-
Notifications
You must be signed in to change notification settings - Fork 1
/
align_sentence.py
264 lines (211 loc) · 8.99 KB
/
align_sentence.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
import pprint as pp
import re
import os, sys
import time
import tensorflow_hub as hub
import numpy as np
import tensorflow_text
import argparse
embed = hub.load("https://tfhub.dev/google/universal-sentence-encoder-multilingual/3")
# from transformers import MarianMTModel, MarianTokenizer
# tokenizer_ja_en = MarianTokenizer.from_pretrained('Helsinki-NLP/opus-mt-ja-en')
# model_ja_en = MarianMTModel.from_pretrained('Helsinki-NLP/opus-mt-ja-en')
import MeCab
wakati = MeCab.Tagger("-Owakati")
from bs4 import BeautifulSoup
from nltk.tokenize import sent_tokenize
import nltk
nltk.download('punkt')
def cos_sim(lhs, rhs):
return np.dot(lhs, rhs) / (np.linalg.norm(lhs) * np.linalg.norm(rhs))
def get_html(txt):
ret = txt[txt.find("<html"):txt.find("</html")] + "</html>"
return ret.replace("<CR>", "\r").replace("<LF>", "\n").replace("<TAB>", " ")
def get_article(html, url, bp, mode = 'ja'):
sp = BeautifulSoup(html, 'html.parser')
ret = "".join([n.get_text() for n in sp.select(bp)])
if len(ret) > 100:
return ret
return ""
def clean_txt(txt):
ret = re.sub(r"^\r?\n?", "", txt.replace(u'\xa0', u' '))
ret = re.sub(r"(\r\n)+", "\n", ret)
ret = re.sub(r"\n+", "\n", ret)
ret = re.sub(r"\r+", "\n", ret)
ret = re.sub(r"^[\.\,]?[\s\r\t\n]*", "", ret)
ret = re.sub(r"[\s\r\t\n]+$", "", ret)
return ret
def to_one_line(txt):
return re.sub("^(\s|\W)+", "", txt.replace("\r", "").replace("\n", "").replace("\t", ""))
def sent_tokenize_ja(txt):
jp_sent_tokenizer = nltk.RegexpTokenizer(u'[^ 「」！？。]*[！？。]')
ary = jp_sent_tokenizer.tokenize(txt)
return [v for v in filter(lambda x: len(x) > 3, [clean_txt(s) for s in ary])]
# If you process news articles, it's better than the above method.
def sent_tokenize_ja_news(txt):
ary = re.split("(。|！|？|\n|\s\s)",txt)
for i in range(len(ary) - 1):
if re.search("(「|「)", ary[i]) and not re.search("(」|」)", ary[i]) and re.search("(」|」)", ary[i+1]) :
ary[i] = ary[i] + ary[i+1]
i+=1
return ["{}。".format(v) for v in filter(lambda x: len(x) > 3, [re.sub("^[\s\r\n]+","",clean_txt(s)) for s in ary])]
def sent_tokenize_en(txt):
ary = sent_tokenize(clean_txt(re.sub("\.", ". ", txt)))
return [v for v in filter(lambda x: len(x) > 3, [clean_txt(s) for s in ary])]
def get_align(ja_sents, en_sents, threshold=0.4, decay=0.05):
ja_vec = embed(ja_sents)
en_vec = embed(en_sents)
offsets = []
for jidx, jv in enumerate(ja_vec):
max_sim, ofs = max([(cos_sim(jv, ev), eidx - jidx) for eidx, ev in enumerate(en_vec)])
if max_sim > threshold and abs(ofs) < 10:
offsets.append(ofs)
if len(offsets):
off_set = sum(offsets) / len(offsets)
else:
off_set = 0
ret = []
jidx = 0
while jidx < len(ja_vec):
jv = ja_vec[jidx]
max_sim, eidx = max(
[((1.0 - min(abs(eidx - (jidx + off_set)) * decay, 1.0)) * cos_sim(jv, ev), eidx) for eidx, ev in
enumerate(en_vec)])
add_below_sim_ja = 0.0
add_below_sim_en = 0.0
if max_sim < threshold:
jidx += 1
continue
if jidx + 1 < len(ja_vec):
add_below_sim_ja = (1.0 - min(abs(eidx - (jidx + off_set)) * decay, 1.0)) * cos_sim(
embed([ja_sents[jidx] + ja_sents[jidx + 1]])[0], en_vec[eidx])
if eidx + 1 < len(en_vec):
add_below_sim_en = (1.0 - min(abs(eidx - (jidx + off_set)) * decay, 1.0)) * cos_sim(ja_vec[jidx], embed(
[en_sents[eidx] + en_sents[eidx + 1]])[0])
if add_below_sim_ja > max_sim + decay * 2 and add_below_sim_ja > threshold + decay:
ret.append((add_below_sim_ja, ja_sents[jidx] + ja_sents[jidx + 1], en_sents[eidx]))
jidx += 1
off_set -= 1.0
elif add_below_sim_en > max_sim + decay * 2 and add_below_sim_en > threshold + decay:
ret.append((add_below_sim_en, ja_sents[jidx], en_sents[eidx] + en_sents[eidx + 1]))
off_set += 1.0
elif max_sim > threshold:
ret.append((max_sim, ja_sents[jidx], en_sents[eidx]))
jidx += 1
return ret
def get_title(html, url, mode):
sp = BeautifulSoup(html, 'html.parser')
return "".join([n.get_text() for n in sp.select('title')])
def get_description(html, url, mode):
sp = BeautifulSoup(html, 'html.parser')
return sp.find('meta', attrs={'property': 'og:description'})["content"]
def count_words(sentence):
return len(nltk.word_tokenize(sentence))
def filter_aligns(align_sents, threshold=0.7, mode="mecab"):
# sort and uniq_en
en_sents = {}
ret = []
if mode != "translate":
threshold = threshold * 0.9
for tpl in sorted(align_sents, reverse=True):
en_s = tpl[2]
if en_s in en_sents:
continue
else:
en_sents[en_s] = 1
ret.append(tpl)
# check word count
txts = [v[1] for v in ret]
# batch = tokenizer_ja_en.prepare_translation_batch(src_texts=txts)
# gen = model_ja_en.generate(**batch)
# txt_en = tokenizer_ja_en.batch_decode(gen, skip_special_tokens=True)
if mode == "translate":
# translated = model_ja_en.generate(**tokenizer_ja_en.prepare_translation_batch(txts))
# txt_en = [tokenizer_ja_en.decode(t, skip_special_tokens=True) for t in translated]
pass
else:
txt_en = [wakati.parse(s).split() for s in txts]
#print(txt_en)
filtered_ret = []
for idx, tpl in enumerate(ret):
if mode == "translate":
ja_en_words = 1.0 + count_words(txt_en[idx])
else:
ja_en_words = 1.0 + len(txt_en[idx])
original_en_words = 1.0 + count_words(tpl[2])
if threshold < ja_en_words / original_en_words and ja_en_words / original_en_words < 1.0 / threshold:
filtered_ret.append(tpl)
return filtered_ret
def filter_sents(sents):
for_filter_sents = ['Share this with', 'Email', 'Twitter', 'Facebook','Pinterest', 'WhatsApp',
'Messenger','Hatena', 'Mixi', 'Line', 'このリンクをコピーする', 'LinkedIn']
def _filter_match(lhs, rhs):
if (rhs.find(lhs) != -1 or lhs.find(rhs) != -1) and abs(len(lhs) - len(rhs)) < 5:
return True
return False
ret = []
for s in sents:
if max([_filter_match(s, fs) for fs in for_filter_sents]):
continue
else:
ret.append(s)
return ret
def get_align_html(ja_html, en_html, url = "", ja_bp="p", en_bp="p"):
align_texts = []
# ja_title = get_title(ja_html, url, "ja")
# en_title = get_title(en_html, url, "en")
#
# sim_title = cos_sim(embed([ja_title])[0], embed([en_title])[0])
# if sim_title > 0.5:
# align_texts.append((sim_title, ja_title, en_title))
# ja_description = get_description(ja_html, url, "ja")
# en_description = get_description(en_html, url, "en")
# sim_description = cos_sim(embed([ja_description])[0], embed([en_description])[0])
# if sim_description > 0.5:
# align_texts.append((sim_description, ja_description, en_description))
ja_article = get_article(ja_html, url, ja_bp)
ja_article = re.sub("(\r\n|\r|\n|\t){2,}", ". ", ja_article)
en_article = get_article(en_html, url, en_bp)
en_article = re.sub("(\r\n|\r|\n|\t){2,}", ". ", en_article)
ja_sents = filter_sents(sent_tokenize_ja(clean_txt(ja_article)))
en_sents = filter_sents(sent_tokenize_en(clean_txt(en_article)))
if len(ja_sents) < 50:
threshold = 0.3
elif len(ja_sents) < 100:
threshold = 0.4
else:
threshold = 0.5
if re.search("抄訳", ja_article):
threshold = 0.6
if len(ja_sents) * 2 < len(en_sents):
threshold = 0.6
align_sents = get_align(ja_sents, en_sents, threshold=threshold)
for tpl in align_sents:
print("align_candidate=\t{}\t{}\t{}\n".format(to_one_line(tpl[1]), to_one_line(tpl[2]), tpl[0]))
align_sents_filter = filter_aligns(align_sents)
print("original={} filter={} \t url={}".format(len(align_sents), len(align_sents_filter), url))
align_texts += align_sents_filter
return align_texts
def main():
parser = argparse.ArgumentParser(description='align en_sentence and ja_sentence')
parser.add_argument('ja_html', help='html file (lang=ja)')
parser.add_argument('en_html', help='html file (lang=en)')
parser.add_argument('--ja_bp', help='BeautifulSoup selector for ja_html')
parser.add_argument('--en_bp', help='BeautifulSoup selector for en_html')
args = parser.parse_args()
ja_html_file = args.ja_html
en_html_file = args.en_html
ja_bp = "p"
if args.ja_bp:
ja_bp = args.ja_bp
en_bp = "p"
if args.en_bp:
en_bp = args.en_bp
ja_html = open(ja_html_file).read()
en_html = open(en_html_file).read()
aligns = get_align_html(ja_html, en_html, ja_bp=ja_bp, en_bp=en_bp)
for tpl in aligns:
print("{}\t{}\t{}\n".format(to_one_line(tpl[1]), to_one_line(tpl[2]), tpl[0]))
return
if __name__ == "__main__":
main()