Skip to content

Kowsher/Bangla-Fasttext

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Bangla FastText Model & Toolkit

We have constructed a dataset that contains Bangla text data for training unsupervised ML model, and it contains around 14 GB of text data. One of the largest in Bengali Language model called BanglaLM: Bangla Corpus For Language Model Research. The Bangla FastText model had been developed based on this dataset. We used google cloud to train model. We developed two models based on skipgram and cbow training method. This is open source python module to use these two models easily. We also developed sentence embedding systems for the using of sklearn classifiers. It showed better perfromance than facebook pretrained fasttext model on Bangla Wikidataset.

Dataset (Bengali)

Link for the BanglaLM dataset :

-> Github: BanglaLM: Bangla Corpus For Language Model Research

-> Kaggle: BanglaLM: Bangla Corpus For Language Model Research

Model link:

Bangla FastText

Installation

To install the latest release, we can do :

!pip install BanglaFastText

or, to get the latest development version of BanglaFastText, we can install from our github repository :

$ https://github.com/Kowsher/Bangla-Fasttext.git
$ cd Bangla-Fasttext
$ sudo pip install .
$ # or :
$ sudo python setup.py install

For further information and introduction see README.md

Getting started

In order to learn word vectors, as described here, BanglaFastText function like this:

import BanglaFastText

#There are two variations of training methods cbow and skipgram. By default, it's cbow method and the model preparation path is set default as current working directory 
>>> Bn = BanglaFastText.BanglaFasttext()

#If want to save the model in manual path, we can by using 'sav_path' parameter.

# Skipgram model :
>>> Bn = BanglaFastText.BanglaFasttext(method='skipgram', save_path = '/content/model')
# 'path' is the directory to save the downloaded model
>>> model = Bn.model_load()

# or, cbow model :
>>> Bn = BanglaFastText.BanglaFasttext(method='cbow', save_path = '/content/model')
>>> model = Bn.model_load()

Where method parameter is to choose the training method and path is to save model.

Loading a model object

If we have already model then we can simply read and load the model as :

# To read a model
>>> Bn = BanglaFastText.BanglaFasttext(model_path = './model_name')

# to load the model as object we can
>>> model = Bn.model_load()

Playing with the parameters

# to get vector of a word
>>> model['দেশ']

# to get most similar words
>>> model.most_similar("দেশ")

# to find word similarity
>>> Bn.word_similarity('কিতাব', 'বই')

# to find sentence similarity
>>> Bn.sent_similarity('আমি দেশকে ভালোবাসি', 'অনেক সুন্দর আমাদের দেশ')

#  for sentence embedding 
>>> corpus = ['আমি দেশকে ভালোবাসি', 'অনেক সুন্দর আমাদের দেশ']
>>> X = Bn.sent_embd(corpus)

Fine Tuning

If we want to fine tuning or update weights by our dataset

>>> corpus = ['আমি দেশকে ভালোবাসি', 'অনেক সুন্দর আমাদের দেশ']
>>> Bn.fine_tuning(corpus, epochs=5)
>>> model = Bn.model_load()

......
>>> tuned_model = Bn.fine_tuning(corpus, epochs=5) # to get the raw model after finetuned, if we want to use it further  

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published