-
Notifications
You must be signed in to change notification settings - Fork 2
/
prep_wikiqa_data.py
436 lines (388 loc) · 16.3 KB
/
prep_wikiqa_data.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
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
#
# This script handles the preparation of the WikiQA dataset.
# It will be used by the tensorflow model.
################################################################################
import cPickle as pickle
import pdb
import numpy as np
from scipy import stats
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
# Each bucket is indicated by three numbers:
# (1) sentence_number, (2) question_length, (3) sentence_length
BUCKETS = [[5, 40, 40], [15, 40, 40], [30, 40, 40]]
# BUCKETS = [[5, 25, 50], [10, 25, 50], [10, 25, 50], [10, 25, 100],
# [20, 25, 50], [20, 25, 100], [30, 25, 50], [30, 25, 100]]
# Bucket collaps means ignore the first dimension of buckets:
# sentence_number.
# This will be used in the bag sampling technique defined
# "my_data_plus.py".
# It coresponds to on with the "if_pad_token_only" option in
# function "cast_data_to_buckets".
BUCKETS_COLLAPSED = [[40, 40]]
# Borrowed from "process_data.py" from the "WikiQACodePackage",
# from http://research.microsoft.com/en-US/downloads/a5c91569-d291-4450-aaab-5bce7995fe0c/default.aspx
class WordVecs(object):
"""
precompute embeddings for word/feature/tweet etc.
"""
def __init__(self, fname, vocab, binary=1, has_header=False):
if binary == 1:
word_vecs = self.load_bin_vec(fname, vocab)
else:
word_vecs = self.load_txt_vec(fname, vocab, has_header)
self.k = len(word_vecs.values()[0])
self.add_unknown_words(word_vecs, vocab, k=self.k)
self.W, self.word_idx_map = self.get_W(word_vecs, k=self.k)
def get_W(self, word_vecs, k=300):
"""
Get word matrix. W[i] is the vector for word indexed by i
"""
vocab_size = len(word_vecs)
word_idx_map = dict()
W = np.zeros(shape=(vocab_size+1, k))
W[0] = np.zeros(k)
i = 1
for word in word_vecs:
W[i] = word_vecs[word]
word_idx_map[word] = i
i += 1
return W, word_idx_map
def load_bin_vec(self, fname, vocab):
"""
Loads 300x1 word vecs from Google (Mikolov) word2vec
"""
word_vecs = {}
with open(fname, "rb") as f:
header = f.readline()
vocab_size, layer1_size = map(int, header.split())
binary_len = np.dtype('float32').itemsize * layer1_size
for line in xrange(vocab_size):
word = []
while True:
ch = f.read(1)
if ch == ' ':
word = ''.join(word)
break
if ch != '\n':
word.append(ch)
if word in vocab:
word_vecs[word] = np.fromstring(f.read(binary_len), dtype='float32')
else:
f.read(binary_len)
logger.info("num words already in word2vec: " + str(len(word_vecs)))
return word_vecs
def load_txt_vec(self, fname, vocab, has_header=False):
"""
Loads 50x1 word vecs from sentiment word embeddings (Tang et al., 2014)
"""
word_vecs = {}
pos = 0
with open(fname, "rb") as f:
if has_header: header = f.readline()
for line in f:
parts = line.strip().split()
word = parts[0]
if word in vocab:
word_vecs[word] = np.asarray(map(float, parts[1:]))
pos += 1
logger.info("num words already in word2vec: " + str(len(word_vecs)))
return word_vecs
def add_unknown_words(self, word_vecs, vocab, min_df=1, k=300):
"""
For words that occur in at least min_df documents, create a separate word vector.
0.25 is chosen so the unknown vectors have (approximately) same variance as pre-trained ones
"""
for word in vocab:
if word not in word_vecs and vocab[word] >= min_df:
#print word
word_vecs[word] = np.random.uniform(-0.25,0.25,k)
# Borrowed from "qa_cnn.py" from the "WikiQACodePackage",from:
# http://research.microsoft.com/en-US/downloads/a5c91569-d291-4450-aaab-5bce7995fe0c/default.aspx
# Slightly changed since LSTM doesn't need pad on the margin but CNN does.
# Also not necessarily truncate or pad sentences.
def get_idx_from_sent(sent, word_idx_map, max_l=None):
"""
Transforms sentence into a list of indices. Pad with zeroes.
"""
x = []
words = sent.split()
for i, word in enumerate(words):
if i >= max_l: break
if word in word_idx_map:
x.append(word_idx_map[word])
while len(x) < max_l:
x.append(0)
return x, len(words)
def visualize_distribution(data):
""" Visualize bucket distribution, a distribution over:
(sentence_number, question_length, sentence_length)"""
bucket_distribution = []
for qid, value in data.iteritems():
q_length = len(value['question'])
s_number = len(value['sentences'])
s_length = max([len(sent) for sent in value['sentences'].itervalues()])
bucket_distribution.append([s_number, q_length, s_length])
# pdb.set_trace()
distribution = np.transpose(np.array(bucket_distribution))
kde = stats.gaussian_kde(distribution)
density = kde(distribution)
idx = density.argsort()
x = distribution[0, idx]
y = distribution[1, idx]
z = distribution[2, idx]
density = density[idx]
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(x, y, z, c=density)
ax.set_xlabel('sentence number')
ax.set_ylabel('question length')
ax.set_zlabel('sentence length')
plt.show()
def make_data(revs, word_idx_map, max_l=None, val_test_splits=[2,3]):
"""
Transforms sentences.
train/val/test:
--- dict using "qid" as key
--- value: "question": list of intergers representing question
--- value: "sentences":
--- --- dict using "aid" as key
--- --- value: list of intergers representing sentence
--- value: "sentence_labels":
--- --- dict using "aid" as key
--- --- value: binary number representing if this sentence is an answer
"""
# Formulate data for answer triggering task
train, val, test = {}, {}, {}
val_split, test_split = val_test_splits
actual_question_max_l, actual_sentence_max_l = 0, 0
for rev in revs:
if rev["split"]==1: data_split = train
elif rev["split"]==val_split: data_split = val
elif rev["split"]==test_split: data_split = test
else: raise ValueError("Unknown data split: %s" % rev["split"])
qid, aid = rev['qid'], rev['aid']
# Add question sentence
if qid not in data_split:
data_split[qid] = {}
question, length = get_idx_from_sent(rev['question'], word_idx_map,
max_l)
actual_question_max_l = max(length, actual_question_max_l)
data_split[qid]['question'] = question
data_split[qid]['question_feats'] = [rev['num_words_q']]
data_split[qid]['sentences'] = {}
data_split[qid]['sentence_feats'] = {}
data_split[qid]['sentence_labels'] = {}
# else: # Just to make sure
# question, length = get_idx_from_sent(rev['question'], word_idx_map, max_l)
# # print question
# # print data_split[qid]['question']
# assert(question == data_split[qid]['question'])
# Add answer sentence
assert(aid not in data_split[qid])
sent, length = get_idx_from_sent(rev['answer'], word_idx_map, max_l)
actual_sentence_max_l = max(length, actual_sentence_max_l)
data_split[qid]['sentences'][aid] = sent
data_split[qid]['sentence_feats'][aid] = [rev['num_words_a']] + rev['features']
data_split[qid]['sentence_labels'][aid] = rev['y']
print('Question length: max %d' % actual_question_max_l)
print('Sentence length: max %d' % actual_sentence_max_l)
# Check data
# 1. make sure qid is consecutive
for data_split in [train, val, test]:
qids = data_split.keys()
qids.sort()
assert(qids == range(min(qids), max(qids)+1))
# 2. make sure aid for each question is consecutive
for data_split in [train, val, test]:
max_num_sentence = 0
min_num_sentence = 9999
for qid, value in data_split.iteritems():
num_sentences = len(value['sentence_labels'])
max_num_sentence = max(num_sentences, max_num_sentence)
min_num_sentence = min(num_sentences, min_num_sentence)
aids_1 = value['sentences'].keys()
aids_1.sort()
aids_2 = value['sentence_labels'].keys()
aids_2.sort()
assert(aids_1 == aids_2)
assert(aids_1 == range(min(aids_1), max(aids_2)+1))
print('Sentence number: max %d; min %d' %
(max_num_sentence, min_num_sentence))
return [train, val, test]
def pad_token_only(m_buckets_collapsed, q, s, l):
""" Pad each sample into a bucket, but only pad short sentences.
This means do not add all paded sentences.
This also means that the first dimension of buckets is useless.
NOTE: use index 0 to pad.
"""
assert(s)
assert(len(s) == len(l))
fit = False
for bid, bucket in enumerate(m_buckets_collapsed):
if len(q) <= bucket[0] and \
len(s[0]) <= bucket[1]:
for i in range(bucket[0] - len(q)):
q.append(0)
assert(len(q) == bucket[0])
for i in range(len(s)):
for j in range(bucket[1] - len(s[i])):
s[i].append(0)
assert(len(s[i]) == bucket[1])
fit = True
break
if not fit:
print("(%d, %d, %d) doesn't fit into any bucket!" %
(len(s), len(q), len(s[0])))
return None
else:
return bid, np.array(q), np.array(s), np.array(l)
def pad(m_buckets, q, s, l, qf, sf):
""" Pad each sample q + sents into a bucket.
NOTE: use index 0 to pad.
"""
assert(s)
assert(len(s) == len(l))
fit = False
for bid, bucket in enumerate(m_buckets):
if len(s) <= bucket[0] and \
len(q) <= bucket[1] and \
len(s[0]) <= bucket[2]:
# Add pad sentence placeholders
for i in range(bucket[0] - len(s)):
s.append([]) # all-padding sentence place taker
l.append(0) # all_padding sentence's label is 0
sf.append([0,0,0]) # all-padding sentence's features are zeros
assert(len(s) == bucket[0])
assert(len(l) == bucket[0])
assert(len(sf) == bucket[0])
# Add pad question tokens
for i in range(bucket[1] - len(q)):
q.append(0)
assert(len(q) == bucket[1])
# Add pad sentence tokens
for i in range(bucket[0]):
for j in range(bucket[2] - len(s[i])):
s[i].append(0)
assert(len(s[i]) == bucket[2])
fit = True
break
if not fit:
print("(%d, %d, %d) doesn't fit into any bucket!" %
(len(s), len(q), len(s[0])))
return None
else:
return bid, np.array(q), np.array(s), np.array(l), np.array(qf), np.array(sf)
def cast_data_to_buckets(data, m_buckets, if_pad_token_only=False):
""" Fit data into buckets, for feeding into NN model
data_tuple_b:
- 1st dimension idx: bucket index
- 2nd dimension idx:
0 : all questions
1 : all sentences
2 : all labels (corresponding to sentences)
3 : all question features
4 : all sentence features
"""
data_tuple_b = [[[], # question list
[], # sentences list
[], # sentence label list
[], # question feature list
[]] # sentence feature list
for b_id in xrange(len(m_buckets))]
for qid, value in data.iteritems():
q = value['question']
qf = value['question_feats']
s, sf, l = [], [], []
assert(value['sentences'].keys() == value['sentence_labels'].keys())
assert(value['sentences'].keys() == value['sentence_feats'].keys())
for aid in value['sentences'].iterkeys():
s.append(value['sentences'][aid])
sf.append(value['sentence_feats'][aid])
l.append(value['sentence_labels'][aid])
# Padding
if not if_pad_token_only:
b_id, question, sentences, label, q_feats, s_feats = \
pad(m_buckets, q, s, l, qf, sf)
else:
#FIXME: hasn't updated to support question features "qf" and
# sentence features "sf"
b_id, question, sentences, label = \
pad_token_only(m_buckets, q, s, l)
data_tuple_b[b_id][0].append(question)
data_tuple_b[b_id][1].append(sentences)
data_tuple_b[b_id][2].append(label)
data_tuple_b[b_id][3].append(q_feats)
data_tuple_b[b_id][4].append(s_feats)
# NOTE: this has been changed since otherwise the return will generate
# errors, when using collapsed buckets.
# for b_id in xrange(len(data_tuple_b)):
# data_tuple_b[b_id][0] = np.array(data_tuple_b[b_id][0])
# data_tuple_b[b_id][1] = np.array(data_tuple_b[b_id][1])
# data_tuple_b[b_id][2] = np.array(data_tuple_b[b_id][2])
return np.array(data_tuple_b)
def find_similar_words(wordvecs):
""" Use loaded word embeddings to find out the most similar words in the
embedded vector space.
"""
from sklearn.metrics import pairwise_distances
from scipy.spatial.distance import cosine
pairwise_sim_mat = 1 - pairwise_distances(wordvecs.W[1:],
metric='cosine',
# metric='euclidean',
)
id2word = {}
for key, value in wordvecs.word_idx_map.iteritems():
assert(value not in id2word)
id2word[value] = key
while True:
word = raw_input("Enter a word ('STOP' to quit): ")
if word == 'STOP': break
try:
w_id = wordvecs.word_idx_map[word]
except KeyError:
print '%s not in the vocabulary.' % word
sim_w_id = pairwise_sim_mat[w_id-1].argsort()[-10:][::-1]
for i in sim_w_id:
print id2word[i+1],
print ''
def prep_data():
"""
The HALF preprocessed data "wiki_cnn.pkl" is generated by
python script "process_data.py" in the "WikiQACodePackage",
specified by path "WikiQACodePackage/code/process_data.py".
It performs basic preprocessing (like tokenize) and also loads
the pretrained Word2Vec results from Google.
The command to run the script is:
$ python -u process_data.py --w2v_fname ../data/GoogleNews-vectors-negative300.bin \
--extract_feat 1 ../data/wiki/WikiQASent-train.txt ../data/wiki/WikiQASent-dev.txt \
../data/wiki/WikiQASent-test.txt ../wiki_cnn.pkl
"""
revs, wordvecs, max_l = pickle.load(open('./data/wiki_cnn.pkl', 'rb'))
# find_similar_words(wordvecs)
# pdb.set_trace()
# NOTE: keep this for being consistent with WikiQA code package
# Truncate and pad all questions or sentences to this length.
max_l = 40
dataset = make_data(revs, wordvecs.word_idx_map, max_l=max_l)
train_data, val_data, test_data = dataset
# visualize_distribution(train_data) # aid choose buckets
# SAMPLEBAG
train_tuple_b = cast_data_to_buckets(train_data, BUCKETS)
# train_tuple_b = cast_data_to_buckets(train_data, BUCKETS_COLLAPSED,
# if_pad_token_only=True)
valid_tuple_b = cast_data_to_buckets(val_data, BUCKETS)
test_tuple_b = cast_data_to_buckets(test_data, BUCKETS)
# Vocabulary
assert('<pad>' not in wordvecs.word_idx_map)
assert(0 not in wordvecs.word_idx_map.values())
word2id = wordvecs.word_idx_map
word2id['<pad>'] = 0
# Embeddings
word_embeddings = wordvecs.W
assert(len(word_embeddings) == len(word2id))
return train_tuple_b, valid_tuple_b, test_tuple_b, word2id, word_embeddings
if __name__ == '__main__':
train_tuple_b, valid_tuple_b, test_tuple_b, word2id, word_embeddings = \
prep_data()
pdb.set_trace()