meitei-senter 1.1.2

A lightweight sentence boundary detector for Meitei Mayek (Manipuri) text

Meitei Senter

A lightweight sentence boundary detector for Meitei Mayek (Manipuri) text.

Features

• 🚀 Lightweight - Only ~1MB model, minimal dependencies
• 🎯 Accurate - 94.7% F-Score on Meitei text
• 🔧 Easy to use - Simple Python API and CLI
• ⚡ Fast - Optimized for quick inference

---

Installation

pip install meitei-senter

Optional: spaCy Backend (for higher accuracy)

pip install meitei-senter[spacy]

---

Quick Start

Python API

from meitei_senter import MeiteiSentenceSplitter

# Initialize the splitter
splitter = MeiteiSentenceSplitter()

# Split text into sentences
text = "ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ ꯫ ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ ꯫"
sentences = splitter.split_sentences(text)

for i, sent in enumerate(sentences, 1):
    print(f"{i}. {sent}")

Output:

1. ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ ꯫
2. ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ ꯫

Command Line

# Interactive mode
meitei-senter --interactive

# Direct text input
meitei-senter --text "ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ ꯫ ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ ꯫"

# Show version
meitei-senter --version

---

Advanced Usage

Using the Convenient Loader

from meitei_senter import load_splitter

# Load with default (delimiter-based) backend
splitter = load_splitter()

# Or with spaCy backend (requires spacy extra)
splitter = load_splitter(use_spacy=True)

sentences = splitter.split_sentences("Your Meitei text here ꯫")

Using Neural Network Mode

from meitei_senter import MeiteiSentenceSplitter

# Enable neural mode for context-aware splitting
splitter = MeiteiSentenceSplitter(use_neural=True)
sentences = splitter.split_sentences(text)

Direct Callable Interface

from meitei_senter import MeiteiSentenceSplitter

splitter = MeiteiSentenceSplitter()

# Call splitter directly
sentences = splitter("ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ... ꯫ ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ ꯫")

With spaCy (Custom Tokenizer)

import spacy
import os
from meitei_senter import MeiteiTokenizer, get_model_path

# Get path to bundled model
model_path = os.path.join(get_model_path(), 'meitei_tokenizer.model')

# Create blank spaCy model with custom tokenizer
nlp = spacy.blank("xx")
nlp.tokenizer = MeiteiTokenizer(model_path, nlp.vocab)

doc = nlp("ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ ꯫")
print([token.text for token in doc])
# Output: ['ꯆꯦ', 'ꯔꯣ', 'ꯀꯤ', 'ꯑꯁꯤ', 'ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥ', 'ꯒꯤ', 'ꯁꯍꯔ', 'ꯅꯤ', '꯫']

---

🌐 REST API Server

Start the FastAPI server for HTTP-based sentence splitting:

Start Server

# Using CLI
meitei-senter-server --port 8000

# Or with uvicorn directly
uvicorn meitei_senter.server:app --host 0.0.0.0 --port 8000

# With auto-reload for development
meitei-senter-server --port 8000 --reload

API Endpoints (POST only)

Endpoint	Method	Description
/	POST	API info
/health	POST	Health check
/split	POST	Split text into sentences
/tokenize	POST	Tokenize text
/docs	GET	Swagger UI

Example Requests

POST /split

curl -X POST "http://localhost:8000/split" \
     -H "Content-Type: application/json" \
     -d '{"text": "ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ ꯫ ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ ꯫"}'

Response:

{
  "sentences": ["ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ ꯑꯣꯀ꯭ꯂꯥꯍꯣꯃꯥꯒꯤ ꯁꯍꯔꯅꯤ꯫", "ꯃꯁꯤ ꯌꯥꯝꯅ ꯆꯥꯎꯏ꯫"],
  "count": 2
}

POST /tokenize

curl -X POST "http://localhost:8000/tokenize" \
     -H "Content-Type: application/json" \
     -d '{"text": "ꯆꯦꯔꯣꯀꯤ ꯑꯁꯤ"}'

Response:

{
  "tokens": ["▁ꯆꯦ", "ꯔꯣ", "ꯀꯤ", "▁ꯑꯁꯤ"],
  "token_ids": [460, 390, 42, 3],
  "count": 4
}

POST /health

curl -X POST "http://localhost:8000/health"

Response:

{"status": "ok", "version": "1.1.0", "model_loaded": true}

---

📊 Model Details

Feature	Specification
Model Size	~1 MB
Tokenizer	SentencePiece (Unigram, 8K vocab)
Architecture	CNN (HashEmbedCNN)
F-Score	94.71%
Precision	93.94%
Recall	95.49%

---

📂 Repository Structure

mni_tokenizer/
├── meitei_senter/              # Main package
│   ├── __init__.py             # Package exports
│   ├── cli.py                  # Command-line interface
│   ├── model.py                # PyTorch model & splitter
│   ├── tokenizer.py            # spaCy tokenizer
│   ├── meitei_tokenizer.model  # SentencePiece model
│   ├── meitei_senter.pth       # PyTorch weights
│   └── meitei_senter.json      # Model config
├── pyproject.toml              # Build configuration
└── README.md                   # This file

---

API Reference

MeiteiSentenceSplitter

Main class for sentence splitting.

MeiteiSentenceSplitter(
    pth_path: str = None,      # Path to PyTorch model
    spm_path: str = None,      # Path to SentencePiece model
    config_path: str = None,   # Path to config JSON
    use_neural: bool = False   # Enable neural network mode
)

Methods:

Method	Description
split_sentences(text)	Split text into list of sentences
tokenize(text)	Tokenize text into pieces and IDs
__call__(text)	Direct callable interface

MeiteiTokenizer

spaCy-compatible tokenizer using SentencePiece.

MeiteiTokenizer(model_path: str, vocab: spacy.Vocab)

load_splitter

Convenience function to load a pre-configured splitter.

load_splitter(use_spacy: bool = False)

---

🔧 Development

# Clone repository
git clone https://github.com/Okramjimmy/mni_tokenizer.git
cd mni_tokenizer

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Build package
python -m build

# Upload to PyPI
twine upload dist/*

---

📜 License

MIT License - see LICENSE for details.

---

📚 Citation

If you use this in your research, please cite:

@software{meitei_senter,
  author = {Okram Jimmy},
  title = {Meitei Senter: Sentence Boundary Detection for Meitei Mayek},
  year = {2024},
  url = {https://github.com/Okramjimmy/mni_tokenizer}
}

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📧 Contact

• Author: Okram Jimmy
• Email: [e-mail]
• GitHub: @Okramjimmy