def eval_sighan_2015_by_bert(sighan_path=sighan_2015_path,
verbose=True,
num_limit_lines=100):
from pycorrector.bert.bert_corrector import BertCorrector
model = BertCorrector()
total_count = 0
right_count = 0
right_rate = 0.0
recall_rate = 0.0
recall_right_count = 0
recall_total_count = 0
start_time = time.time()
with open(sighan_path, 'r', encoding='utf-8') as f:
for line in f:
line = line.strip()
if line.startswith('#'):
continue
parts = line.split()
if len(parts) != 2:
continue
if 0 < num_limit_lines < total_count:
continue
src = parts[0]
trg = parts[1]
pred, pred_detail = model.bert_correct(src)
if src != trg:
recall_total_count += 1
if pred == trg:
right_count += 1
if src != trg:
recall_right_count += 1
if verbose:
print("\nright:")
print(
f'input : {src}\ntruth : {trg}\npredict: {pred} pred_detail: {pred_detail}'
)
else:
if verbose:
print("\nwrong:")
print(
f'input : {src}\ntruth : {trg}\npredict: {pred} pred_detail: {pred_detail}'
)
total_count += 1
spend_time = time.time() - start_time
if total_count > 0:
right_rate = right_count / total_count
if recall_total_count > 0:
recall_rate = recall_right_count / recall_total_count
print(
'right_rate:{}, right_count:{}, total_count:{};\n'
'recall_rate:{}, recall_right_count:{}, recall_total_count:{}, spend_time:{} s'
.format(right_rate, right_count, total_count, recall_rate,
recall_right_count, recall_total_count, spend_time))
def eval_corpus500_by_bert(input_eval_path=eval_data_path, output_eval_path='', verbose=True):
from pycorrector.bert.bert_corrector import BertCorrector
model = BertCorrector()
res = []
corpus = load_json(input_eval_path)
total_count = 0
right_count = 0
right_rate = 0.0
recall_rate = 0.0
recall_right_count = 0
recall_total_count = 0
start_time = time.time()
for data_dict in corpus:
text = data_dict.get('text', '')
correction = data_dict.get('correction', '')
errors = data_dict.get('errors', [])
# pred_detail: list(wrong, right, begin_idx, end_idx)
pred_sentence, pred_detail = model.bert_correct(text)
# compute recall
if errors:
recall_total_count += 1
if errors and pred_detail and correction == pred_sentence:
recall_right_count += 1
# compute precision
if correction == pred_sentence:
right_count += 1
print("\nright:")
print('truth :', text, errors)
print('predict:', pred_sentence, pred_detail)
else:
err_data_dict = copy.deepcopy(data_dict)
err_data_dict['pred_sentence'] = pred_sentence
err_data_dict['pred_errors'] = str(pred_detail)
res.append(err_data_dict)
if verbose:
print("\nwrong:")
print('input :', text)
print('truth :', correction, errors)
print('predict:', pred_sentence, pred_detail)
total_count += 1
spend_time = time.time() - start_time
if total_count > 0:
right_rate = right_count / total_count
if recall_total_count > 0:
recall_rate = recall_right_count / recall_total_count
print('right_rate:{}, right_count:{}, total_count:{};\n'
'recall_rate:{}, recall_right_count:{}, recall_total_count:{}, spend_time:{} s'.format(right_rate,
right_count,
total_count,
recall_rate,
recall_right_count,
recall_total_count,
spend_time))
if output_eval_path:
save_json(res, output_eval_path)
def main(args):
if args.data == 'sighan_15' and args.model == 'rule':
demo()
# Sentence Level: acc:0.173225, precision:0.979592, recall:0.148541, f1:0.257965, cost time:230.92 s
eval.eval_sighan2015_by_model(pycorrector.correct)
if args.data == 'sighan_15' and args.model == 'bert':
# right_rate:0.37623762376237624, right_count:38, total_count:101;
# recall_rate:0.3645833333333333, recall_right_count:35, recall_total_count:96, spend_time:503 s
from pycorrector.bert.bert_corrector import BertCorrector
model = BertCorrector()
eval.eval_sighan2015_by_model(model.bert_correct)
if args.data == 'sighan_15' and args.model == 'macbert':
# Sentence Level: acc:0.914885, precision:0.995199, recall:0.916446, f1:0.954200, cost time:29.47 s
from pycorrector.macbert.macbert_corrector import MacBertCorrector
model = MacBertCorrector()
eval.eval_sighan2015_by_model(model.macbert_correct)
if args.data == 'sighan_15' and args.model == 'ernie':
# right_rate:0.297029702970297, right_count:30, total_count:101;
# recall_rate:0.28125, recall_right_count:27, recall_total_count:96, spend_time:655 s
from pycorrector.ernie.ernie_corrector import ErnieCorrector
model = ErnieCorrector()
eval.eval_sighan2015_by_model(model.ernie_correct)
if args.data == 'corpus500' and args.model == 'rule':
demo()
# right_rate:0.486, right_count:243, total_count:500;
# recall_rate:0.18, recall_right_count:54, recall_total_count:300, spend_time:78 s
eval.eval_corpus500_by_model(pycorrector.correct)
if args.data == 'corpus500' and args.model == 'bert':
# right_rate:0.586, right_count:293, total_count:500;
# recall_rate:0.35, recall_right_count:105, recall_total_count:300, spend_time:1760 s
from pycorrector.bert.bert_corrector import BertCorrector
model = BertCorrector()
eval.eval_corpus500_by_model(model.bert_correct)
if args.data == 'corpus500' and args.model == 'macbert':
# Sentence Level: acc:0.724000, precision:0.912821, recall:0.595318, f1:0.720648, cost time:6.43 s
from pycorrector.macbert.macbert_corrector import MacBertCorrector
model = MacBertCorrector()
eval.eval_corpus500_by_model(model.macbert_correct)
if args.data == 'corpus500' and args.model == 'ernie':
# right_rate:0.598, right_count:299, total_count:500;
# recall_rate:0.41333333333333333, recall_right_count:124, recall_total_count:300, spend_time:6960 s
from pycorrector.ernie.ernie_corrector import ErnieCorrector
model = ErnieCorrector()
eval.eval_corpus500_by_model(model.ernie_correct)
def error(data_json):
with open(data_json, 'r', encoding='utf-8') as f:
data_list = json.load(f)
# model mode
d = BertCorrector()
error_list = []
error_detail = []
for data in data_list:
file_type = data.get('file_type')
file_no = data.get('file_no')
for s in data.get('sentences'):
if s.get('type') == "正文" or (file_type != '3' and s.get('type') == '标题'):
paragraph_no = s.get('paragraph_no')
sentence_no = s.get('sentence_no')
text = s.get('text')
# model mode
corrected_sent, details = d.bert_correct(text)
# corrected_sent, details = pycorrector.correct(text)
new_details = []
# details : ['七', '器', 1, 2]
for detail in details:
new_details.append(find_diff(detail))
orign = detail[0]
fix = detail[1]
s = difflib.SequenceMatcher(None, orign, fix)
for tag, i1, i2, j1, j2 in s.get_opcodes():
if tag == 'replace':
error = '%s%s%s%s%s%s1001,%s' % (file_type,
file_no,
paragraph_no,
sentence_no,
str(detail[2] + 1 + i1).zfill(2),
str(detail[2] + i2).zfill(2),
fix[j1:j2]
)
error_list.append(error)
error_detail.append((error, corrected_sent, new_details))
return error_list, error_detail
def error(data_json):
with open(data_json, 'r', encoding='utf-8') as f:
data_list = json.load(f)
# model mode
d = BertCorrector()
error_list = []
error_detail = []
for data in data_list:
file_type = data.get('file_type')
file_no = data.get('file_no')
first_type = None
first_no = None
local_part = 0
for s in data.get('sentences'):
if s.get('type') == "正文" or (file_type != '3'
and s.get('type') == '标题'):
paragraph_no = s.get('paragraph_no')
sentence_no = s.get('sentence_no')
text = s.get('text')
list_text = []
if s.get('type') == '标题':
spilt_text = text.split("\n")
for item in spilt_text:
list_text.append(item)
else:
list_text.append(text)
for i, item in enumerate(list_text):
# model mode
corrected_sent, details = d.bert_correct(item)
# corrected_sent, details = pycorrector.correct(text)
if first_type != s.get('type') or first_no != paragraph_no:
local_part = 0
first_type = s.get('type')
first_no = paragraph_no
new_details = []
# details : ['七', '器', 1, 2]
for detail in details:
new_details.append(find_diff(detail))
orign = detail[0]
fix = detail[1]
ds = difflib.SequenceMatcher(None, orign, fix)
for tag, i1, i2, j1, j2 in ds.get_opcodes():
if tag == 'replace':
part_text_1 = item[:detail[2] + i1 + 1]
part_text_2 = item[:detail[2] + i2 + 1]
add_1 = seg_char(part_text_1)
add_2 = seg_char(part_text_2)
error = '%s%s%s%s%s%s1,%s' % (
file_type,
file_no,
str(int(paragraph_no) + i).zfill(2),
str(len(add_1) + local_part).zfill(3),
# str(detail[2] + 1 + i1 + local_part).zfill(3),
str(int(paragraph_no) + i).zfill(2),
str(len(add_2) - 1 + local_part).zfill(3),
# str(detail[2] + i2 + local_part).zfill(3),
fix[j1:j2])
error_list.append(error)
error_detail.append(
(error, corrected_sent, new_details))
return error_list, error_detail
# -*- coding: utf-8 -*- # @Author : Dapeng # @File : test.py # @Desc : # @Contact : [email protected] # @Time : 2020/10/12 下午6:59 from pycorrector.bert.bert_corrector import BertCorrector d = BertCorrector() error_sentences = [ '疝気医院那好 为老人让坐，疝気专科百科问答', '少先队员因该为老人让坐', '少 先 队 员 因 该 为 老人让坐', '机七学习是人工智能领遇最能体现智能的一个分知', '今天心情很好', ] for sent in error_sentences: corrected_sent, err = d.bert_correct(sent) print("original sentence:{} => {}, err:{}".format(sent, corrected_sent, err))
'少先队员因该为老人让坐\r\n少 先 队 员 因 该 为 老人让坐\r\n机七学习是人工智能领遇最能体现智能的一个分知\r\n到以深切的问候',
style="margin: 0px; width: 310px; height: 187px;"),
# form.Textbox("bax",
# form.notnull,
# form.regexp('\d+', 'Must be a digit'),
# form.Validator('Must be more than 5', lambda x:int(x)>5)),
form.Textarea(name='output_sentences',
style="margin: 0px; width: 310px; height: 187px;"),
# form.Checkbox('curly'),
# form.Dropdown('french', ['mustard', 'fries', 'wine'])
)
RetFrom = form.Form(form.Textarea(name='output_sentences'))
from pycorrector.bert.bert_corrector import BertCorrector
d = BertCorrector()
class index:
def GET(self):
form = InputForm()
# 确保通过调用它来创建表单的副本（上面一行）
# 否则更改将全局显示
# return form.read()
return render.formtest(form)
def POST(self):
inputform = InputForm()
if not inputform.validates():
return render.formtest(inputform)
else:
