-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathchinese_bert2.py
229 lines (207 loc) · 12.3 KB
/
chinese_bert2.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
import argparse
import os
from scipy.special import softmax
from transformers import AutoTokenizer, BertTokenizer, BertForMaskedLM
import torch
import torch.nn.functional as F
import json
from collections import Counter
from pathlib import Path
from transformers import pipeline
from tqdm import tqdm
from transformers import logging
logging.set_verbosity_error()
local_f1 = Path('data/local_female.txt').read_text().strip().split('\n')
local_m1 = Path('data/local_male.txt').read_text().strip().split('\n')
amb_f1 = Path('data/amb_f1.txt').read_text().strip().split('\n')
amb_m1 = Path('data/amb_m1.txt').read_text().strip().split('\n')
verb_f1 = Path('data/verb_f1.txt').read_text().strip().split('\n')
verb_m1 = Path('data/verb_m1.txt').read_text().strip().split('\n')
in_verb_f1 = Path('data/in_verb_f1.txt').read_text().strip().split('\n')
in_verb_m1 = Path('data/in_verb_m1.txt').read_text().strip().split('\n')
blocking = Path('data/blocking_literature.txt').read_text().strip().split('\n')
animacy_pro = Path('data/inanimate_pron.txt').read_text().strip().split('\n')
animacy_noun = Path('data/inanimate_nouns.txt').read_text().strip().split('\n')
subj_f1 = Path('data/subject_orientation_f1.txt').read_text().strip().split('\n')
subj_m1 = Path('data/subject_orientation_m1.txt').read_text().strip().split('\n')
subj_f1_bias = Path('data/subject_orientation_f1_bias.txt').read_text().strip().split('\n')
subj_m1_bias = Path('data/subject_orientation_m1_bias.txt').read_text().strip().split('\n')
natural_local_m = Path('data/filtered_sents_local_m_binding.txt').read_text().strip().split('\n')
natural_local_f = Path('data/filtered_sents_local_f_binding.txt').read_text().strip().split('\n')
natural_local_verb = Path('data/real_data_lb_verb.txt').read_text().strip().split('\n')
natural_long_verb = Path('data/real_data_ldb_verb.txt').read_text().strip().split('\n')
natural_long_anim = Path('data/real_data_ldb_anim.txt').read_text().strip().split('\n')
# Tokenize input
def get_probability(zh_sents, output, task=None, local=True, antecedent=None, antecedent_list =None, verbose=False):
# Get logits from the model
nlp = pipeline("fill-mask", model=model)
mask = nlp.tokenizer.mask_token
target_dic = {'她': 'f', '他': 'm', '我': 'w', '它': 't', '你': 'n'}
with open(output, 'w', encoding="utf-8") as out_tsv:
c = 0
target_dic = {key:value for key, value in target_dic.items() if value in antecedent_list}
target = [key for key, _ in target_dic.items()]
for s in tqdm(zh_sents):
if task == 'syntax':
ziji_index = s.index('自')
text = f'{s[:ziji_index-1]}{mask}{s[ziji_index:]}'
elif task == 'syntax2':
ziji_index = s.index('自')
text = f'{s[:ziji_index]}{mask}{s[ziji_index:]}'
elif task == 'subject_orientation':
de_id = s.index('的')
# end_id = s.index('是')
end_id = len(s)-1
text = f'假如{s[:-1]}， 那么{s[de_id + 1:end_id]}是关于{mask}的。'
else:
if local:
text = f'假如{s[:-1]},那么{s[2:-3]}{mask}自己。'
else:
text = f'假如{s[:-1]},那么{s[2:-3]}{mask}。'
# text = f'如果{s[:-1]},那么{mask}{s[-5:]}'
predictions = nlp(text, targets=target)
scores = softmax([x['score'] for x in predictions])
p_antecedents = [target_dic[x['token_str']] for x in predictions]
all_prob = {x:y for x, y in zip(p_antecedents, scores)}
prob_to_write = '\t'.join([x+':'+str(y) for x, y in all_prob.items()])
out_tsv.write(f'{text}\t{prob_to_write}\n')
all_prob = sorted(all_prob.items(), key=lambda x: x[1], reverse=True)
if verbose:
print(text, all_prob)
if antecedent=='w':
if all_prob[0][0] == 'w':
c += 1
elif antecedent=='f':
if all_prob[0][0] == 'f':
c += 1
elif antecedent == 'm':
if all_prob[0][0] == 'm':
c += 1
elif antecedent == 'n':
if all_prob[0][0] == 'n':
c += 1
else:
if all_prob[0][0] != 't':
c += 1
print(f'{c}\t{len(zh_sents)}\t{c / len(zh_sents)}')
return c, len(zh_sents)
def test_real_data(input_file, output_file, task =None, verbose=False):
input_sents = [x.split() for x in Path(input_file).read_text().strip().split('\n')]
nlp = pipeline("fill-mask", model=model)
old_target_dic = {'她': 'f', '他': 'm', '我': 'w', '它': 't', 'n': '你'}
mask = nlp.tokenizer.mask_token
with open(output_file, 'w', encoding='utf-8') as out_tsv:
c = 0
for s in tqdm(input_sents):
text = f'如果{s[0][:-1]}，那么{s[2][:-2]}{mask}。'
freq_char = Counter(text)
if task =='animacy':
to_add_antecedent = ['她','他','它']
elif task == 'blocking':
to_add_antecedent = ['她','他','我']
else:
to_add_antecedent = ['她', '他']
antecedent_list =list(set([x for x, y in old_target_dic.items() if freq_char[x]>0 or x in to_add_antecedent]))
target_dic = {x:y for x, y in old_target_dic.items() if x in antecedent_list}
label2target = {y:x for x, y in target_dic.items()}
target = [x for x, _ in target_dic.items()]
predictions = nlp(text, targets=target)
scores = softmax([x['score'] for x in predictions])
p_antecedents = [target_dic[x['token_str']] for x in predictions]
all_prob = {x: y for x, y in zip(p_antecedents, scores)}
prob_to_write = '\t'.join([x + ':' + str(y) for x, y in all_prob.items()])
out_tsv.write(f'{text}\t{prob_to_write}\n')
all_prob = sorted(all_prob.items(), key=lambda x: x[1], reverse=True)
pred = label2target[all_prob[0][0]]
if pred == s[1]:
c += 1
else:
if verbose:
print(text)
print(all_prob)
else:
pass
print(f'{c}\t{len(input_sents)}\t{c / len(input_sents)}')
return c, len(input_sents)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-m', '--model', type=str, help='the evaluated model, available options: [chinesebert, xlm, mbert]', required=True)
args = parser.parse_args()
try:
os.mkdir(f'result/{args.model}')
except:
pass
if args.model =='chinesebert':
model = 'google-bert/bert-base-chinese'
elif args.model =='xlm':
model = 'xlm-roberta-base'
elif args.model == 'mbert':
model = 'google-bert/bert-base-multilingual-cased'
elif args.model == 'xlm-large':
model = 'xlm-roberta-large'
elif args.model == 'roberta':
model = 'hfl/chinese-roberta-wwm-ext-large'
elif args.model == 'lert':
model = 'hfl/chinese-lert-base'
elif args.model =='lert-large':
model = 'hfl/chinese-lert-large'
elif args.model =='bert-wwm':
model = 'hfl/chinese-bert-wwm-ext'
else:
raise ValueError('invalid model name!')
# print('========================REAL DATA==========================================')
# print('real data: local binding, female binder')
# c1, all1 = get_probability(natural_local_f, f'result/{args.model}/natural_local_f1.tsv', 'syntax', antecedent='f', antecedent_list=['f','m'])
# print('real data: local binding, male binder')
# c2, all2 = get_probability(natural_local_m, f'result/{args.model}/natural_local_m1.tsv', 'syntax', antecedent='m', antecedent_list=['f','m'])
#
# print('real data: reflexive verb, local binding')
# c3, all3 = test_real_data('data/real_data_lb_verb.txt', f'result/{args.model}/lb_name.tsv')
# print('real data: non-reflexive verb, long-distance binding')
# c4, all4 = test_real_data('data/real_data_ldb_verb.txt', f'result/{args.model}/ldb_name.tsv')
# print('real data: animacy effect, long-distance binding')
# c5, all5 = test_real_data('data/real_data_ldb_anim.txt', f'result/{args.model}/ldb_anim.tsv', task = 'animacy')
#
# print('real data: blocking effect, long-distance binding')
# c20, all20 = test_real_data('data/real_data_blocking.txt', f'result/{args.model}/natural_blocking.tsv', task='blocking')
#
#
# real_c = c1+c2+c3+c4+c5+c20
# real_all = all1+all2+all3+all4+all5+all20
# print('++++++++++++++++++++++++OVERALL+++++++++++++++++++++++++++++++++++++++++')
# print(f'{real_c}\t{real_all}\t{real_c/real_all}')
print('========================SYNTHETIC DATA======================================')
print('In the local binding setting, the percentage of local binding is: ')
c6, all6 = get_probability(local_f1, f'result/{args.model}/local_f1.tsv', 'syntax2', antecedent='f', antecedent_list=['f','m'])
c7, all7 =get_probability(local_m1, f'result/{args.model}/local_m1.tsv', 'syntax2', antecedent='m', antecedent_list=['f','m'])
print('In ambiguous setting, the percentage of local binding:')
c8, all8 =get_probability(amb_f1, f'result/{args.model}/amb_f1.tsv','syntax2', antecedent='f', antecedent_list=['f','m'])
c9, all9 =get_probability(amb_m1, f'result/{args.model}/amb_m1.tsv', 'syntax2',antecedent='m', antecedent_list=['f','m'])
print((c8 + c9) / (all8 + all9))
print('In externally oriented verb setting, the percentage of local binding:')
c10, all10 =get_probability(verb_f1, f'result/{args.model}/verb_f1.tsv', local=False, antecedent='f', antecedent_list=['f','m'])
c11, all11 =get_probability(verb_m1, f'result/{args.model}/verb_m1.tsv', local=False, antecedent='m', antecedent_list=['f','m'])
print((c10 + c11) / (all10 + all11))
print('In internally oriented verb setting, the percentage of local binding:')
c12, all12 =get_probability(in_verb_f1, f'result/{args.model}/in_verb_f1.tsv', antecedent='f', antecedent_list=['f','m'])
c13, all13 =get_probability(in_verb_m1, f'result/{args.model}/in_verb_m1.tsv', antecedent='m', antecedent_list=['f','m'])
print((c12 + c13) / (all12 + all13))
print('In the blocking effect setting, the percentage of local binding:')
c14, all14 =get_probability(blocking, f'result/{args.model}/blocking.tsv', 'syntax2', antecedent='w', antecedent_list=['f','m', 'w'])
print('In animate setting, the percentage of long-distance binding:')
c15, all15 =get_probability(animacy_noun, f'result/{args.model}/animacy_noun.tsv', local= False, antecedent='t', antecedent_list=['f','m', 't'])
print('In subject orientation, the percentage of local binding:')
c16, all16 =get_probability(subj_f1, f'result/{args.model}/subj_f1.tsv', 'subject_orientation', antecedent='f', antecedent_list=['f','m'])
c17, all17 =get_probability(subj_m1, f'result/{args.model}/subj_m1.tsv', 'subject_orientation', antecedent='m', antecedent_list=['f','m'])
# print('In subject orientation in a gender-biased setting, the percentage of local binding:')
# c18, all18 =get_probability(subj_f1_bias, f'result/{args.model}/subj_f1_bias.tsv', 'subject_orientation', antecedent='f', antecedent_list=['f','m'])
# c19, all19 =get_probability(subj_m1_bias, f'result/{args.model}/subj_m1_bias.tsv', 'subject_orientation', antecedent='m', antecedent_list=['f','m'])
# print(f'{(c16+c18)/(all16+all18)}\t{(c17+c19)/(all17+all19)}')
syn_c = c6+c7+c8+c9+c10+c11+c12+c13+c14+c15+c16+c17
syn_all = all6+all7+all8+all9+all10+all11+all12+all13+all14+all15+all16+all17
print('+++++++++++++++++++++++OVERALL++++++++++++++++++++++++++')
print(f'{syn_c}\t{syn_all}\t{syn_c/syn_all}')
# print(
# f'{round((c1 / all1) * 100, 1)}&{round((c2 / all2) * 100, 1)}&{round((c4 / all4) * 100, 1)}&{round((c3 / all3) * 100, 1)}&{round((c20 / all20), 1)}&{round((c5 / all5) * 100, 1)}&{round((real_c / real_all) * 100, 1)}')
print(
f'{round((c6 / all6) * 100, 1)}&{round((c7 / all7) * 100, 1)}&{round(((c10 + c11) / (all10 + all11)) * 100, 1)}&{round(((c12 + c13) / (all12 + all13)) * 100, 1)}&{round((c14 / all14) * 100, 1)}&{round(((c8+c9) / (all8+all9)) * 100, 1)}&{round((c15 / all15) * 100, 1)}&{round((c16 / all16) * 100, 1)}&{round((c17 / all17) * 100, 1)}&{round((syn_c / syn_all) * 100, 1)}')