🐍 python-vaporetto 🛥

Vaporetto is a fast and lightweight pointwise prediction based tokenizer. This is a Python wrapper for Vaporetto.

Installation

Install pre-built package from PyPI

Run the following command:

$ pip install vaporetto

Build from source

You need to install the Rust compiler following the documentation beforehand. vaporetto uses pyproject.toml, so you also need to upgrade pip to version 19 or later.

$ pip install --upgrade pip

After setting up the environment, you can install vaporetto as follows:

$ pip install git+https://github.com/daac-tools/python-vaporetto

Example Usage

python-vaporetto does not contain model files. To perform tokenization, follow the document of Vaporetto to download distribution models or train your own models beforehand.

Check the version number as shown below to use compatible models:

>>> import vaporetto
>>> vaporetto.VAPORETTO_VERSION
'0.6.5'

Examples:

# Import vaporetto module
>>> import vaporetto

# Load the model file
>>> with open('tests/data/vaporetto.model', 'rb') as fp:
...     model = fp.read()

# Create an instance of the Vaporetto
>>> tokenizer = vaporetto.Vaporetto(model, predict_tags = True)

# Tokenize
>>> tokenizer.tokenize_to_string('まぁ社長は火星猫だ')
'まぁ/名詞/マー 社長/名詞/シャチョー は/助詞/ワ 火星/名詞/カセー 猫/名詞/ネコ だ/助動詞/ダ'

>>> tokens = tokenizer.tokenize('まぁ社長は火星猫だ')

>>> len(tokens)
6

>>> tokens[0].surface()
'まぁ'

>>> tokens[0].tag(0)
'名詞'

>>> tokens[0].tag(1)
'マー'

>>> [token.surface() for token in tokens]
['まぁ', '社長', 'は', '火星', '猫', 'だ']

Note for distributed models

The distributed models are compressed in zstd format. If you want to load these compressed models, you must decompress them outside the API.

>>> import vaporetto
>>> import zstandard  # zstandard package in PyPI

>>> dctx = zstandard.ZstdDecompressor()
>>> with open('tests/data/vaporetto.model.zst', 'rb') as fp:
...    with dctx.stream_reader(fp) as dict_reader:
...        tokenizer = vaporetto.Vaporetto(dict_reader.read(), predict_tags = True)

Note for KyTea's models

You can also use KyTea's models as follows:

>>> with open('path/to/jp-0.4.7-5.mod', 'rb') as fp:  # doctest: +SKIP
...     tokenizer = vaporetto.Vaporetto.create_from_kytea_model(fp.read())

Note: Vaporetto does not support tag prediction with KyTea's models.

Speed Comparison

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contribution

See the guidelines.