def create_from_ud(cls, data_file_list, split_chars=True):
"""Initialize corpus from a path to a file in conllu format
split_chars: if true, split up multisyllabic words into characters.
"""
corpus = POSCorpus()
corpus.sentences = []
for data_file_path in data_file_list:
with open(data_file_path, "r", encoding="utf-8") as data_file:
data = data_file.read()
data = conllu.parse(data)
for token_list in data:
sentence = []
for token in token_list:
pos = token['upostag']
word = token['form']
if split_chars:
for char in word:
sentence.append({'char': chinese_converter.to_simplified(char), 'pos': pos})
else:
sentence.append({'word': chinese_converter.to_simplified(word), 'pos': pos})
if len(sentence) > 0:
corpus.sentences.append(sentence)
return corpus
def test_two_way(self):
with self.subTest("should get same result when translated both ways"):
trad = '皇后與國王在後面共同候車'
result = to_traditional(to_simplified(trad))
self.assertEqual(trad, result)
simp = '皇后与国王在后面共同候车后'
result = to_simplified(to_traditional(simp))
self.assertEqual(simp, result)
def process_data(label, text):
data = []
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese',
do_lower_case=True)
text = re.sub(
'[a-zA-Z0-9 ’!"#$%&\'()*+,-./:;<=>?@?★…【】《》？“”‘’！[\\]^_`{|}~]', '',
text)
text_split = re.split('。', text)
text_split_join = []
temp = ''
for i in text_split:
if len(temp) + len(i) < 511:
temp += i
else:
if len(temp) < 511:
text_split_join.append(temp)
temp = i
if temp and len(temp) < 511:
text_split_join.append(temp)
for text in text_split_join:
text = chinese_converter.to_simplified(text)
encoded_text = tokenizer(text,
padding=True,
return_tensors="pt",
add_special_tokens=True)
encoded_text = np.array(encoded_text["input_ids"])[0]
if label[0:2] == '[]':
label_cls = [0] * len(encoded_text)
else:
if type(label) == type(''):
label = label.split(', ')
label_cls = [0] * len(encoded_text)
for i in label:
i = i.split('\'')
i = chinese_converter.to_simplified(i[1])
encode_i = tokenizer(i,
padding=True,
return_tensors="pt",
add_special_tokens=True)
encode_i = np.array(encode_i['input_ids'][0][1:-1])
for j in range(0, len(encoded_text) - len(encode_i)):
if (encoded_text[j:j + len(encode_i)] == encode_i).all():
for k in range(0, len(encode_i)):
label_cls[j + k] = 1
data.append(
(torch.tensor(encoded_text), torch.tensor(np.array(label_cls))))
return data
def make_data(a, b):
data = []
model = BertModel.from_pretrained('bert-base-chinese')
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese', do_lower_case = True)
for i in range(a, b):
f = open("training_set/" + str(i) + ".txt", "r")
lines = f.readlines()
label = lines[0]
text = lines[1] if len(lines) == 2 else ''.join(lines[2:])
f.close()
text = re.sub('[a-zA-Z0-9 ’!"#$%&\'()*+,-./:;<=>?@?★…【】《》？“”‘’！[\\]^_`{|}~]','',text)
text_split = re.split('。', text)
text_split_join = []
temp = ''
for i in text_split:
if len(temp) + len(i) < 511:
temp += i
else:
if len(temp) < 511:
text_split_join.append(temp)
temp = i
if temp and len(temp) < 511:
text_split_join.append(temp)
for text in text_split_join:
text = chinese_converter.to_simplified(text)
encoded_text = tokenizer(text, padding=True, return_tensors="pt", add_special_tokens=True)#会自动添加ERNIE所需要的特殊token，如[CLS], [SEP]
encoded_text = np.array(encoded_text["input_ids"])[0]#[:MAX_SEQLEN]
#encoded_text = np.pad(encoded_text, (0, MAX_SEQLEN-len(encoded_text)), mode='constant') # 对所有句子都补长至11000，这样会比较费显存；
if label[0:2] == '[]':
label_cls = [0] * len(encoded_text)
else:
if type(label) == type(''):
label = label.split(', ')
label_cls = [0] * len(encoded_text)
for i in label:
i = i.split('\'')
i = chinese_converter.to_simplified(i[1])
encode_i = tokenizer(i, padding=True, return_tensors="pt", add_special_tokens=True)
encode_i = np.array(encode_i['input_ids'][0][1:-1])
for j in range(0, len(encoded_text)-len(encode_i)):
if (encoded_text[j:j+len(encode_i)] == encode_i).all():
for k in range(0, len(encode_i)):
label_cls[j+k] = 1
data.append((torch.tensor(encoded_text), torch.tensor(np.array(label_cls))))
return data
def inference(text, model):
data = []
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese',
do_lower_case=True)
text_split = re.split('。', text)
text_split_join = []
temp = ''
for i in text_split:
if len(temp) + len(i) < 511:
temp += i
else:
if len(temp) < 511:
text_split_join.append(temp)
temp = i
if temp and len(temp) < 511:
text_split_join.append(temp)
for text in text_split_join:
text = re.sub(
'[a-zA-Z0-9 ’!"#$%&\'()*+,-./:;<=>?@?★…【】《》？“”‘’！[\\]^_`{|}~]', '',
text)
text_convert = chinese_converter.to_simplified(text)
encoded_text = tokenizer(text_convert,
padding=True,
return_tensors="pt",
add_special_tokens=True)
encoded_text = np.array(encoded_text["input_ids"])[0]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
outputs = model(input_ids=torch.tensor([encoded_text]).to((device)))
logits = outputs[0][0][1:-1]
index = 0
for i in range(0, len(logits)):
if logits[i][1] > logits[i][0]:
if data and i == index + 1:
data[-1] += text[i]
else:
data.append(text[i])
index = i
data = set(data)
res = []
for i in data:
if 2 <= len(i) <= 4:
res.append(i)
return res
def testing(text, model):
data = []
tokenizer = BertTokenizer.from_pretrained('bert-base-chinese', do_lower_case = True)
text_split = re.split('。', text)
text_split_join = []
temp = ''
for i in text_split:
if len(temp) + len(i) < 511:
temp += i
else:
if len(temp) < 511:
text_split_join.append(temp)
temp = i
if temp and len(temp) < 511:
text_split_join.append(temp)
for text in text_split_join:
text = re.sub('[a-zA-Z0-9 ’!"#$%&\'()*+,-./:;<=>?@?★…【】《》？“”‘’！[\\]^_`{|}~]','',text)
# print(len(text),text)
text_convert = chinese_converter.to_simplified(text)
encoded_text = tokenizer(text_convert, padding=True, return_tensors="pt", add_special_tokens=True)
encoded_text = np.array(encoded_text["input_ids"])[0]
# print(encoded_text)
# predicted_index = encoded_text
# predicted_token = [tokenizer.convert_ids_to_tokens([predicted_index[x]])[0] for x in
# range(1, (len(encoded_text) - 1))]
# print(predicted_token)
outputs = model(input_ids=torch.tensor([encoded_text]).to('cuda:0'))
logits = outputs[0][0][1:-1]
# print(len(logits))
index = 0
for i in range(0, len(logits)):
if logits[i][1] > logits[i][0]:
if data and i == index + 1:
data[-1] += text[i]
else:
data.append(text[i])
index = i
data = set(data)
res = []
for i in data:
if 2 <= len(i) <= 4:
res.append(i)
print(res)
return set(data)
def _to_simplified(self, x):
return [cc.to_simplified(proc) for proc in x]
def test_to_simplified(self):
result = to_simplified('皇后與國王在後面共同候車吃麵')
self.assertEqual(result, '皇后与国王在后面共同候车吃面')
def test_other_language(self):
text = "this is a book."
with self.subTest("should not change text in other languages"):
self.assertEqual(text, to_simplified(text))
self.assertEqual(text, to_traditional(text))
