-
Notifications
You must be signed in to change notification settings - Fork 0
/
preprocessing.py
259 lines (206 loc) · 11 KB
/
preprocessing.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
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
import json
import os
import numpy as np
import codecs
import tokenization
def parse_json(json_path):
ret_sents, ret_labels = dict(), dict()
with codecs.open(json_path, 'r', encoding='ascii') as json_file:
json_data = json.load(json_file)
for dataset in json_data:
ret_sents[dataset] = []
ret_labels[dataset] = []
for example in json_data[dataset]:
ret_sents[dataset].append(example[0].lower())
ret_labels[dataset].append(example[1].lower())
return ret_sents, ret_labels
def get_squad_queries(squad_path, tokenizer):
queries = []
with codecs.open(squad_path, 'r', encoding='utf=8') as json_file:
json_data = json.load(json_file)
for examples in json_data['data']:
for paragraph in examples['paragraphs']:
for qa in paragraph['qas']:
question = qa['question'].lower()
tokens = tokenizer.tokenize(question)
if len(tokens) >= 150:
continue
queries.append(question)
print('squad has '+str(len(queries))+' queries')
np.random.seed(0)
shuffled_ids = np.arange(len(queries))
queries = np.array(queries)[shuffled_ids]
return queries[:10000]
def write_to_oos_bin_oversampled(output_dir, ret_sents, ret_labels, squad_queries):
# write for binary classification: is [in-scope], oos [out-of-scope]
if not os.path.exists(output_dir + ''):
os.mkdir(output_dir + '')
train_sents = np.array(ret_sents['train'] + ret_sents['oos_train'] + list(squad_queries))
train_labels = np.array(['is'] * len(ret_sents['train']) + ['oos'] * len(ret_sents['oos_train'])
+ ['oos'] * len(squad_queries))
np.random.seed(0)
shuffled_ids = np.arange(len(train_sents))
np.random.shuffle(shuffled_ids)
train_sents = train_sents[shuffled_ids]
train_labels = train_labels[shuffled_ids]
print(str(np.sum(train_labels == 'oos')))
print(str(np.sum(train_labels == 'is')))
with open(output_dir + '/sentences.train.in', 'w') as output:
output.write('\n'.join(train_sents))
with open(output_dir + '/sentences.train.out', 'w') as output:
for label in train_labels:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.eval.in', 'w') as output:
output.write('\n'.join(ret_sents['val'] + ret_sents['oos_val']))
with open(output_dir + '/sentences.eval.out', 'w') as output:
for label in ['is'] * len(ret_sents['val']) + ['oos'] * len(ret_sents['oos_val']):
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.test.in', 'w') as output:
output.write('\n'.join(ret_sents['test'] + ret_sents['oos_test']))
with open(output_dir + '/sentences.test.out', 'w') as output:
for label in ['is'] * len(ret_sents['test']) + ['oos'] * len(ret_sents['oos_test']):
output.write('_'.join(label.split()) + '\n')
def write_to_oos_bin(output_dir, ret_sents, ret_labels):
# write for binary classification: is [in-scope], oos [out-of-scope]
if not os.path.exists(output_dir + ''):
os.mkdir(output_dir + '')
train_sents = np.array(ret_sents['train'] + ret_sents['oos_train'])
train_labels = np.array(['is'] * len(ret_sents['train']) + ['oos'] * len(ret_sents['oos_train']))
np.random.seed(0)
shuffled_ids = np.arange(len(train_sents))
np.random.shuffle(shuffled_ids)
train_sents = train_sents[shuffled_ids]
train_labels = train_labels[shuffled_ids]
with open(output_dir + '/sentences.train.in', 'w') as output:
output.write('\n'.join(train_sents))
with open(output_dir + '/sentences.train.out', 'w') as output:
for label in train_labels:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.eval.in', 'w') as output:
output.write('\n'.join(ret_sents['val'] + ret_sents['oos_val']))
with open(output_dir + '/sentences.eval.out', 'w') as output:
for label in ['is'] * len(ret_sents['val']) + ['oos'] * len(ret_sents['oos_val']):
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.test.in', 'w') as output:
output.write('\n'.join(ret_sents['test'] + ret_sents['oos_test']))
with open(output_dir + '/sentences.test.out', 'w') as output:
for label in ['is'] * len(ret_sents['test']) + ['oos'] * len(ret_sents['oos_test']):
output.write('_'.join(label.split()) + '\n')
def write_to_oos_bin_downsampled(output_dir, ret_sents, ret_labels):
# write for binary classification: is [in-scope], oos [out-of-scope]
if not os.path.exists(output_dir + ''):
os.mkdir(output_dir + '')
train_sents = np.array(ret_sents['train'] + ret_sents['oos_train'])
train_labels_orig = np.array(ret_labels['train'] + ret_labels['oos_train'])
np.random.seed(0)
shuffled_ids = np.arange(len(train_sents))
np.random.shuffle(shuffled_ids)
train_sents = train_sents[shuffled_ids]
train_labels_orig = train_labels_orig[shuffled_ids]
down_sample_per_class = 6
is_labels = set(ret_labels['train'])
sampled_num = {label: 0 for label in is_labels}
downsampled_train_sents, downsampled_train_labels = [], []
for ind, sent in enumerate(train_sents):
if train_labels_orig[ind] in sampled_num:
if sampled_num[train_labels_orig[ind]] < down_sample_per_class:
downsampled_train_sents.append(sent)
downsampled_train_labels.append('is')
sampled_num[train_labels_orig[ind]] += 1
else:
downsampled_train_sents.append(sent)
downsampled_train_labels.append('oos')
np.random.seed(0)
shuffled_ids = np.arange(len(downsampled_train_sents))
np.random.shuffle(shuffled_ids)
downsampled_train_sents = np.array(downsampled_train_sents)[shuffled_ids]
downsampled_train_labels = np.array(downsampled_train_labels)[shuffled_ids]
with open(output_dir + '/sentences.train.in', 'w') as output:
output.write('\n'.join(downsampled_train_sents))
with open(output_dir + '/sentences.train.out', 'w') as output:
for label in downsampled_train_labels:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.eval.in', 'w') as output:
output.write('\n'.join(ret_sents['val'] + ret_sents['oos_val']))
with open(output_dir + '/sentences.eval.out', 'w') as output:
for label in ['is'] * len(ret_sents['val']) + ['oos'] * len(ret_sents['oos_val']):
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.test.in', 'w') as output:
output.write('\n'.join(ret_sents['test'] + ret_sents['oos_test']))
with open(output_dir + '/sentences.test.out', 'w') as output:
for label in ['is'] * len(ret_sents['test']) + ['oos'] * len(ret_sents['oos_test']):
output.write('_'.join(label.split()) + '\n')
def write_to_oos_inscope(output_dir, ret_sents, ret_labels):
if not os.path.exists(output_dir + ''):
os.mkdir(output_dir + '')
train_sents = np.array(ret_sents['train'])
train_labels = np.array(ret_labels['train'])
np.random.seed(0)
shuffled_ids = np.arange(len(train_sents))
np.random.shuffle(shuffled_ids)
train_sents = train_sents[shuffled_ids]
train_labels = train_labels[shuffled_ids]
with open(output_dir + '/sentences.train.in', 'w') as output:
output.write('\n'.join(train_sents))
with open(output_dir + '/sentences.train.out', 'w') as output:
for label in train_labels:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.eval.in', 'w') as output:
output.write('\n'.join(ret_sents['val']))
with open(output_dir + '/sentences.eval.out', 'w') as output:
for label in ret_labels['val']:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.test.in', 'w') as output:
output.write('\n'.join(ret_sents['test']))
with open(output_dir + '/sentences.test.out', 'w') as output:
for label in ret_labels['test']:
output.write('_'.join(label.split()) + '\n')
def write_to_oos_train(output_dir, ret_sents, ret_labels, label_set_oos_train_to_ids):
if not os.path.exists(output_dir + ''):
os.mkdir(output_dir + '')
train_sents = np.array(ret_sents['train'] + ret_sents['oos_train'])
train_labels = np.array(ret_labels['train'] + ret_labels['oos_train'])
np.random.seed(0)
shuffled_ids = np.arange(len(train_sents))
np.random.shuffle(shuffled_ids)
train_sents = train_sents[shuffled_ids]
train_labels = train_labels[shuffled_ids]
with open(output_dir + '/sentences.train.in', 'w') as output:
output.write('\n'.join(train_sents))
with open(output_dir + '/sentences.train.out', 'w') as output:
for label in train_labels:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.eval.in', 'w') as output:
output.write('\n'.join(ret_sents['val'] + ret_sents['oos_val']))
with open(output_dir + '/sentences.eval.out', 'w') as output:
for label in ret_labels['val'] + ret_labels['oos_val']:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/sentences.test.in', 'w') as output:
output.write('\n'.join(ret_sents['test'] + ret_sents['oos_test']))
with open(output_dir + '/sentences.test.out', 'w') as output:
for label in ret_labels['test'] + ret_labels['oos_test']:
output.write('_'.join(label.split()) + '\n')
with open(output_dir + '/intent_to_ids.txt', 'w') as output:
for intent in label_set_oos_train_to_ids:
output.write(str('_'.join(intent.split())) + '\n')
sent_lens = []
for dataset in ret_sents:
sent_lens.extend([len(sent) for sent in ret_sents[dataset]])
print('max sent length: ' + str(max(sent_lens)))
if __name__ == '__main__':
ret_sents, ret_labels = parse_json('data/data_imbalanced.json')
label_set = []
for dataset in ret_labels:
label_set.extend(ret_labels[dataset])
label_set = set(label_set)
label_set_oos_train = list(label_set) + ['oos']
label_to_ids = {label: id for id, label in enumerate(list(label_set))}
label_set_oos_train_to_ids = {label: id for id, label in enumerate(label_set_oos_train)}
# write_to_oos_train('data/oos_train', ret_sents, ret_labels, label_set_oos_train_to_ids)
#
# write_to_oos_bin('data/oos_binary/bin', ret_sents, ret_labels)
# write_to_oos_inscope('data/oos_binary/in_scope', ret_sents, ret_labels)
# write_to_oos_bin_downsampled('data/oos_binary/down_sampled_bin', ret_sents, ret_labels)
tokenizer = tokenization.FullTokenizer(vocab_file='/Users/yxu132/data/bert/uncased_L-12_H-768_A-12/vocab.txt', do_lower_case=True)
squad_queries = get_squad_queries('/Users/yxu132/pub-repos/decaNLP/data/squad/train-v1.1.json', tokenizer)
write_to_oos_bin_oversampled('data/oos_binary/over_sampled_bin', ret_sents, ret_labels, squad_queries)