-
Notifications
You must be signed in to change notification settings - Fork 168
/
annotate_ws.py
196 lines (166 loc) · 6.75 KB
/
annotate_ws.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
#!/usr/bin/env python3
# docker run --name corenlp -d -p 9000:9000 vzhong/corenlp-server
# Wonseok Hwang. Jan 6 2019, Comment added
from argparse import ArgumentDefaultsHelpFormatter, ArgumentParser
import os
import records
import ujson as json
from stanza.nlp.corenlp import CoreNLPClient
from tqdm import tqdm
import copy
from wikisql.lib.common import count_lines, detokenize
from wikisql.lib.query import Query
client = None
def annotate(sentence, lower=True):
global client
if client is None:
client = CoreNLPClient(default_annotators='ssplit,tokenize'.split(','))
words, gloss, after = [], [], []
for s in client.annotate(sentence):
for t in s:
words.append(t.word)
gloss.append(t.originalText)
after.append(t.after)
if lower:
words = [w.lower() for w in words]
return {
'gloss': gloss,
'words': words,
'after': after,
}
def annotate_example(example, table):
ann = {'table_id': example['table_id']}
ann['question'] = annotate(example['question'])
ann['table'] = {
'header': [annotate(h) for h in table['header']],
}
ann['query'] = sql = copy.deepcopy(example['sql'])
for c in ann['query']['conds']:
c[-1] = annotate(str(c[-1]))
q1 = 'SYMSELECT SYMAGG {} SYMCOL {}'.format(Query.agg_ops[sql['agg']], table['header'][sql['sel']])
q2 = ['SYMCOL {} SYMOP {} SYMCOND {}'.format(table['header'][col], Query.cond_ops[op], detokenize(cond)) for col, op, cond in sql['conds']]
if q2:
q2 = 'SYMWHERE ' + ' SYMAND '.join(q2) + ' SYMEND'
else:
q2 = 'SYMEND'
inp = 'SYMSYMS {syms} SYMAGGOPS {aggops} SYMCONDOPS {condops} SYMTABLE {table} SYMQUESTION {question} SYMEND'.format(
syms=' '.join(['SYM' + s for s in Query.syms]),
table=' '.join(['SYMCOL ' + s for s in table['header']]),
question=example['question'],
aggops=' '.join([s for s in Query.agg_ops]),
condops=' '.join([s for s in Query.cond_ops]),
)
ann['seq_input'] = annotate(inp)
out = '{q1} {q2}'.format(q1=q1, q2=q2) if q2 else q1
ann['seq_output'] = annotate(out)
ann['where_output'] = annotate(q2)
assert 'symend' in ann['seq_output']['words']
assert 'symend' in ann['where_output']['words']
return ann
def find_sub_list(sl, l):
# from stack overflow.
results = []
sll = len(sl)
for ind in (i for i, e in enumerate(l) if e == sl[0]):
if l[ind:ind + sll] == sl:
results.append((ind, ind + sll - 1))
return results
def check_wv_tok_in_nlu_tok(wv_tok1, nlu_t1):
"""
Jan.2019: Wonseok
Generate SQuAD style start and end index of wv in nlu. Index is for of after WordPiece tokenization.
Assumption: where_str always presents in the nlu.
return:
st_idx of where-value string token in nlu under CoreNLP tokenization scheme.
"""
g_wvi1_corenlp = []
nlu_t1_low = [tok.lower() for tok in nlu_t1]
for i_wn, wv_tok11 in enumerate(wv_tok1):
wv_tok11_low = [tok.lower() for tok in wv_tok11]
results = find_sub_list(wv_tok11_low, nlu_t1_low)
st_idx, ed_idx = results[0]
g_wvi1_corenlp.append( [st_idx, ed_idx] )
return g_wvi1_corenlp
def annotate_example_ws(example, table):
"""
Jan. 2019: Wonseok
Annotate only the information that will be used in our model.
"""
ann = {'table_id': example['table_id'],'phase': example['phase']}
_nlu_ann = annotate(example['question'])
ann['question'] = example['question']
ann['question_tok'] = _nlu_ann['gloss']
# ann['table'] = {
# 'header': [annotate(h) for h in table['header']],
# }
ann['sql'] = example['sql']
ann['query'] = sql = copy.deepcopy(example['sql'])
conds1 = ann['sql']['conds']
wv_ann1 = []
for conds11 in conds1:
_wv_ann1 = annotate(str(conds11[2]))
wv_ann11 = _wv_ann1['gloss']
wv_ann1.append( wv_ann11 )
# Check whether wv_ann exsits inside question_tok
try:
wvi1_corenlp = check_wv_tok_in_nlu_tok(wv_ann1, ann['question_tok'])
ann['wvi_corenlp'] = wvi1_corenlp
except:
ann['wvi_corenlp'] = None
ann['tok_error'] = 'SQuAD style st, ed are not found under CoreNLP.'
return ann
def is_valid_example(e):
if not all([h['words'] for h in e['table']['header']]):
return False
headers = [detokenize(h).lower() for h in e['table']['header']]
if len(headers) != len(set(headers)):
return False
input_vocab = set(e['seq_input']['words'])
for w in e['seq_output']['words']:
if w not in input_vocab:
print('query word "{}" is not in input vocabulary.\n{}'.format(w, e['seq_input']['words']))
return False
input_vocab = set(e['question']['words'])
for col, op, cond in e['query']['conds']:
for w in cond['words']:
if w not in input_vocab:
print('cond word "{}" is not in input vocabulary.\n{}'.format(w, e['question']['words']))
return False
return True
if __name__ == '__main__':
parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter)
parser.add_argument('--din', default='/Users/wonseok/data/WikiSQL-1.1/data', help='data directory')
parser.add_argument('--dout', default='/Users/wonseok/data/wikisql_tok', help='output directory')
parser.add_argument('--split', default='train,dev,test', help='comma=separated list of splits to process')
args = parser.parse_args()
answer_toy = not True
toy_size = 10
if not os.path.isdir(args.dout):
os.makedirs(args.dout)
# for split in ['train', 'dev', 'test']:
for split in args.split.split(','):
fsplit = os.path.join(args.din, split) + '.jsonl'
ftable = os.path.join(args.din, split) + '.tables.jsonl'
fout = os.path.join(args.dout, split) + '_tok.jsonl'
print('annotating {}'.format(fsplit))
with open(fsplit) as fs, open(ftable) as ft, open(fout, 'wt') as fo:
print('loading tables')
# ws: Construct table dict with table_id as a key.
tables = {}
for line in tqdm(ft, total=count_lines(ftable)):
d = json.loads(line)
tables[d['id']] = d
print('loading examples')
n_written = 0
cnt = -1
for line in tqdm(fs, total=count_lines(fsplit)):
cnt += 1
d = json.loads(line)
# a = annotate_example(d, tables[d['table_id']])
a = annotate_example_ws(d, tables[d['table_id']])
fo.write(json.dumps(a) + '\n')
n_written += 1
if answer_toy:
if cnt > toy_size:
break
print('wrote {} examples'.format(n_written))