README.md

metadata

license: cc-by-nc-4.0
tags:
  - legal-nlp
  - multilingual
  - constitutions
  - machine-translation
  - comparative-law
  - mcwc
dataset-name: MCWC

Multilingual Corpus of World’s Constitutions (MCWC)

The MCWC is a curated multilingual corpus of constitutional texts from 191 countries, including both current and historical versions. The dataset provides aligned constitutional content in English, Arabic, and Spanish, enabling comparative legal analysis and multilingual NLP research.

This CSV version is a cleaned, structured, and sentence-aligned representation of the corpus, suitable for machine translation, information retrieval, and a wide range of legal NLP tasks.

📄 Paper (OSACT @ LREC-COLING 2024):
https://aclanthology.org/2024.osact-1.7/

---

Contents of the dataset

MCWC.csv contains:

• Constitutional text in three languages: English, Arabic, Spanish
• Pairwise sentence alignments using a stable AlignID
• Country and continent metadata
• Cleaned and normalised text for direct use in NLP pipelines
• Translations for constitutions originally missing in Arabic or Spanish, generated using a fine-tuned NMT model
• One row per sentence, enabling easy filtering and cross-lingual matching

Typical columns include:

• country
• continent
• align_id
• text_en
• text_ar
• text_es
• constitution_year
• language_source (original vs. machine-translated)

---

Motivation

Constitutions form the legal and philosophical foundation of nation-states. However, constitutional research is often limited by:

• fragmented access to multilingual versions
• inconsistent formatting and metadata
• lack of high-quality sentence alignment
• scarcity of constitutional text in Arabic and Spanish

The MCWC addresses these gaps by offering a unified and multilingual corpus built for computational analysis, comparative constitutional research, and machine translation development.

The acronym MCWC is pronounced “Makkuk”, an Arabic word for space shuttle, chosen to symbolise the corpus’s role in enabling cross-lingual travel between legal systems.

---

Data sources and preparation

The corpus integrates material from:

• the Comparative Constitutions Project
• the Constitute Project
• Wikipedia and various governmental archives

Raw texts—primarily in XML—were normalised, cleaned, and sentence-segmented. For constitutions lacking Arabic or Spanish versions, we generated translations using a fine-tuned state-of-the-art MT model.

A custom parser aligned sentences across languages using structural cues (article numbers, section labels, etc.), with each sentence assigned a persistent AlignID for alignment across all three languages.

To support comparative analysis, we applied:

• continent classification via gazetteer matching
• tokenisation and normalisation
• vocabulary overlap statistics
• TF-IDF cosine-similarity analysis (English)

The paper provides further detail, including heatmaps of by-continent overlap and cross-lingual similarity.

---

Dataset statistics

• 223 constitutions
• 191 countries
• 95 constitutions available in all three languages
• 52,177 aligned English–Arabic sentences
• 48,892 aligned English–Spanish sentences
• 27,352 aligned Arabic–Spanish sentences
• 236,156 parallel machine-generated sentences (SeamlessM4T-v2)

Across continents:

• Africa: 65 constitutions
• Asia: 54
• Europe: 49
• North America: 26
• South America: 15
• Oceania: 14

Average lengths and TTR values are reported in the paper.

---

Intended uses

The MCWC is designed to support:

• multilingual machine translation
• legal text analysis and retrieval
• constitutional comparison and modelling
• multilingual topic modelling
• cross-lingual semantic similarity
• legal corpus linguistics
• training/testing MT systems for legal text

Users may also employ the dataset to study variation in legal language across jurisdictions and historical periods.

---

Translation and evaluation

For missing Arabic/Spanish texts, English constitutions were translated using Facebook’s Seamless-m4t-v2-large.

Evaluation included:

• BLEU = 0.68 on 500 manually inspected En–Ar and En–Es pairs
• Human annotation by two Arabic NLP experts
  • Cohen’s κ = 0.30 (due to label imbalance)
  • Krippendorff’s α ≈ 0.90 (robust reliability)

Fine-tuned Marian NMT models trained on the MCWC consistently improved translation quality across all six language pairs. These models are available on HuggingFace.

---

Train / Validation / Test Splits

The original MCWC paper does not define an official train–validation–test split.
To support reproducibility and facilitate machine-learning experiments, we provide a recommended 80/10/10 split using a fixed random seed. This ensures consistent partitioning whilst keeping the dataset fully shuffled.

The split is applied over the complete CSV file (MCWC.csv) and does not use stratification, as constitutional texts vary considerably across countries and languages.

import pandas as pd
from sklearn.model_selection import train_test_split

# Load the full corpus
df = pd.read_csv("MCWC.csv")

# First split: train vs temporary (validation + test)
train_df, temp_df = train_test_split(
    df,
    test_size=0.2,     # 20% goes to val+test
    random_state=42,
    shuffle=True
)

# Second split: create separate validation and test sets
val_df, test_df = train_test_split(
    temp_df,
    test_size=0.5,     # half of 20% → 10% each
    random_state=42,
    shuffle=True
)

# Save the splits
train_df.to_csv("MCWC_train.csv", index=False)
val_df.to_csv("MCWC_val.csv", index=False)
test_df.to_csv("MCWC_test.csv", index=False)

print("Train:", len(train_df))
print("Validation:", len(val_df))
print("Test:", len(test_df))

---

Ethical considerations and limitations

• Source texts originate from open datasets (CCP, Constitute Project) and governmental archives.
• This CSV is a processed and aligned derivative, not a redistribution of the original XML files.
• Translation quality varies depending on source material and MT system limitations.
• Some constitutions contain archaic, historical, or culturally-specific legal phrasing that may not translate literally.
• Differences in translation providers (HeinOnline, IDEA, OUP, etc.) may introduce stylistic or interpretative variance.

Users requiring original XML files should retrieve them directly from the Constitute Project.

---

Citation

If you use this dataset, please cite:

El-Haj, M. & Ezzini, S. (2024).
The Multilingual Corpus of World’s Constitutions (MCWC).
OSACT Workshop @ LREC-COLING 2024.
https://aclanthology.org/2024.osact-1.7/

---

Contact

For questions or collaboration:

Mo El-Haj
UCREL NLP Group
Lancaster University & VinUniversity
Email: m.el-haj@lancaster.ac.uk / elhaj.m@vinuni.edu.vn