Neural Machine Translation System: English-to-Russian

💖 Support My GITHUB Open Projects

Complete implementation of English-to-Russian neural machine translation system using attention mechanisms and transformer architectures.

Overview

This project develops and evaluates a neural machine translation (NMT) system for English-to-Russian translation using modern deep learning approaches. The system fine-tunes the Helsinki-NLP/opus-mt-en-ru pre-trained model on the opusbooks dataset.

Features

• Transformer-based architecture with attention mechanisms
• Fine-tuned Helsinki-NLP/opus-mt-en-ru model
• Training on 10,000+ sentence pairs from opusbooks
• Evaluation using multiple metrics (BLEU, chrF, TER)
• Handling of Russian morphological complexity
• Integration with Hugging Face Transformers library

Performance Metrics

• SacreBLEU Score: 22.6
• chrF Score: 48.5
• Training: Kaggle T4 GPU environment
• Dataset: OpenParallel opusbooks (EN-RU)

Installation

pip install torch transformers datasets evaluate

Dataset • Source: OpenParallel opusbooks corpus • Language Pair: English → Russian • Training Samples: 10,000+ • Format: JSON with ‘src’ and ‘tgt’ fields Architecture • Model: Sequence-to-Sequence with Transformer encoder-decoder • Encoder: Multi-head self-attention (12 heads) • Decoder: Masked multi-head attention with cross-attention • Vocabulary: BPE tokenization (32k tokens) Related Publications • Development and Evaluation of an English-to-Russian Neural Machine Translation System • Developpement et Evaluation d’un Systeme de Traduction Automatique Neuronale • Systeme de traduction automatique neuronal du russe vers l’anglais (Conference)

Hugging Face Model Card

See: https://huggingface.co/Helsinki-NLP/opus-mt-en-ru

License

MIT License - See LICENSE file for details

Author

Dominique S. Loyer

• ORCID: https://orcid.org/0009-0003-9713-7109
• GitHub: https://github.com/DominiqueLoyer
• Affiliation: Universite du Quebec a Montreal (UQAM)

Challenges & Solutions

Russian Morphological Complexity

• Solution: Fine-tuning on morphologically rich dataset
• Pre-trained model handles inflectional suffixes
• Subword tokenization captures Russian morphology

Data Quality

• Solution: Filtering of low-quality alignments
• Validation on in-domain test set
• Manual review of translation outputs

Future Work

• Multilingual support (French, German)
• Bidirectional translation (RU-EN)
• Domain-specific fine-tuning
• Integration with back-translation for data augmentation
• Evaluation on news domain benchmark

Contributing

Contributions welcome! Areas of interest:

• Back-translation data augmentation
• Domain adaptation experiments
• Evaluation on standard benchmarks

---

Last Updated: January 6, 2026 Status: Ready for production use

---

Bleu Score

Bleu Score