bengali-bpe 0.1.6

A Byte Pair Encoding (BPE) library for the Bengali language.

Bengali BPE

bengali_bpe is a Python library for Byte Pair Encoding (BPE) specifically designed for the Bengali language.
It enables you to train BPE models on Bengali text, encode words and sentences into subword units, and decode them back.
This helps improve NLP model performance for Bengali text processing, tokenization, and embedding preparation.

---

✨ Features

• 🧠 Train a Byte Pair Encoding model on Bengali text corpus
• 🔠 Encode Bengali sentences or words into subword tokens
• 🔁 Decode subword tokens back into full Bengali words
• ⚙️ Simple, lightweight, and easy to integrate into your NLP pipelines
• 🪶 Supports Bengali Unicode normalization

---

📦 Installation

Install directly from PyPI:

pip install bengali_bpe

---

🚀 Usage Examples

from bengali_bpe import BengaliBPE
from bengali_bpe.utils import normalize_bengali_text

# Sample Bengali corpus
corpus = [
    "বাংলা ভাষা সুন্দর",
    "আমি বাংলা পড়ি",
    "বাংলা ভয়ানক নয়"
]

# Normalize text
corpus = [normalize_bengali_text(sentence) for sentence in corpus]

# Initialize and train the model
bpe = BengaliBPE(num_merges=10)
bpe.train(corpus)

# Encode a sentence
sentence = "বাংলা ভাষা সুন্দর"
encoded = bpe.encode(sentence)
print("Encoded:", encoded)

# Decode back
decoded = bpe.decode(encoded)
print("Decoded:", decoded)

---

Output

Encoded: [['বা', 'ংলা'], ['ভা', 'ষা'], ['সু', 'ন্', 'দর']]
Decoded: বাংলা ভাষা সুন্দর

Encode and Decode a Single Word

from bengali_bpe import BengaliBPE

bpe = BengaliBPE(num_merges=5)
bpe.train(["বাংলা ভাষা সুন্দর"])
encoded_word = bpe.encode_word("বাংলা")
print("Encoded Word:", encoded_word)

decoded_word = bpe.decode([encoded_word])
print("Decoded Word:", decoded_word)

---

Output

Encoded Word: ['বা', 'ংলা']
Decoded Word: বাংলা

Normalize Bengali Text

from bengali_bpe.utils import normalize_bengali_text

text = "বাংলা    ভাষা    সুন্দর।।"
print(normalize_bengali_text(text))

---

Output

বাংলা ভাষা সুন্দর।।

---

Example: Training and Applying BPE on a Bengali Paragraph

from bengali_bpe import BengaliBPE
from bengali_bpe.utils import normalize_bengali_text

text = """বাংলা একটি মধুর ভাষা। এটি বিশ্বের অন্যতম প্রাচীন ও সমৃদ্ধ ভাষাগুলোর একটি।
বাংলা ভাষার ইতিহাস ও ঐতিহ্য হাজার বছরের পুরোনো।"""

corpus = [normalize_bengali_text(text)]
bpe = BengaliBPE(num_merges=15)
bpe.train(corpus)

encoded = bpe.encode("বাংলা একটি মধুর ভাষা")
print("Encoded:", encoded)

decoded = bpe.decode(encoded)
print("Decoded:", decoded)

---

Full Example: Combine All Steps

from bengali_bpe import BengaliBPE
from bengali_bpe.utils import normalize_bengali_text

corpus = [
    "আমি বাংলা ভাষা ভালোবাসি",
    "বাংলা একটি সুন্দর ভাষা",
    "বাংলা আমাদের মাতৃভাষা"
]

corpus = [normalize_bengali_text(c) for c in corpus]
bpe = BengaliBPE(num_merges=12)
bpe.train(corpus)

sentence = "আমি বাংলা ভালোবাসি"
encoded = bpe.encode(sentence)
decoded = bpe.decode(encoded)

print("Original:", sentence)
print("Encoded:", encoded)
print("Decoded:", decoded)

---

Output

Original: আমি বাংলা ভালোবাসি
Encoded: [['আ', 'মি'], ['বা', 'ংলা'], ['ভা', 'লো', 'বা', 'সি']]
Decoded: আমি বাংলা ভালোবাসি

---

Example Use Cases

Use Case	Description
🔤 Subword Tokenization	Split Bengali words into meaningful subword units for NLP models
🧩 Embedding Preparation	Generate stable subword tokens for embedding or transformer-based models
🧠 Text Compression	Apply BPE for efficient text representation
📚 Data Preprocessing	Clean and normalize Bengali text before training models

---

API References

Function	Description
train(corpus)	Train the BPE model on a list of Bengali sentences
encode(text)	Encode an entire Bengali sentence into subword tokens
encode_word(word)	Encode a single Bengali word
decode(encoded_words)	Decode BPE tokens back to full Bengali text
normalize_bengali_text(text)	Normalize and clean Bengali text (NFC normalization)

---

Example Project Structure

bengali_bpe/
├── bengali_bpe/
│   ├── __init__.py
│   ├── encoder.py
│   └── utils.py
├── tests/
│   └── test_encoder.py
├── README.md
├── setup.py
└── LICENSE

---

Developers

Firoj Ahmmed Patwary
BSc & MSc in Statistics, Jagannath University
MSc in Data Science, Freie Universität Berlin
Researcher in Data Science, Machine Learning, NLP, and Explainable AI

🌐 Website: www.firoj.net

📧 Email: [e-mail]