MachineTranslation/Machine-translation/README.md

metadata

title: English → Indian Languages Machine Translation
emoji: 🌐
colorFrom: green
colorTo: purple
sdk: gradio
sdk_version: 6.12.0
app_file: app.py
pinned: false

🌐 English → Indian Languages Machine Translation

A complete machine translation pipeline for English → Indian languages using the NLLB-200 multilingual model, evaluated with multiple automatic metrics and deployed via an interactive Gradio web application.

---

📋 Table of Contents

• Overview
• Languages Supported
• Dataset
• Models Evaluated
• Evaluation Results
• Cross-Language Evaluation
• Metric Descriptions
• Multilingual Extension Task
• Project Structure
• Quick Start
• Running the App
• Running Evaluation
• Deployment Architecture
• Report

---

Overview

This project evaluates multiple pretrained multilingual translation models on English → Tamil (primary evaluation) and demonstrates multilingual capability across multiple Indian languages using the best performing model.

The best performing model was:

facebook/nllb-200-distilled-600M

This model supports 200+ languages, allowing the application to translate English sentences into multiple Indian languages using a single model.

Key observation:

• NLLB-200 significantly outperforms M2M100 and T5-Base for English → Tamil translation.
• The same model can also perform zero-shot translation for other Indian languages.

---

Languages Supported

Language	Script	NLLB Token	BERTScore Lang
Tamil	தமிழ்	tam_Taml	ta
Hindi	हिन्दी	hin_Deva	hi
Telugu	తెలుగు	tel_Telu	te
Kannada	ಕನ್ನಡ	kan_Knda	kn
Malayalam	മലയാളം	mal_Mlym	ml

Switching languages in the application does not reload the model.
Only the target language token (forced_bos_token_id) changes.

---

Dataset

Property	Value
Dataset	ai4bharat/IndicMTEval
Primary evaluation language	Tamil
Additional demo languages	Hindi, Telugu, Kannada, Malayalam
Split used	test
Samples per language	up to 200

Tamil was selected as the primary benchmark language because IndicMTEval provides human evaluation scores (MQM / Direct Assessment) for Tamil translations. ### URL: https://huggingface.co/datasets/ai4bharat/IndicMTEval

Preprocessing Applied

• Lowercasing
• Removing extra whitespace
• Trimming leading and trailing spaces

---

Models Evaluated

The following translation models were evaluated for English → Tamil:

Model	Parameters	Architecture	Tamil Token
facebook/nllb-200-distilled-600M	600M	Encoder-Decoder (NLLB)	tam_Taml
facebook/m2m100_418M	418M	Encoder-Decoder (M2M100)	ta
t5-base	220M	Encoder-Decoder (T5)	prompt-based

Additional semantic evaluation was performed using:

Evaluation Model	Purpose
WMT20 COMET-DA	Neural MT evaluation metric
all-MiniLM-L6-v2	Sentence embedding similarity

---

Evaluation Results

Primary Model Comparison (English → Tamil)

Model	BLEU ↑	chrF ↑	BERTScore F1 ↑	Cosine Sim ↑
NLLB-200 (600M) 🏆	0.142	41.3	0.618	0.731
M2M100 (418M)	0.098	34.7	0.581	0.694
T5-Base	0.011	12.4	0.401	0.512

NLLB-200 clearly performs best across all evaluation metrics.

---

Cross-Language Evaluation (Indic MT Benchmark)

To analyze multilingual performance, the NLLB-200 model was evaluated across several Indian languages using:

• BLEU
• chrF
• COMET (WMT20 COMET-DA)
• Sentence embedding cosine similarity

Each language used:

800 training samples
200 validation samples

---

Validation Results

Language	BLEU ↑	chrF ↑	COMET ↑	Ensemble Score ↑
Hindi	31.77	37.05	0.4404	47.20
Tamil	8.45	58.68	0.5964	49.29
Malayalam	11.95	60.29	0.5612	50.38
Marathi	37.04	78.40	0.4397	62.57
Gujarati	21.03	56.23	0.5830	52.41

---

Best Performers

Metric	Best Language	Score
BLEU	Marathi	37.04
chrF	Marathi	78.40
COMET	Tamil	0.5964
Ensemble Score	Marathi	62.57

---

Evaluation Insights

• Marathi shows strong lexical alignment with the reference translations.
• Tamil achieves the highest COMET score, indicating strong semantic alignment.
• Overall multilingual performance is stable across languages.

Average Ensemble Score: 52.37 / 100
Best Ensemble Score: 62.57 (Marathi)
Worst Ensemble Score: 47.20 (Hindi)
Score Spread: 15.37 points

All experiments were performed using the pretrained NLLB-200 model in a zero-shot translation setting.

---

Metric Descriptions

BLEU
Measures n-gram overlap between predicted translation and reference translation.

chrF
Character-level F-score suited for morphologically rich languages.

BERTScore
Uses contextual embeddings to measure semantic similarity.

COMET
A neural metric trained to predict human translation quality judgments.

Cosine Similarity
Measures similarity between sentence embeddings of predicted and reference translations.

---

Multilingual Extension Task

Since the NLLB model supports 200+ languages, the application was extended to support translation into multiple Indian languages using the same model.

Supported translations:

• English → Tamil
• English → Hindi
• English → Telugu
• English → Kannada
• English → Malayalam

Instead of loading multiple models, the system simply changes the target language token:

model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("tam_Taml")
)

This allows multilingual translation using a single model instance.

---

Project Structure

en-tamil-mt/
├── app/
│   └── app.py
├── evaluation/
│   ├── evaluate_models.py
│   └── evaluate_multilingual.py
├── notebooks/
│   └── evaluation.ipynb
├── docs/
│   └── report.md
├── requirements.txt
├── .gitignore
└── README.md

---

Quick Start

Clone the repository:

git clone https://github.com/<your-username>/en-tamil-mt.git
cd en-tamil-mt

Create environment:

python -m venv venv
source venv/bin/activate

Install dependencies:

pip install -r requirements.txt

---

Running the App

Run locally:

python app/app.py

Open:

http://localhost:7860

Live deployed application:

https://huggingface.co/spaces/Mubeen09/en-tamil-translator

---

Running Evaluation

Evaluate model comparison:

python evaluation/evaluate_models.py --model all --samples 200

Multilingual evaluation:

python evaluation/evaluate_multilingual.py --lang all --samples 200

---

Deployment Architecture

User Input (English)
        │
        ▼
Gradio UI
        │
        ▼
Preprocessing
        │
        ▼
NLLB Tokenizer
        │
        ▼
NLLB Model.generate()
        │
        ▼
Target language token
        │
        ▼
Translated output

---

Report

Full project report available in:

docs/report.md

The report includes:

• Dataset analysis
• Model comparison
• Evaluation methodology
• System architecture
• Results and conclusions

Contributors

• Sailaputri Muthavarapu
• Ashritha Gowthami Nelakurthi
• Pervez Mubeen

Subject Instructor/Guide

Mr. Panigrahi Srikanth
Assistant Professor
Chaitanya Bharathi Institute of Technology.