-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataset.py
215 lines (167 loc) · 8.16 KB
/
dataset.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
import os
import re
import json
import tqdm
import spacy
from pathlib import Path
from sklearn.model_selection import train_test_split
from torchtext.data import Field, Example, Dataset
# import constants
from constants import *
class AnswerVerbalizationDataset(object):
TOKENIZE_SEQ = lambda self, x: x.replace("?", " ?").\
replace(".", " .").\
replace(",", " ,").\
replace("'", " '").\
split()
def __init__(self):
self.train_path = str(ROOT_PATH) + '/data/' + args.dataset + '/preprocessed_train.json'
self.test_path = str(ROOT_PATH) + '/data/' + args.dataset + '/preprocessed_test.json'
self.prepare_data = {
VQUANDA: self._prepare_vquanda,
PARAQA: self._prepare_paraqa,
VANILLA: self._prepare_vanilla
}
self.load_data_and_fields()
def _prepare_vquanda(self, data):
input_data = []
for d in data:
uid = d['uid']
preprrocessed = d[PREPROCESSED]
# preprocess
question = preprrocessed[QUESTION].lower()
logical_form = preprrocessed[LOGICAL_FORM].lower()
answer_verbalization = preprrocessed[ANSWER].lower()
negative_lf = preprrocessed[NEGATIVE]
# tokenize
input = self.TOKENIZE_SEQ(question)
logical_form = self.TOKENIZE_SEQ(logical_form)
answer_verbalization = self.TOKENIZE_SEQ(answer_verbalization)
while len(input) < args.max_input_size:
input.append(PAD_TOKEN)
while len(logical_form) < args.max_input_size:
logical_form.append(PAD_TOKEN)
input_data.append([input, logical_form, logical_form, ['1'], answer_verbalization])
for negative in negative_lf[:1]:
negative = self.TOKENIZE_SEQ(negative.lower())
dec_logical_form = [PAD_TOKEN for _ in range(args.max_input_size)]
input_data.append([input, negative, dec_logical_form, ['0'], answer_verbalization])
return input_data
def _prepare_paraqa(self, data):
input_data = []
for d in data:
uid = d['uid']
preprrocessed = d[PREPROCESSED]
question = preprrocessed[QUESTION].lower()
logical_form = preprrocessed[LOGICAL_FORM].lower()
answers = preprrocessed[ANSWER]
negative_lf = preprrocessed[NEGATIVE]
input = self.TOKENIZE_SEQ(question)
logical_form = self.TOKENIZE_SEQ(logical_form)
while len(input) < args.max_input_size:
input.append(PAD_TOKEN)
while len(logical_form) < args.max_input_size:
logical_form.append(PAD_TOKEN)
for answer in answers:
answer_verbalization = self.TOKENIZE_SEQ(answer.lower())
input_data.append([input, logical_form, logical_form, ['1'], answer_verbalization])
for negative in negative_lf:
negative = self.TOKENIZE_SEQ(negative.lower())
dec_logical_form = [UNK_TOKEN for _ in range(args.max_input_size)]
input_data.append([input, negative, dec_logical_form, ['0'], answer_verbalization])
return input_data
def _prepare_vanilla(self, data):
input_data = []
for d in data:
uid = d['question_id']
preprrocessed = d[PREPROCESSED]
question = preprrocessed[QUESTION].lower()
logical_form = preprrocessed[LOGICAL_FORM].lower()
answer_verbalization = preprrocessed[ANSWER].lower()
negative_lf = preprrocessed[NEGATIVE]
input = self.TOKENIZE_SEQ(question)
logical_form = self.TOKENIZE_SEQ(logical_form)
answer_verbalization = self.TOKENIZE_SEQ(answer_verbalization)
negative_lf = self.TOKENIZE_SEQ(negative_lf)
if len(input) > args.max_input_size or\
len(logical_form) > args.max_input_size or\
len(negative_lf) > args.max_input_size:
continue
while len(input) < args.max_input_size:
input.append(PAD_TOKEN)
while len(logical_form) < args.max_input_size:
logical_form.append(PAD_TOKEN)
input_data.append([input, logical_form, logical_form, ['1'], answer_verbalization])
while len(negative_lf) < args.max_input_size:
negative_lf.append(PAD_TOKEN)
dec_logical_form = [PAD_TOKEN for _ in range(args.max_input_size)]
input_data.append([input, negative_lf, dec_logical_form, ['0'], answer_verbalization])
return input_data
def _make_torchtext_dataset(self, data, fields):
examples = [Example.fromlist(i, fields) for i in data]
return Dataset(examples, fields)
def load_data_and_fields(self, cover_entities=False, query_as_input=False):
train, test, val = [], [], []
# read data
with open(self.train_path) as json_file:
train = json.load(json_file)
with open(self.test_path) as json_file:
test = json.load(json_file)
test, val = train_test_split(test, test_size=0.4, shuffle=False)
train = self.prepare_data[args.dataset](train)
val = self.prepare_data[args.dataset](val)
test = self.prepare_data[args.dataset](test)
# create fields
self.input_field = Field(init_token=START_TOKEN,
eos_token=CTX_TOKEN,
pad_token=PAD_TOKEN,
unk_token=UNK_TOKEN,
lower=True,
batch_first=True)
self.lf_field = Field(init_token=START_TOKEN,
eos_token=CTX_TOKEN,
pad_token=PAD_TOKEN,
unk_token=UNK_TOKEN,
lower=True,
batch_first=True)
self.sim_field = Field(init_token='0',
eos_token='0',
pad_token=PAD_TOKEN,
unk_token='0',
batch_first=True)
self.decoder_field = Field(init_token=START_TOKEN,
eos_token=END_TOKEN,
pad_token=PAD_TOKEN,
unk_token=UNK_TOKEN,
lower=True,
batch_first=True)
fields_tuple = [(INPUT, self.input_field),
(ST_LOGICAL_FORM, self.lf_field),
(DEC_LOGICAL_FORM, self.lf_field),
(SIMILARITY_THRESHOLD, self.sim_field),
(DECODER, self.decoder_field)]
# create toechtext datasets
self.train_data = self._make_torchtext_dataset(train, fields_tuple)
self.val_data = self._make_torchtext_dataset(val, fields_tuple)
self.test_data = self._make_torchtext_dataset(test, fields_tuple)
# build vocabularies
self.input_field.build_vocab(self.train_data, self.val_data, self.test_data, min_freq=0, vectors='glove.840B.300d')
self.lf_field.build_vocab(self.train_data, self.val_data, self.test_data, min_freq=0, vectors='glove.840B.300d')
self.sim_field.build_vocab(self.train_data, self.val_data, self.test_data, min_freq=0)
self.decoder_field.build_vocab(self.train_data, self.val_data, self.test_data, min_freq=0)
def get_data(self):
return self.train_data, self.val_data, self.test_data
def get_fields(self):
return {
INPUT: self.input_field,
LOGICAL_FORM: self.lf_field,
SIMILARITY_THRESHOLD: self.sim_field,
DECODER: self.decoder_field,
}
def get_vocabs(self):
return {
INPUT: self.input_field.vocab,
LOGICAL_FORM: self.lf_field.vocab,
SIMILARITY_THRESHOLD: self.sim_field.vocab,
DECODER: self.decoder_field.vocab
}