Bengali Natural Language Processing(BNLP)

BNLP is a natural language processing toolkit for Bengali Language. This tool will help you to tokenize Bengali text, Embedding Bengali words, Bengali POS Tagging, Construct Neural Model for Bengali NLP purposes.

Contents

• Current Features
• Installation
• Pretrained Model
• Tokenization
• Embedding
• POS Tagging
• Issue
• Contributor Guide
• Contributor List

Current Features

• Bengali Tokenization

  • SentencePiece Tokenizer
  • Basic Tokenizer
  • NLTK Tokenizer

• Bengali Word Embedding

  • Bengali Word2Vec
  • Bengali Fasttext
  • Bengali GloVe

• Bengali POS Tagging

Installation

• pypi package installer(python 3.5, 3.6, 3.7 tested okay)

  pip install bnlp_toolkit

• Local

  $git clone https://github.com/sagorbrur/bnlp.git
  $cd bnlp
  $python setup.py install
  

Pretrained Model

Download Link

• Bengali SentencePiece
• Bengali Word2Vec
• Bengali FastText
• Bengali GloVe Wordvectors
• Bengali POS Tag model

Training Details

• All three model trained with Bengali Wikipedia Dump Dataset
  • Bengali Wiki Dump
• SentencePiece Training Vocab Size=50000
• Fasttext trained with total words = 20M, vocab size = 1171011, epoch=50, embedding dimension = 300 and the training loss = 0.318668,
• Word2Vec word embedding dimension = 300
• To Know Bengali GloVe Wordvector and training process follow this repository
• Bengali CRF POS Tagging was training with nltr dataset with 80% accuracy.

Tokenization

• Bengali SentencePiece Tokenization

  • tokenization using trained model

    from bnlp.sentencepiece_tokenizer import SP_Tokenizer
    
    bsp = SP_Tokenizer()
    model_path = "./model/bn_spm.model"
    input_text = "আমি ভাত খাই। সে বাজারে যায়।"
    tokens = bsp.tokenize(model_path, input_text)
    print(tokens)

  • Training SentencePiece

    from bnlp.sentencepiece_tokenizer import SP_Tokenizer
    
    bsp = SP_Tokenizer(is_train=True)
    data = "test.txt"
    model_prefix = "test"
    vocab_size = 5
    bsp.train_bsp(data, model_prefix, vocab_size) 

• Basic Tokenizer

  from bnlp.basic_tokenizer import BasicTokenizer
  basic_t = BasicTokenizer(False)
  raw_text = "আমি বাংলায় গান গাই।"
  tokens = basic_t.tokenize(raw_text)
  print(tokens)
  
  # output: ["আমি", "বাংলায়", "গান", "গাই", "।"]

• NLTK Tokenization

  from bnlp.nltk_tokenizer import NLTK_Tokenizer
  
  text = "আমি ভাত খাই। সে বাজারে যায়। তিনি কি সত্যিই ভালো মানুষ?"
  bnltk = NLTK_Tokenizer(text)
  word_tokens = bnltk.word_tokenize()
  sentence_tokens = bnltk.sentence_tokenize()
  print(word_tokens)
  print(sentence_tokens)
  
  # output
  # word_token: ["আমি", "ভাত", "খাই", "।", "সে", "বাজারে", "যায়", "।", "তিনি", "কি", "সত্যিই", "ভালো", "মানুষ", "?"]
  # sentence_token: ["আমি ভাত খাই।", "সে বাজারে যায়।", "তিনি কি সত্যিই ভালো মানুষ?"]

Word Embedding

• Bengali Word2Vec

  • Generate Vector using pretrain model

    from bnlp.bengali_word2vec import Bengali_Word2Vec
    
    bwv = Bengali_Word2Vec()
    model_path = "model/bengali_word2vec.model"
    word = 'আমার'
    vector = bwv.generate_word_vector(model_path, word)
    print(vector.shape)
    print(vector)

  • Find Most Similar Word Using Pretrained Model

    from bnlp.bengali_word2vec import Bengali_Word2Vec
    
    bwv = Bengali_Word2Vec()
    model_path = "model/bengali_word2vec.model"
    word = 'আমার'
    similar = bwv.most_similar(model_path, word)
    print(similar)

  • Train Bengali Word2Vec with your own data

    from bnlp.bengali_word2vec import Bengali_Word2Vec
    bwv = Bengali_Word2Vec(is_train=True)
    data_file = "test.txt"
    model_name = "test_model.model"
    vector_name = "test_vector.vector"
    bwv.train_word2vec(data_file, model_name, vector_name)
    

• Bengali FastText

  • Generate Vector Using Pretrained Model

    from bnlp.bengali_fasttext import Bengali_Fasttext
    
    bft = Bengali_Fasttext()
    word = "গ্রাম"
    model_path = "model/bengali_fasttext.bin"
    word_vector = bft.generate_word_vector(model_path, word)
    print(word_vector.shape)
    print(word_vector)
    

  • Train Bengali FastText Model

    from bnlp.bengali_fasttext import Bengali_Fasttext
    
    bft = Bengali_Fasttext(is_train=True)
    data = "data.txt"
    model_name = "saved_model.bin"
    epoch = 50
    bft.train_fasttext(data, model_name, epoch)

• Bengali GloVe Word Vectors

  We trained glove model with bengali data(wiki+news articles) and published bengali glove word vectors
  You can download and use it on your different machine learning purposes.

  from bnlp.glove_wordvector import BN_Glove
  glove_path = "bn_glove.39M.100d.txt"
  word = "গ্রাম"
  bng = BN_Glove()
  res = bng.closest_word(glove_path, word)
  print(res)
  vec = bng.word2vec(glove_path, word)
  print(vec)

Bengali POS Tagging

• Bengali CRF POS Tagging

  NB: not in pypi yet. comming in next version

  • Find Pos Tag Using Pretrained Model

    from bnlp.bengali_pos import BN_CRF_POS
    bn_pos = BN_CRF_POS()
    model_path = "model/bn_pos_model.pkl"
    text = "আমি ভাত খাই।"
    res = bn_pos.pos_tag(model_path, text)
    print(res)
    # [('আমি', 'PPR'), ('ভাত', 'NC'), ('খাই', 'VM'), ('।', 'PU')]

  • Train POS Tag Model

    from bnlp.bengali_pos import BN_CRF_POS
    bn_pos = BN_CRF_POS()
    model_name = "pos_model.pkl"
    tagged_sentences = [[('রপ্তানি', 'JJ'), ('দ্রব্য', 'NC'), ('-', 'PU'), ('তাজা', 'JJ'), ('ও', 'CCD'), ('শুকনা', 'JJ'), ('ফল', 'NC'), (',', 'PU'), ('আফিম', 'NC'), (',', 'PU'), ('পশুচর্ম', 'NC'), ('ও', 'CCD'), ('পশম', 'NC'), ('এবং', 'CCD'),('কার্পেট', 'NC'), ('৷', 'PU')], [('মাটি', 'NC'), ('থেকে', 'PP'), ('বড়জোর', 'JQ'), ('চার', 'JQ'), ('পাঁচ', 'JQ'), ('ফুট', 'CCL'), ('উঁচু', 'JJ'), ('হবে', 'VM'), ('৷', 'PU')]]
    
    bn_pos.training(model_name, tagged_sentences)

Issue

• if ModuleNotFoundError: No module named 'fasttext' problem arise please do the next line

pip install fasttext

• if nltk issue arise please do the following line before importing bnlp

import nltk
nltk.download("punkt")

Contributor Guide

Check CONTRIBUTING.md page for details.

Thanks To

• Semantics Lab

Contributor List

• Sagor Sarker
• Faruk Ahmad
• Mehadi Hasan Menon
• Kazal Chandra Barman
• Md Ibrahim