def update_tf(id_):
result = gql_client.execute(note_from_id, variable_values={"id": id_})
note = result["note"]
html = note["contentHtml"]
text = html_text.extract_text(html)
tokenizer = WordTokenizer("sudachi", mode="C", with_postag=True)
words = tokenizer.tokenize(text)
hiragana_re = re.compile("[\u3041-\u309F]+")
number_re = re.compile("[0-9,.]+")
filtered_words = list(
filter(
# lambda x: len(x) > 3 or not hiragana_re.fullmatch(x),
lambda x: (len(x) > 3 if hiragana_re.fullmatch(x) else len(x) > 1)
and not number_re.fullmatch(x),
# map(lambda x: x.normalized_form, filter(lambda x: x.postag in ["名詞", "動詞"], words)),
map(lambda x: x.surface, filter(lambda x: x.postag in ["名詞"],
words)),
))
num_words = len(filtered_words)
word_count = Counter(filtered_words)
word_freq_list = list(
map(lambda k: (k, word_count[k] / num_words), word_count))
word_freq = dict(word_freq_list)
tf_tsv_key = get_tf_tsv_key(id_)
tf_tsv = "\n".join(map(lambda x: "\t".join(map(str, x)), word_freq_list))
private_bucket.put_object(Body=tf_tsv.encode("utf-8"), Key=tf_tsv_key)
return word_freq
async def batch_tokenize(params: TokenizeParameter, request: Request):
if params.texts is None:
raise HTTPException(status_code=400, detail="texts is required.")
cache_key = generate_cache_key(params)
if cache_key in request.app.state.cache:
logging.info(f"Hit cache: {cache_key}")
tokenizer = request.app.state.cache[cache_key]
else:
logging.info(f"Create tokenizer: {cache_key}")
try:
tokenizer = WordTokenizer(
tokenizer=params.tokenizer,
user_dictionary_path=params.user_dictionary_path,
system_dictionary_path=params.system_dictionary_path,
model_path=params.model_path,
mode=params.mode,
dictionary_format=params.dictionary_format,
)
request.app.state.cache[cache_key] = tokenizer
except Exception:
raise HTTPException(status_code=400,
detail="fail to initialize tokenizer")
tokens_list = [[token.dict() for token in tokenizer.tokenize(text)]
for text in params.texts]
return {"tokens_list": tokens_list}
def tokenize(params: TokenizeParameter, request: Request):
if params.texts is not None:
message = (
"A parameter `texts` is now unacceptable for /api/v1/tokenize."
" Please use /api/v1/batch_tokenize instead.")
raise HTTPException(status_code=400, detail=message)
if params.text is None:
raise HTTPException(status_code=400,
detail="text or texts is required.")
cache_key = generate_cache_key(params)
if cache_key in request.app.state.cache:
logging.info(f"Hit cache: {cache_key}")
tokenizer = request.app.state.cache[cache_key]
else:
logging.info(f"Create tokenizer: {cache_key}")
try:
tokenizer = WordTokenizer(
tokenizer=params.tokenizer,
user_dictionary_path=params.user_dictionary_path,
system_dictionary_path=params.system_dictionary_path,
model_path=params.model_path,
mode=params.mode,
dictionary_format=params.dictionary_format,
)
request.app.state.cache[cache_key] = tokenizer
except Exception:
raise HTTPException(status_code=400,
detail="fail to initialize tokenizer")
tokens = [token.dict() for token in tokenizer.tokenize(params.text)]
return {"tokens": tokens}
def test_tokenize(raw_texts: List[str], tokenizer_params: Dict):
tokenizer_name = tokenizer_params["tokenizer"]
tokenizer = WordTokenizer(**tokenizer_params)
expect = [
Token.from_dict(token_param)
for token_param in read_lines(tokenizer_name)[0]
]
result = tokenizer.tokenize(raw_texts[0])
assert expect == result
def test_tokenize_with_character(raw_texts: List[str], tokenizer_params: Dict):
if tokenizer_params["tokenizer"] == "kytea" and sys.version_info < (3, 7):
pytest.skip("KyTea doesn't work in Python3.6")
tokenizer_name = tokenizer_params["tokenizer"]
tokenizer = WordTokenizer(**tokenizer_params)
expect = [
Token.from_dict(token_param)
for token_param in read_lines(tokenizer_name)[0]
]
result = tokenizer.tokenize(raw_texts[0])
assert expect == result
def preprocess_data(df):
# split
df['text'] = df['text'].apply(neologdn.normalize)
le = preprocessing.LabelEncoder()
df['label'] = le.fit_transform(df['label'])
df_train, df_test, y_train, y_test = train_test_split(df,
df['label'].values,
test_size=0.2,
random_state=42,
stratify=df['label'])
# tokenize
tokenizer = WordTokenizer('MeCab')
docs_train = np.array([
' '.join(map(str, tokenizer.tokenize(text)))
for text in df_train['text']
])
docs_test = np.array([
' '.join(map(str, tokenizer.tokenize(text)))
for text in df_test['text']
])
# tfidf: Don't use df_test for fitting
count_vec = CountVectorizer(min_df=2,
max_features=20000,
ngram_range=(1, 3))
bags_train = count_vec.fit_transform(docs_train)
bags_test = count_vec.transform(docs_test)
tfidf = TfidfTransformer(use_idf=True, norm='l2', smooth_idf=True)
tf_idf_train = tfidf.fit_transform(bags_train)
tf_idf_test = tfidf.transform(bags_test)
X_train = pd.DataFrame(tf_idf_train.toarray())
X_test = pd.DataFrame(tf_idf_test.toarray())
return X_train.reset_index(drop=True), X_test.reset_index(
drop=True), y_train, y_test
def tokenize(params: TokenizeParameter, request: Request):
if params.text is not None:
texts = [params.text]
elif params.texts is not None:
texts = params.texts
else:
raise HTTPException(status_code=400,
detail="text or texts is required.")
mode = params.mode.lower()
model_path = ("data/model.spm" if params.tokenizer.lower()
== "sentencepiece" else None) # NOQA
signature = f"{params.tokenizer}.{model_path}.{mode}"
if signature in request.app.tokenizers:
logging.info(f"Hit cache: {signature}")
tokenizer = request.app.tokenizers[signature]
else:
logging.info(f"Create tokenizer: {signature}")
try:
tokenizer = WordTokenizer(
tokenizer=params.tokenizer,
with_postag=True,
model_path=model_path,
mode=mode,
)
request.app.tokenizers[signature] = tokenizer
except Exception:
raise HTTPException(status_code=400,
detail="fail to initialize tokenizer")
results = [[{
"surface": t.surface,
"part_of_speech": t.postag
} for t in tokenizer.tokenize(text)] for text in texts]
return {"tokens": results}
" has been trained with n_features=%d" % (
n_features, expected_n_features))
X = X * self._idf_diag
return X
if __name__ == '__main__':
df = load_pandas_df(nrows=10)
# Normalization
df['text'] = df['text'].apply(neologdn.normalize)
tokenizer = WordTokenizer('MeCab')
docs = np.array([
' '.join(map(str, tokenizer.tokenize(text))) for text in df['text']
])
print(docs.shape)
# (10,)
count_vec = CountVectorizer(min_df=2,
max_features=20000,
ngram_range=(1, 3))
bags = count_vec.fit_transform(docs)
print(bags.toarray().shape)
print(bags.toarray())
"""
(10, 445)
[[1 0 1 ... 0 0 0]
[1 0 0 ... 0 0 0]
try:
_tokenizer = WordTokenizer("Sudachi", mode="A", with_postag=True)
word_tokenizers.append(_tokenizer)
except (ImportError, KeyError):
print("Skip: ", "Sudachi")
print("Finish creating word tokenizers")
print()
# ref: https://ja.wikipedia.org/wiki/東京特許許可局
document = "東京特許許可局（とうきょうとっきょきょかきょく） 日本語の早口言葉。" # NOQA
document += "なお実際に特許に関する行政を行うのは特許庁であり、過去にこのような役所が存在したことは一度も無い。" # NOQA
print("Given document: {}".format(document))
sentences = sentence_tokenizer.tokenize(document)
for sentence_id, sentence in enumerate(sentences):
print("#{}: {}".format(sentence_id, sentence))
for word_tokenizer in word_tokenizers:
print("Tokenizer: {}".format(word_tokenizer.name))
result = word_tokenizer.tokenize(sentence)
result = [str(r) for r in result]
print(" ".join(result))
print()
word_tokenizer = WordTokenizer("whitespace")
sentence = "私 は 猫 だ ニャン"
print(word_tokenizer.tokenize(sentence))
from konoha import WordTokenizer
from utils_nlp.dataset.livedoor import load_pandas_df
if __name__ == '__main__':
df = load_pandas_df(nrows=10)
text = df['text'][0][:30]
print(text)
# 友人代表のスピーチ、独女はどうこなしている？もうすぐジューン
tokenizer_m = WordTokenizer('MeCab')
print(tokenizer_m.tokenize(text))
# [友人, 代表, の, スピーチ, 、, 独, 女, は, どう, こなし, て, いる, ？, もうすぐ, ジューン]
tokenizer_s = WordTokenizer('Sudachi', mode='A', with_postag=True)
print(tokenizer_s.tokenize(text))
# [友人 (名詞), 代表 (名詞), の (助詞), スピーチ (名詞), 、 (補助記号), 独女 (名詞), は (助詞), どう (副詞), こなし (動詞), て (助詞), いる (動詞), ？ (補助記号), もう (副詞), すぐ (副詞), ジューン (名詞)]
df['sep_text'] = [tokenizer_m.tokenize(text) for text in df['text']]
print(df.head())
from konoha import WordTokenizer
import neologdn
import numpy as np
from utils_nlp.dataset.livedoor import load_pandas_df
from utils_nlp.features import scdv
from utils_nlp.models.pretrained_embeddings.word2vec import load_pretrained_vectors
if __name__ == '__main__':
df = load_pandas_df(nrows=10)
# Normalization
df['text'] = df['text'].apply(neologdn.normalize)
tokenizer = WordTokenizer('MeCab')
docs = np.array(
[map(str, tokenizer.tokenize(text)) for text in df['text']])
print(docs.shape)
# (10,)
word_vec = load_pretrained_vectors('data')
scdv = scdv.create(docs, word_vec, n_components=10)
print(scdv.shape)
# (10, 3000)
