toiro 0.0.11

A comparison tool of Japanese tokenizers

toiro

Toiro is a comparison tool of Japanese tokenizers.

• Compare the processing speed of tokenizers
• Compare the words segmented in tokenizers
• Compare the performance of tokenizers by benchmarking application tasks (e.g., text classification)

It also provides useful functions for natural language processing in Japanese.

• Data downloader for Japanese text corpora
• Preprocessor of these corpora
• Text classifier for Japanese text (e.g., SVM, BERT)

Installation

Python 3.10+ is required. You can install toiro with the following command. Janome is included in the default installation.

pip install toiro

Adding a tokenizer to toiro

If you want to add a tokenizer to toiro, please install it individually. This is an example of adding SudachiPy and nagisa to toiro.

pip install sudachipy sudachidict_core
pip install nagisa

How to install other tokenizers

mecab-python3

pip install mecab-python3

GiNZA

pip install spacy ginza

spaCy

pip install spacy[ja]

KyTea

You need to install KyTea. Please refer to here.

pip install kytea

Juman++ v2

You need to install Juman++ v2. Please refer to here.

pip install pyknp

SentencePiece

pip install sentencepiece

fugashi-ipadic

pip install fugashi ipadic

fugashi-unidic

pip install fugashi unidic-lite

tinysegmenter

pip install tinysegmenter3

tiktoken

pip install tiktoken

If you want to install all the tokonizers at once, please use the following command.

pip install toiro[all_tokenizers]

Getting started

You can check the available tokonizers in your Python environment.

from toiro import tokenizers

available_tokenizers = tokenizers.available_tokenizers()
print(available_tokenizers)

Toiro supports 12 different Japanese tokonizers and 1 BPE tokenizer. This is an example of adding SudachiPy and nagisa.

{'nagisa': {'is_available': True, 'version': '0.2.7'},
 'janome': {'is_available': True, 'version': '0.3.10'},
 'mecab-python3': {'is_available': False, 'version': False},
 'sudachipy': {'is_available': True, 'version': '0.4.9'},
 'spacy': {'is_available': False, 'version': False},
 'ginza': {'is_available': False, 'version': False},
 'kytea': {'is_available': False, 'version': False},
 'jumanpp': {'is_available': False, 'version': False},
 'sentencepiece': {'is_available': False, 'version': False},
 'fugashi-ipadic': {'is_available': False, 'version': False},
 'fugashi-unidic': {'is_available': False, 'version': False},
 'tinysegmenter': {'is_available': False, 'version': False},
 'tiktoken': {'is_available': False, 'version': False}}

Download the livedoor news corpus and compare the processing speed of tokenizers.

from toiro import tokenizers
from toiro import datadownloader

# A list of avaliable corpora in toiro
corpora = datadownloader.available_corpus()
print(corpora)
#=> ['livedoor_news_corpus', 'yahoo_movie_reviews', 'amazon_reviews']

# Download the livedoor news corpus and load it as pandas.DataFrame
corpus = corpora[0]
datadownloader.download_corpus(corpus)
train_df, dev_df, test_df = datadownloader.load_corpus(corpus)
texts = train_df[1]

# Compare the processing speed of tokenizers
report = tokenizers.compare(texts)
#=> [1/3] Tokenizer: janome
#=> 100%|███████████████████| 5900/5900 [00:07<00:00, 746.21it/s]
#=> [2/3] Tokenizer: nagisa
#=> 100%|███████████████████| 5900/5900 [00:15<00:00, 370.83it/s]
#=> [3/3] Tokenizer: sudachipy
#=> 100%|███████████████████| 5900/5900 [00:08<00:00, 696.68it/s]
print(report)
{'execution_environment': {'python_version': '3.7.8.final.0 (64 bit)',
  'arch': 'X86_64',
  'brand_raw': 'Intel(R) Core(TM) i7-7700K CPU @ 4.20GHz',
  'count': 8},
 'data': {'number_of_sentences': 5900, 'average_length': 37.69593220338983},
 'janome': {'elapsed_time': 9.114670515060425},
 'nagisa': {'elapsed_time': 15.873093605041504},
 'sudachipy': {'elapsed_time': 9.05256724357605}}

# Compare the words segmented in tokenizers
text = "都庁所在地は新宿区。"
tokenizers.print_words(text, delimiter="|")
#=>        janome: 都庁|所在地|は|新宿|区|。
#=>        nagisa: 都庁|所在|地|は|新宿|区|。
#=>     sudachipy: 都庁|所在地|は|新宿区|。

Run toiro in Docker

You can use all tokenizers by building a docker container from Docker Hub.

docker run --rm -it taishii/toiro /bin/bash

How to run the Python interpreter in the Docker container

Run the Python interpreter.

root@cdd2ad2d7092:/workspace# python3

Compare the words segmented in tokenizers

>>> from toiro import tokenizers
>>> text = "都庁所在地は新宿区。"
>>> tokenizers.print_words(text, delimiter="|")
 mecab-python3: 都庁|所在地|は|新宿|区|。
        janome: 都庁|所在地|は|新宿|区|。
        nagisa: 都庁|所在|地|は|新宿|区|。
     sudachipy: 都庁|所在地|は|新宿区|。
         spacy: 都庁|所在|地|は|新宿|区|。
         ginza: 都庁|所在地|は|新宿区|。
         kytea: 都庁|所在|地|は|新宿|区|。
       jumanpp: 都庁|所在|地|は|新宿|区|。
 sentencepiece: ▁|都|庁|所在地|は|新宿|区|。
fugashi-ipadic: 都庁|所在地|は|新宿|区|。
fugashi-unidic: 都庁|所在|地|は|新宿|区|。
 tinysegmenter: 都庁所|在地|は|新宿|区|。
 tiktoken_gpt4o: 都|�|�|所在地|は|新|宿|区|。
 tiktoken_gpt5: 都|�|�|所在地|は|新|宿|区|。

Get more information about toiro

The slides at PyCon JP 2020

• Speaker Deck
• PyConJP2020_Online.ipynb

Tutorials in Japanese

• 01_getting_started_ja.ipynb
• 05_svm_vs_bert_benchmarking_application_tasks_ja.ipynb