pongsakornsemsuwan / deepcut

A Thai word tokenization library using Deep Neural Network

Deepcut

A Thai word tokenization library using Deep Neural Network.

What's new?

• v0.5.2.0: Better weight matrix
• v0.5.1.0: Faster tokenization by code refactorization from our new contributor: Titipat Achakulvisut

Performance

The Convolutional Neural network is trained from 90% of NECTEC's BEST corpus (consists of 4 sections, article, news, novel and encyclopedia) and test on the rest 10%. It is a binary classification model trying to predict whether a character is the beginning of word or not. The results calculated from only 'true' class are as follow

• f1 score: 98.1%
• precision score: 97.8%
• recall score: 98.5%

Installation

Install using pip for stable release,

pip install deepcut

For latest development release,

pip install git+git://github.com/rkcosmos/deepcut.git

Or clone the repository and install using setup.py

python setup.py install

Make sure you are using tensorflow backend in Keras by making sure ~/.keras/keras.json is as follows (see also https://keras.io/backend/)

{
  "floatx": "float32",
  "epsilon": 1e-07,
  "backend": "tensorflow",
  "image_data_format": "channels_last"
}

We do not add tensorflow in automatic installation process because it has cpu and gpu version. Installing cpu version to everyone might break those who already have gpu version installed. So please install tensorflow yourself following this guildline https://www.tensorflow.org/install/.

Docker

Install Docker on your machine

For Linux:

curl -sSL https://get.docker.com | sudo sh
docker build -t deepcut .

For other OS: see https://docs.docker.com/engine/installation/

To run this Docker image:

docker run --rm -it deepcut

It will open a shell for us to play with deepcut.

Usage

import deepcut
deepcut.tokenize('ตัดคำได้ดีมาก')

Output will be in list format

['ตัด','คำ','ได้','ดี','มาก']

Notes

Some texts might not be segmented as we would expected (e.g. 'โรงเรียน' -> ['โรง', 'เรียน']), this is because of

• BEST corpus (training data) tokenizes word this way (They use 'Compound words' as a criteria for segmentation)

• They are unseen/new words -> Ideally, this would be cured by having better corpus but it's not very practical so I am thinking of doing semi-supervised learning to incorporate new examples.

Any suggestion and comment are welcome, please post it in issue section.

Contributors

• Rakpong Kittinaradorn
• Korakot Chaovavanich
• Titipat Achakulvisut
• Chanwit Kaewkasi

Partner Organizations

• True Corporation

And we are open for contribution and collaboration.