-
Notifications
You must be signed in to change notification settings - Fork 0
/
ngram_encoder.py
207 lines (183 loc) · 6.92 KB
/
ngram_encoder.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
# coding=utf-8
# Copyright 2022 The NALM Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""ptb2016 processing code for ngram based permutation dataset.
This script collect ngrams from the train_path. The data from train_path,
valid_path and test_path will be subsequently shuffled and displaced based
on the collected ngrams.
See "Assessing Non-autoregressive Alignment in Neural Machine Translation via
Word Reordering" for the full description of the data processing.
"""
import random, copy, string
import pickle, os
prune_threshold = 2 # the higher the threshold, the less word order retained
ngram_set_path = './ptbngrambank/ptb2016ngramsets'+'_'+str(prune_threshold)
ratios = [0.4,0.6,0.8]
flip_prob = [0.5]
train_path='./tmp/t2t_datagen/ptb2016/ptb2016.train.txt'
valid_path='./tmp/t2t_datagen/ptb2016/ptb2016.valid.txt'
test_path ='./tmp/t2t_datagen/ptb2016/ptb2016.test.txt'
destination_path='./tmp/t2t_datagen/ptb2016/'
def load_data(path):
with open(path) as f:
text = f.read()
sents = text.split('\n')
sents = [sent.split(' ')[:-1] for sent in sents]
return sents
if os.path.exists(ngram_set_path):
with open(ngram_set_path,'rb') as f:
ngramsets = pickle.load(f)
else:
import nltk
sents = load_data(train_path)
#below code adapted from https://www.geeksforgeeks.org/n-gram-language-modelling-with-nltk/
stop_words = set(nltk.corpus.stopwords.words('english'))
string.punctuation = string.punctuation +'"'+'"'+'-'+'''+'''+'—'
removal_list = list(stop_words) + list(string.punctuation)+ ['lt','rt']
bigram,trigram,quadgram=[],[],[]
for sentence in sents:
sentence = list(map(lambda x:x.lower(),sentence))
for word in sentence:
if word== '.':
sentence.remove(word)
bigram.extend(list(nltk.ngrams(sentence, 2)))
trigram.extend(list(nltk.ngrams(sentence, 3)))
quadgram.extend(list(nltk.ngrams(sentence, 4)))
def remove_stopwords(x):
y = []
for pair in x:
count = 0
for word in pair:
if word in removal_list:
count = count or 0
else:
count = count or 1
if (count==1):
y.append(pair)
return (y)
bigram=nltk.FreqDist(remove_stopwords(bigram))
trigram=nltk.FreqDist(remove_stopwords(trigram))
quadgram=nltk.FreqDist(remove_stopwords(quadgram))
#end adaptation
ngramsets={}
ngramsets['bi']=set([x for x in bigram if bigram[x]>prune_threshold])
ngramsets['tri']=set([x for x in trigram if trigram[x]>prune_threshold])
ngramsets['quad']=set([x for x in quadgram if quadgram[x]>prune_threshold])
with open(ngram_set_path,'wb') as f:
pickle.dump(ngramsets,f)
def find_gram(sent):
for i in range(len(sent)):
ngram = tuple(map(lambda x:x.lower(),sent[i:i+4]))
if ngram in ngramsets['quad']:
return True, i, i+4
for i in range(len(sent)):
ngram = tuple(map(lambda x:x.lower(),sent[i:i+3]))
if ngram in ngramsets['tri']:
return True, i, i+3
for i in range(len(sent)):
ngram = tuple(map(lambda x:x.lower(),sent[i:i+2]))
if ngram in ngramsets['bi']:
return True, i, i+2
return False, 0, 0
def ngramize_body(sent):
continue_gram,i,j = find_gram(sent)
if continue_gram:
return ngramize_body(sent[:i]),[' '.join(sent[i:j])],ngramize_body(sent[j:])
return sent
def traverse_tree(node):
if isinstance(node,tuple):
tmp=[]
for n in node:
if isinstance(n,list):
tmp.extend(n)
else:
tmp.extend(traverse_tree(n))
return tmp
else:
return node
def ngramize(sent):
sent_tree = ngramize_body(sent)
return traverse_tree(sent_tree)
def isngram(token):
return ' ' in token
def shuffle_ngram(sent, ratio):
# this guarantee displacement of ratio *len(sent) tokens in sent
# if ratio*len(sent) <= 1, return completely shuffled sent
# return error if len(sent) == 1
if len(sent)==1:
tmp=sent[0].split(' ')
random.shuffle(tmp)
return tmp
if round(len(sent)*ratio)<=1:
random.shuffle(sent)
return sent
shuff_id = random.sample([i for i in range(len(sent))],round(len(sent)*ratio))
ngram_id = sorted(shuff_id)
while any(ngram_id[i] == shuff_id[i] for i in range(len(ngram_id))):
random.shuffle(ngram_id)
tmp=copy.deepcopy(sent)
for i,j in zip(ngram_id,shuff_id):
tmp[i]=sent[j]
return tmp
def shuffle_text_by_ngram(sents,ratio):
return [shuffle_ngram(ngramize(sent),ratio) for sent in sents]
def generate_shuffle_text_by_ngram(path,sents,ratio):
if os.path.exists(path):
print('skipping '+path + ': already exists')
else:
text = shuffle_text_by_ngram(sents,ratio)
for i, sent in enumerate(text):
text[i]=' '.join(sent)
with open(path,'w') as f:
f.write('\n'.join(text))
def displace_ngram(sent, prob):
#prob: probability that a displacement occurs
def find_init(seg):
return random.randint(0,len(seg)-2)
def flip(biconstituent):
# biconstituent is a list of 2 item
assert len(biconstituent)==2
if random.random() > prob:
biconstituent.reverse()
return ' '.join(biconstituent)
while len(sent)>1:
if len(sent)==2:
sent = [flip(sent)]
continue
i = find_init(sent)
sent=sent[:i]+[flip(sent[i:i+2])]+sent[i+2:]
return sent[0] if len(sent)==1 else ''
def displace_text_by_ngram(sents, prob):
return [displace_ngram(ngramize(sent),prob) for sent in sents]
def generate_adjacency_displaced_text(path,sents,prob):
if os.path.exists(path):
print('skipping '+path + ': already exists')
else:
text=displace_text_by_ngram(sents,prob)
with open(path,'w') as f:
f.write('\n'.join(text))
def main():
train=load_data(train_path)
valid=load_data(valid_path)
test=load_data(test_path)
for r in ratios:
generate_shuffle_text_by_ngram(destination_path+'ptb2016.train.'+'shuffle.'+str(r)+'_p'+str(prune_threshold)+'.txt',train,r)
generate_shuffle_text_by_ngram(destination_path+'ptb2016.valid.'+'shuffle.'+str(r)+'_p'+str(prune_threshold)+'.txt',valid,r)
generate_shuffle_text_by_ngram(destination_path+'ptb2016.test.'+'shuffle.'+str(r)+'_p'+str(prune_threshold)+'.txt',test,r)
for p in flip_prob:
generate_adjacency_displaced_text(destination_path+'ptb2016.train.'+'displace.'+str(p)+'_p'+str(prune_threshold)+'.txt',train,p)
generate_adjacency_displaced_text(destination_path+'ptb2016.valid.'+'displace.'+str(p)+'_p'+str(prune_threshold)+'.txt',valid,p)
generate_adjacency_displaced_text(destination_path+'ptb2016.test.'+'displace.'+str(p)+'_p'+str(prune_threshold)+'.txt',test,p)
if __name__ == "__main__":
main()