-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathbatch_generation.py
271 lines (227 loc) · 12.5 KB
/
batch_generation.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
260
261
262
263
264
265
266
267
268
269
270
271
# processing training and test data
# use entity pairs to find their co-occurred sentences in the corpus
# and then generate positive and negative training samples
import torch
import numpy as np
import torch.nn.functional as F
def process_training_data(sentences, rep_words, topic_hier, max_seq_length, ent_sent_index, tokenizer):
parent_list = [x for x in topic_hier if x != 'ROOT']
sentences_index = []
final_data = []
real_rep_words = {}
for key in rep_words:
real_rep_words[key] = []
print("collecting positive samples!")
for parent in parent_list:
for child in topic_hier[parent]:
# print(child)
count = 10
for b in rep_words[child]:
if b not in ent_sent_index:
continue
for a in rep_words[parent]:
if a not in ent_sent_index:
continue
cooccur = ent_sent_index[a].intersection(ent_sent_index[b])
if len(cooccur) > 0:
if a not in real_rep_words[parent]:
real_rep_words[parent].append(a)
if b not in real_rep_words[child]:
real_rep_words[child].append(b)
for sen in cooccur:
sentences_index.append(sen)
s = sentences[sen]
s = '[CLS] '+s
s = s.split(' ')
if a not in s or b not in s:
continue
p_index = s.index(a)
c_index = s.index(b)
s[p_index] = '[MASK]'
s[c_index] = '[MASK]'
s = ' '.join(s).replace('_', ' ').replace('-lrb-','(').replace('-rrb-',')').split(' ')
input_id = tokenizer.encode(' '.join(s))
tokened_text = tokenizer.tokenize(' '.join(s))
mask_id = [x for x in range(len(input_id)) if input_id[x]==103]
# if len(mask_id) < 2:
# print(' '.join(s))
# print(a)
# print(b)
if len(input_id) > max_seq_length:
if mask_id[1] - mask_id[0] >= max_seq_length:
continue
else:
input_id = input_id[mask_id[0]:mask_id[1]+1]
p_index = 0 if p_index<c_index else mask_id[1] - mask_id[0]
c_index = 0 if c_index<p_index else mask_id[1] - mask_id[0]
else:
p_index = mask_id[0] if p_index<c_index else mask_id[1]
c_index = mask_id[0] if c_index<p_index else mask_id[1]
input_id = torch.tensor(input_id)
r_sentence = F.pad(torch.tensor(input_id),(0,max_seq_length-len(input_id)), "constant", 0)
attention_mask = torch.cat((torch.ones_like(input_id), torch.zeros(max_seq_length-len(input_id), dtype=torch.int64)),dim=0)
p_mask = np.zeros((max_seq_length))
p_mask[p_index] = 1
c_mask = np.zeros((max_seq_length))
c_mask[c_index] = 1
final_data.append([r_sentence, p_mask, c_mask, attention_mask, 0])
final_data.append([r_sentence, c_mask, p_mask, attention_mask, 1])
pos_len = len(final_data)
print(f"positive data number: {pos_len}")
print("collecting negative samples from siblings!")
for parent in parent_list:
for child in topic_hier[parent]:
# print(child)
count = 10
for b in rep_words[child]:
if b not in ent_sent_index:
continue
for a in rep_words[child]:
if a == b:
continue
if a not in ent_sent_index:
continue
cooccur = ent_sent_index[a].intersection(ent_sent_index[b])
if len(cooccur) > 0:
for sen in cooccur:
if sen in sentences_index:
continue
if np.random.random(1) > 0.1:
continue
# sentences_index.append(sen)
s = sentences[sen]
s = '[CLS] '+s
s = s.split(' ')
if a not in s or b not in s:
continue
p_index = s.index(a)
c_index = s.index(b)
s[p_index] = '[MASK]'
s[c_index] = '[MASK]'
s = ' '.join(s).replace('_', ' ').replace('-lrb-','(').replace('-rrb-',')').split(' ')
input_id = tokenizer.encode(' '.join(s))
mask_id = [x for x in range(len(input_id)) if input_id[x]==103]
if len(input_id) > max_seq_length:
if mask_id[1] - mask_id[0] >= max_seq_length:
continue
else:
input_id = input_id[mask_id[0]:mask_id[1]+1]
p_index = 0 if p_index<c_index else mask_id[1] - mask_id[0]
c_index = 0 if c_index<p_index else mask_id[1] - mask_id[0]
else:
p_index = mask_id[0] if p_index<c_index else mask_id[1]
c_index = mask_id[0] if c_index<p_index else mask_id[1]
input_id = torch.tensor(input_id)
r_sentence = F.pad(torch.tensor(input_id),(0,max_seq_length-len(input_id)), "constant", 0)
attention_mask = torch.cat((torch.ones_like(input_id), torch.zeros(max_seq_length-len(input_id), dtype=torch.int64)),dim=0)
p_mask = np.zeros((max_seq_length))
p_mask[p_index] = 1
c_mask = np.zeros((max_seq_length))
c_mask[c_index] = 1
final_data.append([r_sentence, p_mask, c_mask, attention_mask, 2])
print(len(final_data))
print("collecting negative samples from corpus!")
while len(final_data) < pos_len * 2:
# if len(final_data) % 100 == 0:
# print(pos_len)
# print(len(final_data))
sen = np.random.choice(len(sentences))
# remove positive sentences
if sen in sentences_index:
continue
s = sentences[sen]
s = '[CLS] '+s
s = sentences[sen].split(' ')
#randomly choose pairs
entities = [x for x in s if "_" in x]
if len(entities) < 2:
continue
p_index, c_index = np.random.choice(len(entities),2)
while c_index == p_index:
# continue
c_index = np.random.choice(len(entities))
s[p_index] = '[MASK]'
s[c_index] = '[MASK]'
s = ' '.join(s).replace('_', ' ').replace('-lrb-','(').replace('-rrb-',')').split(' ')
input_id = tokenizer.encode(' '.join(s))
mask_id = [x for x in range(len(input_id)) if input_id[x]==103]
if len(input_id) > max_seq_length:
if mask_id[1] - mask_id[0] >= max_seq_length:
continue
else:
input_id = input_id[mask_id[0]:mask_id[1]+1]
p_index = 0 if p_index<c_index else mask_id[1] - mask_id[0]
c_index = 0 if c_index<p_index else mask_id[1] - mask_id[0]
else:
p_index = mask_id[0] if p_index<c_index else mask_id[1]
c_index = mask_id[0] if c_index<p_index else mask_id[1]
input_id = torch.tensor(input_id)
r_sentence = F.pad(torch.tensor(input_id),(0,max_seq_length-len(input_id)), "constant", 0)
attention_mask = torch.cat((torch.ones_like(input_id), torch.zeros(max_seq_length-len(input_id), dtype=torch.int64)),dim=0)
p_mask = np.zeros((max_seq_length))
p_mask[p_index] = 1
c_mask = np.zeros((max_seq_length))
c_mask[c_index] = 1
final_data.append([r_sentence, p_mask, c_mask, attention_mask, 2])
return final_data, sentences_index
def process_test_data(sentences, test_topic_rep_words, test_cand, max_seq_length, ent_sent_index, ename2embed_bert, tokenizer):
print("collecting positive samples!")
final_data = []
count = 10
for b in test_topic_rep_words:
if b not in ent_sent_index or b not in ename2embed_bert:
continue
for a in test_cand:
if a not in ent_sent_index or a not in ename2embed_bert:
continue
if a == b:
continue
cooccur = ent_sent_index[a].intersection(ent_sent_index[b])
if len(cooccur) > 0:
for sen in cooccur:
s = sentences[sen]
s = '[CLS] '+s
s = s.split(' ')
if a not in s or b not in s:
continue
p_index = s.index(a)
c_index = s.index(b)
s[p_index] = '[MASK]'
s[c_index] = '[MASK]'
s = ' '.join(s).replace('_', ' ').replace('-lrb-','(').replace('-rrb-',')').split(' ')
input_id = tokenizer.encode(' '.join(s))
mask_id = [x for x in range(len(input_id)) if input_id[x]==103]
if len(input_id) > max_seq_length:
if mask_id[1] - mask_id[0] >= max_seq_length:
continue
else:
input_id = input_id[mask_id[0]:mask_id[1]+1]
p_index = 0 if p_index<c_index else mask_id[1] - mask_id[0]
c_index = 0 if c_index<p_index else mask_id[1] - mask_id[0]
else:
p_index = mask_id[0] if p_index<c_index else mask_id[1]
c_index = mask_id[0] if c_index<p_index else mask_id[1]
input_id = torch.tensor(input_id)
r_sentence = F.pad(torch.tensor(input_id),(0,max_seq_length-len(input_id)), "constant", 0)
attention_mask = torch.cat((torch.ones_like(input_id), torch.zeros(max_seq_length-len(input_id), dtype=torch.int64)),dim=0)
p_mask = np.zeros((max_seq_length))
p_mask[p_index] = 1
c_mask = np.zeros((max_seq_length))
c_mask[c_index] = 1
final_data.append([r_sentence, p_mask, c_mask, attention_mask, a])
final_data.append([r_sentence, c_mask, p_mask, attention_mask, a])
return final_data
def generate_batch(batch):
input_ids = torch.tensor([np.array(entry[0]) for entry in batch])
entity1_mask = torch.tensor([entry[1] for entry in batch])
entity2_mask = torch.tensor([entry[2] for entry in batch])
attention_mask = torch.tensor([np.array(entry[3]) for entry in batch])
labels = torch.tensor([entry[4] for entry in batch])
return input_ids, entity1_mask, entity2_mask, attention_mask, labels
def generate_test_batch(batch):
input_ids = torch.tensor([np.array(entry[0]) for entry in batch])
entity1_mask = torch.tensor([entry[1] for entry in batch])
entity2_mask = torch.tensor([entry[2] for entry in batch])
attention_mask = torch.tensor([np.array(entry[3]) for entry in batch])
entity = [entry[4] for entry in batch]
return input_ids, entity1_mask, entity2_mask, attention_mask, entity