographic data** across up to 24 official EU languages.  
It supports two primary benchmark tasks:  
1. **Gender Classification**: Predict the MEP’s gender from a debate speech.  
2. **Vote Prediction**: Predict a FOR/AGAINST vote from the topic and speech (optionally with demographic context).

This dataset is based on publicly available EP plenary records and enhanced with metadata scraped from Wikipedia and other official sources.

- **Curated by:** Jinrui Yang, Xudong Han, Timothy Baldwin  
- **Funded by:** Melbourne Research Scholarship; LIEF HPC-GPGPU Facility (LE170100200)  
- **Shared by:** University of Melbourne   
- **Language(s) (NLP):** Up to 24 official EU languages (Language column)  
- **License:** CC BY-NC 4.0 (to confirm)

### Dataset Sources

- **Repository:** [https://huggingface.co/datasets/unimelb-nlp/EuroParlVote](https://huggingface.co/datasets/unimelb-nlp/EuroParlVote)  
- **Paper:** _Demographics and Democracy: Benchmarking LLMs’ Gender Bias and Political Leaning in European Parliament_ (EMNLP 2024, camera-ready)  
- **Demo:** [[EuroParlVote website](https://parlvote-demo.vercel.app/)] 

## Uses

### Direct Use

- Benchmark LLM fairness/bias in political discourse.
- Multilingual political text classification and vote prediction.
- Study demographic effects (gender, group) on model behavior.

### Out-of-Scope Use

- Real-time vote forecasting or influencing political processes.
- Targeting individuals or groups.
- Disinformation or harassment.

## Dataset Structure

### File Structure

The dataset is split into **train**, **dev**, and **test** (~8:1:1).  
This example shows the `dev_set.csv` structure.

### Columns

| Column             | Type    | Description |
|--------------------|---------|-------------|
| `Chapter`          | float   | Debate chapter number |
| `Chapter_ID`       | string  | Debate chapter unique identifier |
| `Act_ID`           | string  | Legislative act ID (may be "MISSING") |
| `Report_ID`        | string  | Parliamentary report ID |
| `Debate_ID`        | string  | Unique debate ID + language suffix |
| `Vote_ID`          | int     | Unique roll-call vote ID |
| `Vote_Description` | string  | English description of the vote topic |
| `Vote_Timestamp`   | string  | Date-time of the vote |
| `Language`         | string  | ISO language code of the speech |
| `Speaker`          | string  | Speaker’s full name |
| `MEP_ID`           | int     | Unique MEP identifier |
| `Party`            | string  | Party affiliation (if available) |
| `Role`             | string  | Role in debate (e.g., rapporteur) |
| `CODICT`           | int     | Speaker unique code |
| `Speaker_Type`     | string  | Type of speaker (e.g., MEP, Chair) |
| `Start_Time`       | string  | Start time (uniform in this split) |
| `End_Time`         | string  | End time (uniform in this split) |
| `Title_[XX]`       | string  | Debate title in language `XX` (24 variants, e.g., Title_EN, Title_FR, Title_DE) |
| `Speech`           | string  | Full debate speech text |
| `position`         | string  | Vote label: FOR / AGAINST |
| `country_code_x`   | string  | Country code (original source) |
| `group_code`       | string  | Political group code (8 possible) |
| `first_name`       | string  | MEP first name |
| `last_name`        | string  | MEP last name |
| `country_code_y`   | string  | Country code (from demographic scrape) |
| `date_of_birth`    | string  | Date of birth (YYYY-MM-DD) |
| `email`            | string  | Public MEP email (if available) |
| `facebook`         | string  | Facebook profile URL (if available) |
| `twitter`          | string  | Twitter/X profile URL (if available) |
| `gender`           | string  | Binary label: MALE / FEMALE |

**Note:** Title columns cover all official EU languages; `Speech` is in the original debate language (`Language`).

### Label Distribution (dev split)

- **position**: `FOR` and `AGAINST` are balanced in dev/test.
- **gender**: MALE, FEMALE.

## Dataset Creation

### Curation Rationale

Existing multilingual political datasets rarely link actual speeches to **real-world vote outcomes** and demographics, making fairness and bias studies difficult. This dataset bridges that gap.

### Source Data

- Votes from **HowTheyVote.eu**.
- Debates aligned via vote metadata references.
- Demographics from Wikipedia & official EP sources.

### Processing

- Removed abstentions & missing topic/speech.
- Gender inferred from pronouns and manually checked.

### Annotation

- Gender labels created via semi-automatic heuristics, with manual validation.
- Vote labels come directly from official roll-call data.

### Sensitive Information

- Contains names, countries, political groups of public figures (MEPs).
- Binary gender labels do not reflect all identities.

## Bias, Risks, and Limitations

- Binary gender assumption.
- Political group may not fully capture ideology.
- Translation hurts performance; originals recommended.
- Biases in speeches may reflect political context, not individual ideology.

## Citation

**BibTeX:**
```bibtex
@inproceedings{yang2024europarlvote,
  title={Demographics and Democracy: Benchmarking LLMs’ Gender Bias and Political Leaning in European Parliament},
  author={Yang, Jinrui and Han, Xudong and Baldwin, Timothy},
  booktitle={Proceedings of the 8th International Conference on Natural Language and Speech Processing},
  year={2025}
}