Preparing russian hockey news for machine learning.
Unify -> Simplify -> Preprocess text and feed your neural model.
Khl is available on PyPI:
$ pip install khl
It requires Python 3.8+ to run.
To get started right away with basic usage:
from khl import text_to_codes
coder = {
'': 0, # placeholder
'???': 1, # unknown
'.': 2,
'и': 3,
'в': 4,
'-': 5,
':': 6,
'матч': 7,
'за': 8,
'забить': 9,
'гол': 10,
'per': 11, # person entity
'org': 12, # organization entity
'loc': 13, # location entity
'date': 14, # date entity
'против': 15,
'год': 16,
'pers': 17, # few persons entity
'orgs': 18, # few organizations entity
'свой': 19
}
text = """
1 апреля 2023 года в Москве в матче ⅛ финала против „Спартака” Иван Иванов забил свой 100—й гол за карьеру.
«Динамо Мск» - «Спартак» 2:1 ОТ (1:0 0:1 0:0 1:0) Голы забили: Иванов, Петров и Сидоров.
"""
codes = text_to_codes(
text=text,
coder=coder,
stop_words_=["за", "и", "свой"], # stop words to drop
replace_ners_=True, # replace named entities ("Иван Иванов" -> "per", "Спартак" -> "org", "Москва" -> "loc")
replace_dates_=True, # replace dates ("1 апреля 2023 года" -> "date")
replace_penalties_=True, # replace penalties ("5+20" -> "pen")
exclude_unknown=True, # drop lemma that not presented in coder
max_len=20, # get sequence of codes of length 20
)
# codes = [0, 0, 0, 14, 4, 13, 4, 7, 15, 12, 11, 9, 10, 2, 18, 10, 9, 6, 17, 2]
text_to_codes
is a very high level function. What's happens under hood see in Lower level usage.
coder
is just a dictionary where each lemma is represented with unique integer code.
Note that first two elements are reserved for placeholder and unknown elements.
It is possible to get coder
from frequency dictionary file (see in Get lemmas coder).
Frequency dictionary file is a json-file with dictionary where key is lemma and value is how many times this lemma occurred in your whole dataset.
Preferably it should be sorted in descending order of values.
example_frequency_dictionary.json
:
{
".": 1000,
"и": 500,
"в": 400,
"-": 300,
":": 300,
"матч": 290,
"за": 250,
"забить": 240,
"гол": 230,
"per": 200,
"org": 150,
"loc": 150,
"date": 100,
"против": 90,
"год": 70,
"pers": 40,
"orgs": 30,
"свой": 20
}
You could make and use your own frequency dictionary or download this dictionary created by myself.
from khl import stop_words
from khl import utils
from khl import preprocess
coder = preprocess.get_coder("example_frequency_dictionary.json")
text = """
1 апреля 2023 года в Москве в матче ⅛ финала против „Спартака” Иван Иванов забил свой 100—й гол за карьеру.
«Динамо Мск» - «Спартак» 2:1 ОТ (1:0 0:1 0:0 1:0) Голы забили: Иванов, Петров и Сидоров.
"""
unified_text = utils.unify(text)
# "1 апреля 2023 года в Москве в матче 1/8 финала против 'Спартака' Иван Иванов забил свой 100-й гол за карьеру. 'Динамо Мск' - 'Спартак' 2:1 ОТ (1:0 0:1 0:0 1:0) Голы забили: Иванов, Петров и Сидоров."
simplified_text = utils.simplify(
text=unified_text,
replace_ners_=True,
replace_dates_=True,
replace_penalties_=True,
)
# 'date в loc в матче финала против org per забил свой гол за карьеру. org org Голы забили: per per per.'
lemmas = preprocess.lemmatize(text=simplified_text, stop_words_=stop_words)
# ['date', 'в', 'loc', 'в', 'матч', 'финал', 'против', 'org', 'per', 'забить', 'гол', 'карьера', '.', 'orgs', 'гол', 'забить', ':', 'pers', '.']
codes = preprocess.lemmas_to_codes(
lemmas=lemmas,
coder=coder,
exclude_unknown=True,
max_len=20,
)
# [0, 0, 0, 14, 4, 13, 4, 7, 15, 12, 11, 9, 10, 2, 18, 10, 9, 6, 17, 2]
print(
preprocess.codes_to_lemmas(codes=codes, coder=coder)
)
# ['', '', '', 'date', 'в', 'loc', 'в', 'матч', 'против', 'org', 'per', 'забить', 'гол', '.', 'orgs', 'гол', 'забить', ':', 'pers', '.']