jagger-python

Python binding for Jagger(C++ implementation of Pattern-based Japanese Morphological Analyzer) : https://www.tkl.iis.u-tokyo.ac.jp/~ynaga/jagger/index.en.html

Install

$ python -m pip install jagger

This does not install model files.

You can download precompiled KWDLC model from https://github.com/lighttransport/jagger-python/releases/download/v0.1.0/model_kwdlc.tar.gz (Note that KWDLC has unclear license/TermOfUse. Use it at your own risk)

Example

import jagger

model_path = "model/kwdlc/patterns"

tokenizer = jagger.Jagger()
tokenizer.load_model(model_path)

text = "吾輩は猫である。名前はまだない。"
toks = tokenizer.tokenize(text)

for tok in toks:
    print(tok.surface(), tok.feature())
print("EOS")

"""
吾輩    名詞,普通名詞,*,*,吾輩,わがはい,代表表記:我が輩/わがはい カテゴリ:人
は      助詞,副助詞,*,*,は,は,*
猫      名詞,普通名詞,*,*,猫,ねこ,*
である  判定詞,*,判定詞,デアル列基本形,だ,である,*
。      特殊,句点,*,*,。,。,*
名前    名詞,普通名詞,*,*,名前,なまえ,*
は      助詞,副助詞,*,*,は,は,*
まだ    副詞,*,*,*,まだ,まだ,*
ない    形容詞,*,イ形容詞アウオ段,基本形,ない,ない,*
。      特殊,句点,*,*,。,。,*
EOS
"""

# print tags
for tok in toks:
    # print tag(split feature() by comma)
    print(tok.surface())
    for i in range(tok.n_tags()):
        print("  tag[{}] = {}".format(i, tok.tag(i)))
print("EOS")

Batch processing(experimental)

tokenize_batch tokenizes multiple lines(delimited by newline('\n', '\r', or '\r\n')) at once. Splitting lines is done in C++ side.

import jagger

model_path = "model/kwdlc/patterns"

tokenizer = jagger.Jagger()
tokenizer.load_model(model_path)

text = """
吾輩は猫である。
名前はまだない。
明日の天気は晴れです。
"""

# optional: set C++ threads(CPU cores) to use
# default: Use all CPU cores.
# tokenizer.set_threads(4)

toks_list = tokenizer.tokenize_batch(text)

for toks in toks_list:
    for tok in toks:
        print(tok.surface(), tok.feature())

Train a model.

Pyhthon interface for training a model is not provided yet. For a while, you can build C++ trainer cli using CMake(Windows supported). See train/ for details.

Limitation

Single line string must be less than 262,144 bytes(~= 87,000 UTF-8 Japanese chars).

Jagger version

Jagger version used in this Python binding is

2023-02-18

For developer

Edit dev_mode=True in to enable asan + debug build

Run python script with

$ LD_PRELOAD=$(gcc -print-file-name=libasan.so) python FILE.py

or

$ LD_PRELOAD=$(clang -print-file-name=libclang_rt.asan-x86_64.so) python FILE.py

Releasing

Version is created automatically using setuptools_scm.

• tag it: git tag vX.Y.Z
• push tag: git push --tags

TODO

• Provide a model file trained from Wikipedia, UniDic, etc(clearer & permissive licencing&TermOfUse).
  • Use GiNZA for morphological analysis.
• Split feature vector(CSV) considering quote char when extracting tags.
  • e.g. 'a,b,"c,d",e' => ["a", "b", "c,d", "e"]
• Optimize C++ <-> Python interface
  • string_view(or read-only string literal) for tag str.
  • pickle support(for exchanging Python object when using multiprocessing)
    • https://pybind11.readthedocs.io/en/latest/advanced/classes.html#pickling-support

License

Python binding is available under 2-clause BSD licence.

Jagger and ccedar_core.h is licensed under GPLv2/LGPLv2.1/BSD triple licenses.

Third party licences

• stack_container.h: BSD like license.
• nanocsv.h MIT license.