-
Notifications
You must be signed in to change notification settings - Fork 0
/
corpus33.py
394 lines (310 loc) · 14.5 KB
/
corpus33.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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
# -*- coding:utf-8 -*-
# Author: Roger
# Created by Roger on 2017/10/24
from __future__ import absolute_import
import codecs
import math
try:
import simplejson as json
except:
import json
import torch
from torch.autograd import Variable
from layers import Dictionary, Constants
def convert2longtensor(x):
return torch.LongTensor(x)
def convert2variable(x, device=-1, volatile=True):
if device >= 0:
x = x.cuda(device)
return Variable(x, volatile=volatile)
class Evidence(object):
def __init__(self, e_key, e_text, e_text_index, e_feature, starts, ends):
self.e_key = e_key # String
self.e_text = e_text # list(string)
self.e_text_index = e_text_index # torch.LongTensor
self.e_feature = e_feature # torch.LongTensor
self.starts = starts # list(int)
self.ends = ends # list(int)
def __iter__(self):
for d in [self.e_key, self.e_text, self.e_text_index, self.e_feature, self.starts, self.ends]:
yield d
@staticmethod
def load_one_evidence(evidence, word_dict, pos_dict, ner_dict):
e_key = evidence['e_key']
e_text = evidence["evidence_tokens"]
if 'answer_starts' in evidence:
if len(evidence['answer_starts']) == 0:
starts = [-1]
else:
starts = evidence['answer_starts']
else:
starts = [-1]
if 'answer_ends' in evidence:
if len(evidence['answer_ends']) == 0:
ends = [-1]
else:
ends = evidence['answer_ends']
else:
ends = [-1]
# if starts[0] == -1 or ends[0] == -1:
# return None
e_text_index = convert2longtensor(word_dict.convert_to_index(e_text, Constants.UNK_WORD))
e_pos = evidence['evidence_pos']
e_ner = evidence['evidence_ners']
qe_feature = convert2longtensor(evidence["qecomm"])
e_ner_index = convert2longtensor(ner_dict.convert_to_index(e_ner, Constants.UNK_WORD))
e_pos_index = convert2longtensor(pos_dict.convert_to_index(e_pos, Constants.UNK_WORD))
e_feature = torch.stack([e_pos_index, e_ner_index, qe_feature], dim=1)
return Evidence(e_key, e_text, e_text_index, e_feature, starts, ends)
@staticmethod
def batchify(data):
e_key, e_real_text, e_text_index, e_feature_index, starts, ends = zip(*data)
e_feature_size = e_feature_index[0].size()[1]
e_lens = [e_text_index[i].size(0) for i in range(len(data))]
max_e_length = max(e_lens)
e_text = e_text_index[0].new(len(data), max_e_length).fill_(Constants.PAD)
e_feature = e_feature_index[0].new(len(data), max_e_length, e_feature_size).fill_(Constants.PAD)
for i in range(len(data)):
length = e_text_index[i].size(0)
e_text[i, :].narrow(0, 0, length).copy_(e_text_index[i])
e_feature[i, :, :].narrow(0, 0, length).copy_(e_feature_index[i])
'''
start_position = []
end_position = []
for s,e in zip(starts, ends):
start_position.append(convert2longtensor(s))
end_position.append(convert2longtensor(e))
'''
start_position = starts
end_position = ends
e_lens = convert2longtensor(e_lens)
return e_text, e_feature, e_lens, start_position, end_position, e_key, e_real_text
class Question(object):
def __init__(self, q_key, q_text, q_text_index, q_feature):
self.q_key = q_key # String
self.q_text = q_text # list(string)
self.q_text_index = q_text_index # torch.LongTensor
self.q_feature = q_feature # torch.LongTensor
def __iter__(self):
for d in [self.q_key, self.q_text, self.q_text_index, self.q_feature]:
yield d
@staticmethod
def batchify(data):
q_key, q_real_text, q_text_index, q_featurq_index = zip(*data)
q_featurq_size = q_featurq_index[0].size()[1]
q_lens = [q_text_index[i].size(0) for i in range(len(data))]
max_q_length = max(q_lens)
q_text = q_text_index[0].new(len(data), max_q_length).fill_(Constants.PAD)
q_feature = q_featurq_index[0].new(len(data), max_q_length, q_featurq_size).fill_(Constants.PAD)
for i in range(len(data)):
length = q_text_index[i].size(0)
q_text[i, :].narrow(0, 0, length).copy_(q_text_index[i])
q_feature[i, :, :].narrow(0, 0, length).copy_(q_featurq_index[i])
q_lens = convert2longtensor(q_lens)
return q_text, q_feature, q_lens, q_key, q_real_text
@staticmethod
def load_one_question(data, word_dict, pos_dict, ner_dict):
q_key = data['q_key']
q_text = data["question_tokens"]
q_text_index = convert2longtensor(word_dict.convert_to_index(q_text, Constants.UNK_WORD))
q_ner = data["question_ners"]
q_pos = data["question_pos"]
q_ner_index = convert2longtensor(ner_dict.convert_to_index(q_ner, Constants.UNK_WORD))
q_pos_index = convert2longtensor(pos_dict.convert_to_index(q_pos, Constants.UNK_WORD))
q_feature = torch.stack([q_pos_index, q_ner_index], dim=1)
return Question(q_key, q_text, q_text_index, q_feature)
class WebQACorpus(object):
def __init__(self, filename, batch_size=64, device=-1, volatile=False,
word_dict=None, ner_dict=None, pos_dict=None):
if word_dict is None:
self.word_d, self.pos_dict, self.ner_dict = self.load_word_dictionary(filename)
else:
self.word_d = word_dict
self.ner_dict = ner_dict
self.pos_dict = pos_dict
question_dict, evidence_dict, train_pair = self.load_data_file(filename,
word_dict=self.word_d,
ner_dict=self.ner_dict,
pos_dict=self.pos_dict)
self.question_dict = question_dict # {q_key: [question, [eid]]}
self.evidence_dict = evidence_dict # {eid: evidence}
self.data = train_pair # (q_key, eid)
self.batch_size = batch_size
self.device = device
self.volatile = volatile
def __sizeof__(self):
return len(self.data)
def __len__(self):
return len(self.data)
def cpu(self):
self.device = -1
def cuda(self, device=0):
self.device = device
def set_device(self, device=-1):
self.device = device
def set_batch_size(self, batch_size=50):
self.batch_size = batch_size
def _question_evidence(self, question_ids, evidence_ids):
questions = [self.question_dict[qid][0] for qid in question_ids]
evidences = [self.evidence_dict[eid] for eid in evidence_ids]
q_text, q_feature, q_lens, q_key, q_real_text = Question.batchify(questions)
e_text, e_feature, e_lens, start_position, end_position, e_key, e_real_text = Evidence.batchify(evidences)
q_text, q_feature, q_lens = [convert2variable(x, self.device, self.volatile)
for x in [q_text, q_feature, q_lens]]
e_text, e_feature, e_lens, = [convert2variable(x, self.device, self.volatile)
for x in [e_text, e_feature, e_lens]]
return q_text, e_text, start_position, end_position, q_lens, e_lens, q_feature, \
e_feature, q_key, e_key, q_real_text, e_real_text
def _batchify(self, data):
question_ids, evidence_ids = zip(*data)
return self._question_evidence(question_ids, evidence_ids)
def next_batch(self, shuffle=True):
num_batch = int(math.ceil(len(self.data) / float(self.batch_size)))
if not shuffle:
data = self.data
random_indexs = torch.range(0, num_batch - 1)
else:
data = [self.data[index] for index in torch.randperm(len(self.data))]
random_indexs = torch.randperm(num_batch)
for index, i in enumerate(random_indexs):
start, end = i * self.batch_size, (i + 1) * self.batch_size
_batch_size = len(data[start:end])
batch_data = self._batchify(data[start:end])
q_text, e_text, start_position, end_position = batch_data[:4]
q_lens, e_lens, q_feature, e_feature = batch_data[4:8]
q_keys, e_keys = batch_data[8:10]
yield Batch(q_text, e_text, start_position, end_position,
q_lens, e_lens, q_feature, e_feature,
_batch_size, q_keys, e_keys)
def next_question(self):
for qid in self.question_dict.keys():
_, evidence_ids = self.question_dict[qid]
_batch_size = len(evidence_ids)
if _batch_size == 0:
continue
batch_data = self._question_evidence([qid] * _batch_size, evidence_ids)
q_text, e_text, start_position, end_position = batch_data[:4]
q_lens, e_lens, q_feature, e_feature = batch_data[4:8]
q_keys, e_keys, q_real_text, e_real_text = batch_data[8:]
yield BatchQuestion(q_text, e_text, start_position, end_position,
q_lens, e_lens, q_feature, e_feature,
_batch_size, q_keys, e_keys, e_real_text, q_real_text[0])
@staticmethod
def load_one_line_json(line, word_dict, pos_dict, ner_dict):
data = json.loads(line)
question = Question.load_one_question(data, word_dict, pos_dict, ner_dict)
evidences = list()
for evidence in data["evidences"]:
evidence_data = Evidence.load_one_evidence(evidence, word_dict, pos_dict, ner_dict)
if evidence_data is None:
continue
evidences.append(evidence_data)
return question, evidences
@staticmethod
def load_data_file(filename, word_dict, pos_dict, ner_dict):
question_dict = dict()
evidence_dict = dict()
train_pair = list()
count = 0
with codecs.open(filename, 'r', 'utf8') as fin:
for line in fin:
count += 1
all_evidence = list()
question, evidences = WebQACorpus.load_one_line_json(line, word_dict, pos_dict, ner_dict)
for e in evidences:
eid = "%s||%s" % (question.q_key, e.e_key)
evidence_dict[eid] = e
all_evidence.append(eid)
if e.starts[0] == -1 or e.ends[0] == -1:
# Skip No Answer Evidence when Train
continue
if len(e.starts) != len(e.ends): # 把长度不一样的去掉
continue
train_pair.append((question.q_key, eid))
question_dict[question.q_key] = [question, all_evidence]
if count % 5000 == 0:
print(count)
print('load data from %s, get %s qe pairs. ' %(filename, len(train_pair)))
return question_dict, evidence_dict, train_pair
@staticmethod
def load_word_dictionary(filename, word_dict=None, pos_dict=None, ner_dict=None):
if word_dict is None:
word_dict = Dictionary()
word_dict.add_specials([Constants.PAD_WORD, Constants.UNK_WORD, Constants.BOS_WORD, Constants.EOS_WORD],
[Constants.PAD, Constants.UNK, Constants.BOS, Constants.EOS])
if pos_dict is None:
pos_dict = Dictionary()
pos_dict.add_specials([Constants.PAD_WORD, Constants.UNK_WORD],
[Constants.PAD, Constants.UNK])
if ner_dict is None:
ner_dict = Dictionary()
ner_dict.add_specials([Constants.PAD_WORD, Constants.UNK_WORD],
[Constants.PAD, Constants.UNK])
with codecs.open(filename, 'r') as fin:
for line in fin:
data = json.loads(line)
for token in data["question_tokens"]:
word_dict.add(token)
for pos in data['question_pos']:
pos_dict.add(pos)
for ner in data['question_ners']:
ner_dict.add(ner)
for evidence in data["evidences"]:
for token in evidence["evidence_tokens"]:
word_dict.add(token)
for pos in evidence['evidence_pos']:
pos_dict.add(pos)
for ner in evidence['evidence_ners']:
ner_dict.add(ner)
return word_dict, pos_dict, ner_dict
@staticmethod
def load_pos_dictionary():
return Dictionary()
@staticmethod
def load_ner_dictionary():
return Dictionary()
class Batch(object):
def __init__(self, q_text, e_text, start, end,
q_lens, e_lens, q_feature, e_feature,
batch_size, q_keys, e_keys):
self.q_text = q_text
self.e_text = e_text
self.start_position = start
self.end_position = end
self.q_lens = q_lens
self.e_lens = e_lens
self.q_feature = q_feature
self.e_feature = e_feature
self.batch_size = batch_size
self.pred = None
self.q_keys = q_keys
self.e_keys = e_keys
class BatchQuestion(object):
def __init__(self, q_text, e_text, start, end,
q_lens, e_lens, q_feature, e_feature,
batch_size, q_keys, e_keys,
evidence_raw_text=None, question_raw_text=None):
self.q_text = q_text
self.e_text = e_text
self.start_position = start
self.end_position = end
self.q_lens = q_lens
self.e_lens = e_lens
self.q_feature = q_feature
self.e_feature = e_feature
self.batch_size = batch_size
self.pred = None
self.q_keys = q_keys
self.e_keys = e_keys
self.evidence_raw_text = evidence_raw_text
self.question_raw_text = question_raw_text
def test():
corpus = WebQACorpus("data/baidu_data.json")
for data in corpus.next_question():
for index, (start, end, leng) in enumerate(torch.cat([data.start_position.unsqueeze(-1),
data.end_position.unsqueeze(-1),
data.e_lens.unsqueeze(-1)],
1)):
print(''.join(data.evidence_raw_text[index][start.data[0]:end.data[0] + 1]))
if __name__ == "__main__":
test()