AttaCut: Fast and Reasonably Accurate Word Tokenizer for Thai

How does AttaCut look like?

TL;DR: 3-Layer Dilated CNN on syllable and character features. It’s 6x faster than DeepCut (SOTA) while its WL-f1 on BEST is 91%, only 2% lower.

Installation

$ pip install attacut

Remarks: Windows users need to install PyTorch before the command above. Please consult PyTorch.org for more details.

Usage

Command-Line Interface

$ attacut-cli -h
AttaCut: Fast and Reasonably Accurate Word Tokenizer for Thai

Usage:
  attacut-cli <src> [--dest=<dest>] [--model=<model>]
  attacut-cli [-v | --version]
  attacut-cli [-h | --help]

Arguments:
  <src>             Path to input text file to be tokenized

Options:
  -h --help         Show this screen.
  --model=<model>   Model to be used [default: attacut-sc].
  --dest=<dest>     If not specified, it'll be <src>-tokenized-by-<model>.txt
  -v --version      Show version

High-Level API

from attacut import tokenize, Tokenizer

# tokenize `txt` using our best model `attacut-sc`
words = tokenize(txt)

# alternatively, an AttaCut tokenizer might be instantiated directly, allowing
# one to specify whether to use `attacut-sc` or `attacut-c`.
atta = Tokenizer(model="attacut-sc")
words = atta.tokenize(txt)

For better efficiency, we recommend using attacut-cli. Please consult our Google Colab tutorial for more detials.

Benchmark Results

Belows are brief summaries. More details can be found on our benchmarking page.

Tokenization Quality

Speed

Retraining on Custom Dataset

Please refer to our retraining page

Related Resources

Acknowledgements

This repository was initially done by Pattarawat Chormai, while interning at Dr. Attapol Thamrongrattanarit's NLP Lab, Chulalongkorn University, Bangkok, Thailand. Many people have involed in this project. Complete list of names can be found on Acknowledgement.

PyThaiNLP / attacut

readme