drelhaj/marian-finetuned-mcwc-en-to-es

Marian MT fine-tuned on the Multilingual Corpus of World’s Constitutions (MCWC)

This model is a fine-tuned version of Helsinki-NLP/opus-mt-en-es, adapted using high-quality sentence-aligned constitutional text from the Multilingual Corpus of World’s Constitutions (MCWC)
📄 MCWC paper (OSACT 2024): https://aclanthology.org/2024.osact-1.7/

This variant handles: English → Spanish translation.

---

Overview

The MCWC provides a curated multilingual collection of constitutional texts from countries across the world. The corpus emphasises data cleanliness, high-quality sentence alignment, and detailed metadata (including country and continent mappings). It supports research in:

• legal and constitutional NLP
• comparative constitutional studies
• multilingual machine translation
• cross-lingual semantic analysis

This model was fine-tuned on the English–Spanish segment of the MCWC, enabling translation that is more attuned to legal and constitutional language than general-purpose MT systems.

---

Intended use

This model is suitable for tasks such as:

• translating constitutional or legal documents
• cross-lingual legal text comparison
• multilingual information extraction
• downstream legal NLP tasks requiring domain-specific MT

It is not intended for casual or conversational translation, as it is optimised for formal and legal text.

---

Training data

The model was trained on the MCWC’s English–Spanish aligned sentence pairs.
The MCWC dataset includes:

• cleaned constitutional text
• high-quality sentence segmentation
• pairwise alignments
• country and regional metadata

More details may be found in the accompanying paper:

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

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• optimizer: {'name': 'AdamWeightDecay', 'learning_rate': {'module': 'keras.optimizers.schedules', 'class_name': 'PolynomialDecay', 'config': {'initial_learning_rate': 5e-05, 'decay_steps': 3666, 'end_learning_rate': 0.0, 'power': 1.0, 'cycle': False, 'name': None}, 'registered_name': None}, 'decay': 0.0, 'beta_1': 0.9, 'beta_2': 0.999, 'epsilon': 1e-08, 'amsgrad': False, 'weight_decay_rate': 0.01}
• training_precision: mixed_float16

Training results

Train Loss	Validation Loss	Epoch
1.1192	1.0032	0
0.9233	0.9729	1
0.8292	0.9642	2

Framework versions

• Transformers 4.33.3
• TensorFlow 2.13.0
• Datasets 2.14.5
• Tokenizers 0.13.3