Skip to content

yogeshhk/Inglish

BranchesTags

Repository files navigation

Inglish with an 'I' for/on/of India

Technical English to Inglish Translation System

A comprehensive framework for translating technical English content to Indian languages (Hinglish, Minglish, etc.) while preserving domain-specific terminology in their original form through intelligent code-mixing.

🎯 Problem Statement

Pure machine translation of technical content to Indian languages produces incomprehensible text. For example:

English: "This class has 4 member variables."
❌ Poor Hindi: "इस क्लास में 4 सदस्य चर हैं।"
✅ Desired Hinglish (Roman): "iis class mein chaar member variables hai"
✅ Desired Hinglish (Devanagari): "इस क्लास में चार मेंबर व्हेरिएबल्स है।"

This project implements a hybrid translation approach that:

  • Preserves technical terms in English (the language professionals actually use)
  • Translates context into native language for accessibility
  • Produces natural code-mixed output reflecting authentic communication patterns

🏗️ Architecture

┌─────────────────┐
│  English Input  │
└────────┬────────┘
         │
    ┌────▼─────────────────┐
    │  Term Extraction     │
    │  (Rule-Based + NER)  │
    └────┬─────────────────┘
         │
    ┌────▼──────────────────┐
    │  Term Guarding        │
    │  [bracket terms]      │
    └────┬──────────────────┘
         │
    ┌────▼──────────────────┐
    │  LLM Translation      │
    │  (with constraints)   │
    └────┬──────────────────┘
         │
    ┌────▼──────────────────┐
    │  Post-Processing      │
    │  & Validation         │
    └────┬──────────────────┘
         │
    ┌────▼──────────────────┐
    │  Script Conversion    │
    │  (Roman/Devanagari)   │
    └────┬──────────────────┘
         │
    ┌────▼──────────────────┐
    │  Bilingual Output     │
    └───────────────────────┘

🚀 Quick Start

Installation

# Clone repository
git clone https://github.com/yourusername/inglish-translator.git
cd inglish-translator

# Install dependencies
pip install -r requirements.txt

# Download required models
python scripts/download_models.py

Basic Usage

from inglish_translator import InglishtranslationPipeline, TranslationConfig

# Configure pipeline
config = TranslationConfig(
    domain="programming",
    target_language="hi",  # Hindi
    output_format="both"   # Roman + Devanagari
)

# Initialize pipeline
pipeline = InglishtranslationPipeline(config)

# Translate
text = "The for loop iterates over the array of integers."
result = pipeline.translate(text)

print("Roman:", result['roman'])
print("Devanagari:", result['devanagari'])

Output:

Roman: for loop array of integers ke upar iterate karta hai
Devanagari: फॉर लूप ऐरे ऑफ इंटीजर्स के ऊपर iterate करता है

📁 Project Structure

inglish-translator/
├── README.md
├── requirements.txt
├── setup.py
├── .gitignore
│
├── src/
│   ├── __init__.py
│   ├── pipeline.py              # Main translation pipeline
│   ├── term_extractor.py        # Term extraction module
│   ├── translator.py            # LLM-based translator
│   ├── script_converter.py      # Roman/Devanagari conversion
│   └── utils.py                 # Utility functions
│
├── data/
│   ├── glossaries/
│   │   ├── programming.yaml     # Programming domain terms
│   │   ├── physics.yaml         # Physics domain terms
│   │   └── finance.yaml         # Finance domain terms
│   │
│   ├── datasets/
│   │   ├── train/
│   │   │   ├── programming_train.json
│   │   │   └── physics_train.json
│   │   ├── eval/
│   │   │   ├── programming_eval.json
│   │   │   └── physics_eval.json
│   │   └── test/
│   │       └── programming_test.json
│   │
│   └── patterns/
│       └── regex_patterns.json  # Domain-specific regex patterns
│
├── models/
│   └── .gitkeep                 # Downloaded models go here
│
├── benchmarks/
│   ├── baseline_benchmark.py    # Simple rule-based baseline
│   ├── llm_benchmark.py         # LLM-based benchmark
│   └── hybrid_benchmark.py      # Hybrid approach benchmark
│
├── tests/
│   ├── test_term_extractor.py
│   ├── test_translator.py
│   ├── test_script_converter.py
│   └── test_pipeline.py
│
├── scripts/
│   ├── download_models.py       # Download pretrained models
│   ├── prepare_data.py          # Data preprocessing
│   └── evaluate.py              # Evaluation script
│
├── notebooks/
│   ├── 01_data_exploration.ipynb
│   ├── 02_term_extraction_analysis.ipynb
│   └── 03_translation_quality.ipynb
│
└── docs/
    ├── ARCHITECTURE.md
    ├── API_REFERENCE.md
    ├── CONTRIBUTING.md
    └── research_framework.md    # Full research documentation

📊 Datasets

Training Dataset Format

{
  "dataset_name": "programming_hinglish_train",
  "domain": "programming",
  "language_pair": "en_hi",
  "size": 1000,
  "samples": [
    {
      "id": "prog_001",
      "english": "The for loop iterates over the array.",
      "hinglish_roman": "for loop array ke upar iterate karta hai",
      "hinglish_devanagari": "फॉर लूप ऐरे के ऊपर iterate करता है",
      "technical_terms": ["for loop", "array", "iterate"],
      "domain": "programming"
    }
  ]
}

Benchmark Datasets

We provide curated datasets for three domains:

  1. Programming (1,000 train / 200 eval / 200 test)

    • Source: Technical documentation, Stack Overflow, coding tutorials
    • Terms: functions, variables, classes, loops, arrays, etc.

  2. Physics (800 train / 150 eval / 150 test)

    • Source: Textbooks, research papers (simplified)
    • Terms: force, momentum, velocity, energy, quantum, etc.

  3. Finance (600 train / 100 eval / 100 test)

    • Source: Financial reports, investment guides
    • Terms: ROI, equity, derivatives, portfolio, etc.

🔬 Benchmarking

Baseline Methods

We implement three baseline approaches:

1. Rule-Based Baseline (Simplest)

python benchmarks/baseline_benchmark.py \
  --dataset data/datasets/eval/programming_eval.json \
  --domain programming \
  --output results/baseline_results.json

Performance:

  • Terminology Consistency: 95%
  • Grammatical Fluency: 42%
  • Overall BLEU: 0.18
  • Speed: ~500 sentences/sec

2. LLM-Based Baseline

python benchmarks/llm_benchmark.py \
  --dataset data/datasets/eval/programming_eval.json \
  --model gpt-3.5-turbo \
  --output results/llm_results.json

Performance:

  • Terminology Consistency: 68%
  • Grammatical Fluency: 87%
  • Overall BLEU: 0.34
  • Speed: ~5 sentences/sec

3. Hybrid Approach (Recommended)

python benchmarks/hybrid_benchmark.py \
  --dataset data/datasets/eval/programming_eval.json \
  --domain programming \
  --output results/hybrid_results.json

Performance:

  • Terminology Consistency: 96%
  • Grammatical Fluency: 85%
  • Overall BLEU: 0.41
  • Speed: ~50 sentences/sec

Evaluation Metrics

python scripts/evaluate.py \
  --predictions results/hybrid_results.json \
  --references data/datasets/eval/programming_eval.json \
  --metrics all

Metrics Computed:

  • BLEU: N-gram overlap (0.3-0.4 is good)
  • METEOR: Alignment with synonyms
  • BERTScore: Semantic similarity
  • Terminology Consistency: % of technical terms preserved
  • Constraint Preservation: % of bracketed terms intact
  • Fluency Score: Perplexity-based fluency measure

📈 Sample Results

Input

English: "In object-oriented programming, a class is a blueprint for creating objects. 
Each object has member variables and methods."

Baseline (Rule-Based)

Roman: object-oriented programming mein, class blueprint hai create karne ke liye 
objects. Har object ke member variables aur methods hote hain.

Fluency: ★★☆☆☆ (2/5)
Consistency: ★★★★★ (5/5)

LLM-Based

Roman: programming mein, ek varg vastu banane ke liye ek naksha hota hai. 
Har vastu ke sadsya charo aur vidhi hote hain.

Fluency: ★★★★☆ (4/5)
Consistency: ★★★☆☆ (3/5) - Lost technical terms!

Hybrid (Recommended)

Roman: object-oriented programming mein, class objects create karne ke liye 
blueprint hai. Har object ke member variables aur methods hote hain.

Devanagari: ऑब्जेक्ट-ओरिएंटेड प्रोग्रामिंग में, क्लास ऑब्जेक्ट्स create 
करने के लिए blueprint है। हर ऑब्जेक्ट के मेंबर व्हेरिएबल्स और मेथड्स होते हैं।

Fluency: ★★★★☆ (4/5)
Consistency: ★★★★★ (5/5)

🛠️ Domain Glossaries

Glossaries are stored in YAML format with domain-specific terms:

# data/glossaries/programming.yaml
domain: programming
version: "1.0"

terms:
  - term: "variable"
    preserve: true
    context: "data storage"
    
  - term: "function"
    preserve: true
    context: "reusable code block"
    
  - term: "class"
    preserve: true
    context: "object template"
    
  - term: "array"
    preserve: true
    context: "sequence container"

compound_terms:
  - "for loop"
  - "while loop"
  - "if statement"
  - "member variable"

patterns:
  - regex: '\w+\(\)'
    type: "function_call"
    action: "preserve"

Adding New Domains

  1. Create glossary: data/glossaries/your_domain.yaml
  2. Add patterns for term detection
  3. Prepare training data in standard format
  4. Run benchmark to evaluate

🧪 Testing

# Run all tests
pytest tests/

# Run specific test
pytest tests/test_term_extractor.py -v

# Run with coverage
pytest --cov=src tests/

📊 Performance Comparison

Approach Consistency Fluency BLEU Speed (sent/sec)
Rule-Based 95% 42% 0.18 500
LLM-Only 68% 87% 0.34 5
Hybrid 96% 85% 0.41 50

🎯 Use Cases

  1. Technical Documentation Translation

    • Convert English docs to Hinglish for wider accessibility
    • Maintain technical precision while improving comprehension

  2. Educational Content

    • Textbook translation for regional language students
    • Online course material localization

  3. Professional Communication

    • Technical reports and presentations
    • Code documentation and comments

  4. Social Media & Forums

    • Technical discussions on platforms
    • Tutorial videos and blog posts

🔮 Roadmap

  • Basic rule-based term extraction
  • LLM integration with constraints
  • Script conversion (Roman/Devanagari)
  • Programming domain glossary
  • Fine-tuned NER model for technical terms
  • Physics and Finance domain expansion
  • Interactive web interface
  • REST API deployment
  • Mobile app integration
  • Marathi, Tamil, Telugu support
  • Community glossary contribution system

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Ways to contribute:

  • Add domain glossaries
  • Improve term extraction patterns
  • Contribute training data
  • Report issues and bugs
  • Improve documentation

📄 License

This project is licensed under the MIT License - see LICENSE file for details.

🙏 Acknowledgments

  • AI4Bharat for Indic language NLP tools
  • Research community for code-mixing translation work
  • Chinglish precedent demonstrating commercial viability
  • Indian tech community for authentic code-mixing patterns

📚 Citations

If you use this work, please cite:

@software{inglish_translator_2026,
  title={Inglish Translator: Technical English to Indian Language Translation},
  author={Your Name},
  year={2026},
  url={https://github.com/yourusername/inglish-translator}
}

📞 Contact

Made with ❤️ for democratizing technical knowledge across India

About

Inglish with an 'I' for/on/of India

Resources

Readme

Contributing

Contributing
Activity

Stars

1 star

Watchers

0 watching

Forks

2 forks
Report repository

Releases

No releases published

Packages

No packages published

Languages