sergtaima / razdel Goto Github PK

razdel — rule-based system for Russian sentence and word tokenization.

>>> from razdel import tokenize

>>> tokens = list(tokenize('Кружка-термос на 0.5л (50/64 см³, 516;...)'))
>>> tokens
[Substring(0, 13, 'Кружка-термос'),
 Substring(14, 16, 'на'),
 Substring(17, 20, '0.5'),
 Substring(20, 21, 'л'),
 Substring(22, 23, '(')
 ...]
 
>>> [_.text for _ in tokens]
['Кружка-термос', 'на', '0.5', 'л', '(', '50/64', 'см³', ',', '516', ';', '...', ')']

>>> from razdel import sentenize

>>> text = '''
... - "Так в чем же дело?" - "Не ра-ду-ют".
... И т. д. и т. п. В общем, вся газета
... '''

>>> list(sentenize(text))
[Substring(1, 23, '- "Так в чем же дело?"'),
 Substring(24, 40, '- "Не ра-ду-ют".'),
 Substring(41, 56, 'И т. д. и т. п.'),
 Substring(57, 76, 'В общем, вся газета')]

razdel supports Python 3.7+ and PyPy 3.

$ pip install razdel

Materials are in Russian:

• Razdel page on natasha.github.io
• Razdel section of Datafest 2020 talk

Unfortunately, there is no single correct way to split text into sentences and tokens. For example, one may split «Как же так?! Захар...» — воскликнут Пронин. into three sentences ["«Как же так?!", "Захар...»", "— воскликнут Пронин."] while razdel splits it into two ["«Как же так?!", "Захар...» — воскликнут Пронин."]. What would be the correct way to tokenizer т.е.? One may split in into т.|е., razdel splits into т|.|е|..

razdel tries to mimic segmentation of these 4 datasets: SynTagRus, OpenCorpora, GICRYA and RNC. These datasets mainly consist of news and fiction. razdel rules are optimized for these kinds of texts. Library may perform worse on other domains like social media, scientific articles, legal documents.

We measure absolute number of errors. There are a lot of trivial cases in the tokenization task. For example, text чуть-чуть?! is not non-trivial, one may split it into чуть|-|чуть|?|! while the correct tokenization is чуть-чуть|?!, such examples are rare. Vast majority of cases are trivial, for example text в 5 часов ... is correctly tokenized even via Python native str.split into в| |5| |часов| |.... Due to the large number of trivial case overall quality of all segmenators is high, it is hard to compare differentiate between for examlpe 99.33%, 99.95% and 99.88%, so we report the absolute number of errors.

errors — number of errors per 1000 tokens/sentencies. For example, consider etalon segmentation is что-то|?, prediction is что|-|то?, then the number of errors is 3: 1 for missing split то? + 2 for extra splits что|-|то.

time — seconds taken to process whole dataset.

spacy_tokenize, aatimofeev and others a defined in naeval/segment/models.py, for links to models see Naeval registry. Tables are computed in naeval/segment/main.ipynb.

• Chat — https://telegram.me/natural_language_processing
• Issues — https://github.com/natasha/razdel/issues
• Commercial support — https://lab.alexkuk.ru

Dev env

python -m venv ~/.venvs/natasha-razdel
source ~/.venvs/natasha-razdel/bin/activate

pip install -r requirements/dev.txt
pip install -e .

Test

make test
make int  # 2000 integration tests

Release

# Update setup.py version

git commit -am 'Up version'
git tag v0.5.0

git push
git push --tags

mystem errors on syntag

# see naeval/data
cat syntag_tokens.txt | razdel-ctl sample 1000 | razdel-ctl gen | razdel-ctl diff --show moses_tokenize | less

Non-trivial token tests

pv data/*_tokens.txt | razdel-ctl gen --recall | razdel-ctl diff space_tokenize > tests.txt
pv data/*_tokens.txt | razdel-ctl gen --precision | razdel-ctl diff re_tokenize >> tests.txt

Update integration tests

cd tests/data/
pv sents.txt | razdel-ctl up sentenize > t; mv t sents.txt

razdel and moses diff

cat data/*_tokens.txt | razdel-ctl sample 1000 | razdel-ctl gen | razdel-ctl up tokenize | razdel-ctl diff moses_tokenize | less

razdel performance

cat data/*_tokens.txt | razdel-ctl sample 10000 | pv -l | razdel-ctl gen | razdel-ctl diff tokenize | wc -l