🌐 nlp-low-resource-mt-aug

Back-Translation and Data Augmentation for Low-Resource Machine Translation
English → Malayalam Neural Machine Translation using mBART-50, NLLB-200, Helsinki-NLP (Opus-MT), and mT5-Base

---

📌 Overview

Neural Machine Translation (NMT) has advanced significantly for high-resource languages, but low-resource languages like Malayalam still suffer from a critical lack of high-quality parallel corpora. This project tackles that challenge head-on.

We evaluate four NMT architectures on a constrained baseline of 8,000 English–Malayalam sentence pairs, then apply a Back-Translation (BT) based augmentation pipeline to generate 10,000 additional synthetic parallel sentences, significantly improving translation quality across all models.

---

🧠 Models Evaluated

Model	Architecture	Parameters	Key Trait
Helsinki-NLP/opus-mt-en-ml	MarianMT Transformer	~76M	Lightweight, translation-optimized
facebook/mbart-large-50	Seq2Seq Denoising Autoencoder	600M+	Strong morphological understanding
facebook/nllb-200-distilled-600M	Mixture-of-Experts Transformer	600M	State-of-the-art zero-shot for LRLs
google/mt5-base	Text-to-Text Transfer Transformer	580M	Generic generation, flexible

---

📊 Results

BLEU Score Comparison (200-sample test set)

Model	Baseline BLEU	Augmented BLEU	Gain (ΔBLEU)
mT5-Base	1.03	2.15	+1.12
mBART-50	1.44	3.80	+2.36
NLLB-200	2.86	5.12	+2.26
Helsinki-NLP	4.40	6.85	+2.45

✅ Helsinki-NLP achieved the best performance despite having the fewest parameters — demonstrating that task-specific optimization beats over-parameterization in low-resource settings.

---

⚙️ Methodology

Back-Translation Pipeline

Monolingual Malayalam Data (ML)
        ↓
   Back-Translation Model
        ↓
  Generate Synthetic English (EN*)
        ↓
  Synthetic Pairs (EN* ↔ ML)
        ↓
  Fine-tune EN → ML model

1. Baseline Training — All four models fine-tuned on 8,000 EN–ML sentence pairs
2. Back-Translation — Malayalam monolingual data fed through mBART/NLLB to generate synthetic English
3. Augmented Training — Models retrained on 18,000 pairs (8k original + 10k synthetic)
4. Evaluation — BLEU and chrF scores on held-out test set

---

📁 Project Structure

nlp-low-resource-mt-aug/
│
├── google_mt5_base.ipynb          # mT5-Base fine-tuning notebook
├── mbart.ipynb                    # mBART-50 + back-translation notebook
├── NLLB_200.ipynb                 # NLLB-200 fine-tuning notebook
├── Helsinki_NLP_opus_mt_en_ml.ipynb  # MarianMT fine-tuning notebook
│
├── reports/
│   ├── mt5_enml_report.txt
│   └── ...
│
└── README.md

---

🗃️ Dataset

• Source: English to Malayalam Machine Translation Dataset by P. Modi (Kaggle, 2021)
• Baseline corpus: 8,000 EN–ML sentence pairs
• After augmentation: ~18,000 pairs (10,000 synthetic via Back-Translation)
• Split: Train / Eval / Test

---

🔍 Key Findings

• mT5-Base suffered catastrophic hallucination at baseline (BLEU 1.03), repeating tokens like "വിശ്വാസം" over 30 times
• Helsinki-NLP achieved the highest semantic accuracy even at baseline, correctly rendering complex phrases
• mBART-50 showed mixed-script outputs (Malayalam + Roman) before augmentation
• All models improved meaningfully after BT-based augmentation, with the augmented dataset converging faster and reaching lower validation loss

---

🚀 Getting Started

Prerequisites

pip install transformers datasets sacrebleu kagglehub sentencepiece accelerate

Run a Notebook

Open any of the .ipynb files in Google Colab (GPU recommended):

• Helsinki_NLP_opus_mt_en_ml.ipynb — Best starting point (lightest model)
• mbart.ipynb — Includes the full back-translation pipeline
• NLLB_200.ipynb — Best synthetic data quality
• google_mt5_base.ipynb — Generic seq2seq baseline

---

📐 Evaluation Metrics

• BLEU (Bilingual Evaluation Understudy) — n-gram precision with brevity penalty
• chrF (Character n-gram F-score) — Better suited for morphologically rich languages like Malayalam

---

👥 Authors

Name	Institution
Raj Narayanan	Lovely Professional University, Phagwara
Enjula Uchoi	Lovely Professional University, Phagwara
Kanika Debbarma	National Institute of Technology, Arunachal Pradesh
Avnish Thakur	Lovely Professional University, Phagwara

---

📄 Citation

If you use this work, please cite:

@article{narayanan2024backtranslation,
  title     = {Back-Translation and Data Augmentation for Low-Resource Machine Translation},
  author    = {Raj Narayanan and Enjula Uchoi and Kanika Debbarma and Avnish Thakur},
  year      = {2024},
  note      = {Lovely Professional University / NIT Arunachal Pradesh}
}

---

📚 References

• Fadaee et al., "Data Augmentation for Low-Resource NMT," ACL 2017
• Kimera et al., "Data Augmentation With Back-Translation for Low Resource Languages," NLPIR 2024
• Costa-jussà et al., "No Language Left Behind," Meta AI 2022
• Tiedemann & Thottingal, "OPUS-MT," EAMT 2020
• Dabre et al., "IndicTrans," ACL 2022

---

💡 This project contributes to the broader goal of making machine translation accessible to underrepresented language communities.