-
Notifications
You must be signed in to change notification settings - Fork 0
/
rewriter.py
302 lines (260 loc) · 14.3 KB
/
rewriter.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
from utils.data_utils import Candidate, OutputExample
from collections import defaultdict
from cbr import build_ngrams
import logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
class GenderRewriter:
def __init__(self, cbr_model, morph_rewriter, rbr_model, neural_rewriter,
gender_identifier,
ranker=None,
first_person_only=False):
"""
Args:
- cbr_model (default dict): The cbr model where the
keys are (source_word, target_word_gender) and vals
are target_words.
- morph_rewriter (MorphR obj): The morph rewriter.
- rbr_model (default dict): The rbr model where the keys are
(target_word_gender, source_pattern) and vals are dicts
of target_patterns
- neural_rewriter (NeuralR obj): The neural rewriter.
- ranker (Ranker object): The ranker.
- first_person_only (bool): To do *only* first person gender
rewriting.
"""
self.cbr_model = cbr_model
self.morph_rewriter = morph_rewriter
self.ranker = ranker
self.gender_identifier = gender_identifier
self.rbr_model = rbr_model
self.neural_rewriter = neural_rewriter
self.first_person_only = first_person_only
def get_base_and_clitic_target_gender(self, tag, speaker_gender,
listener_gender):
base_word_gender, clitic_gender = tag.split('+')
target_word_gender, target_clitic_gender = base_word_gender, clitic_gender
if base_word_gender == '1M' and speaker_gender == '1F':
target_word_gender = '1F'
elif base_word_gender == '1F' and speaker_gender == '1M':
target_word_gender = '1M'
elif base_word_gender == '2M' and listener_gender == '2F':
target_word_gender = '2F'
elif base_word_gender == '2F' and listener_gender == '2M':
target_word_gender = '2M'
if clitic_gender == '1M' and speaker_gender == '1F':
target_clitic_gender = '1F'
elif clitic_gender == '1F' and speaker_gender == '1M':
target_clitic_gender = '1M'
elif clitic_gender == '2M' and listener_gender == '2F':
target_clitic_gender = '2F'
elif clitic_gender == '2F' and listener_gender == '2M':
target_clitic_gender = '2M'
return f'{target_word_gender}+{target_clitic_gender}'
def rewrite(self, dataset, speaker_gender, listener_gender=None,
use_cbr=True, pick_top_mle=True, reduce_cbr_noise=True,
use_morph=True, use_rbr=True, use_neural=True):
"""
Args:
- dataset (Dataset object): contains list the input examples.
The input examples *must* contain bert
predicted gender tags.
- speaker_gender (str): M or F.
- listener_gender (str): M or F or None (if we're doing 1st per
gender rewriting only).
- use_cbr (bool): to use cbr model or not.
- use_morph (bool): to use the morphological rewriter.
- use_rbr (bool): to use rbr model or not.
- use_neural (bool): to use the neural model or not.
Returns:
- candidates (list): candidate objects.
"""
candidates = []
inf_analysis_stats = defaultdict(int)
oov_stats = defaultdict(int)
if not self.first_person_only:
# Adding person annotations to be compatible with token annoations
speaker_gender = '1' + speaker_gender
listener_gender = '2' + listener_gender
else:
speaker_gender = '1' + speaker_gender
assert listener_gender == None
for i, example in enumerate(dataset):
src_tokens = example.src_tokens
pred_tags = self.gender_identifier.predict_sentence(src_tokens)
# building ngrams for the CBR model if needed
if use_cbr:
tokens_ngrams = build_ngrams(src_tokens,
ngrams=self.cbr_model.ngrams,
pad_left=True)
candidate_sentence = []
candidate_targets = []
proposed_by = []
for j, (token, tag) in enumerate(zip(src_tokens, pred_tags)):
# if the token tag is B+B or if it matches
# the speaker gender tag or the listener gender, pass the
# token as it is
if (tag == 'B+B' or
tag == f'B+{listener_gender}' or
tag == f'{listener_gender}+B' or
tag == f'B+{speaker_gender}' or
tag == f'{speaker_gender}+B' or
tag == f'{speaker_gender}+{listener_gender}' or
tag == f'{listener_gender}+{speaker_gender}' or
tag == f'{listener_gender}+{listener_gender}' or
tag == f'{speaker_gender}+{speaker_gender}'):
candidate_sentence.append(token)
proposed_by.append('NA')
inf_analysis_stats['reg_passes'] += 1
else:
# Getting the target gender based on the provided
# user preferences and predicted token tag
if self.first_person_only:
target_gender = self.get_base_and_clitic_target_gender(tag,
listener_gender=None,
speaker_gender=speaker_gender)
else:
target_gender = self.get_base_and_clitic_target_gender(tag,
listener_gender=listener_gender,
speaker_gender=speaker_gender)
is_oov = True
# use CBR model to get gender alts.
if use_cbr:
cbr_candidates = self.cbr_model[(tokens_ngrams[j],
target_gender)]
inf_analysis_stats['cbr_triggers'] += 1
if cbr_candidates:
# if there are multiple
# alternatives and pick_top_mle, get the most
# probable alternative. Otherwise, create a mask
# sentence and expand targets
is_oov = False
if len(cbr_candidates) > 1:
if pick_top_mle:
rewritten_token = max(cbr_candidates.items(),
key=lambda x: x[1])[0]
candidate_sentence.append(rewritten_token)
proposed_by.append('CBR')
else:
candidate_sentence.append('[MASK]')
token_targets = list(cbr_candidates.keys())
# removing the token itself if it appears
# within the targets
if reduce_cbr_noise and token in token_targets:
token_targets.remove(token)
candidate_targets.append(token_targets)
proposed_by.append('CBR')
else:
# if there is a single alternative, return it.
# but if the generated option is equal
# to the input token, don't return it and
# consider the token to be an OOV.
if reduce_cbr_noise and list(cbr_candidates.keys())[0] == token:
is_oov = True
else:
candidate_sentence.append(list(cbr_candidates.keys())[0])
proposed_by.append('CBR')
else:
oov_stats['cbr_oov'] += 1
# use morphR if CBR fail
# or as a stand alone model
if ((use_cbr and use_morph and is_oov) or
(use_morph and is_oov)):
inf_analysis_stats['morph_triggers'] += 1
morph_res = self.morph_rewriter.rewrite(token,
tag,
target_gender)
if morph_res:
rewritten_token = morph_res['rewritten_token']
token_candidates = morph_res['proposals']
proposal_src = morph_res['proposed_by']
is_oov = False
candidate_sentence.append(rewritten_token)
proposed_by.append(proposal_src)
# If there are multiple targets, expand targets
if len(token_candidates) != 0:
assert rewritten_token == '[MASK]'
candidate_targets.append(token_candidates)
else:
oov_stats['morph_oov'] += 1
# use RBR if morph or CBR (or both) fail
# or as a stand alone model
if ((use_cbr and use_morph and use_rbr and is_oov) or
(use_rbr and is_oov)):
rbr_candidates = self.rbr_model[(target_gender, token,
tag)]
if rbr_candidates:
inf_analysis_stats['rbr_triggers'] += 1
is_oov = False
if len(rbr_candidates) > 1:
candidate_sentence.append('[MASK]')
candidate_targets.append(rbr_candidates)
proposed_by.append('RBR')
else:
rewritten_token = rbr_candidates[0]
candidate_sentence.append(rewritten_token)
proposed_by.append('RBR')
else:
oov_stats['rbr_oov'] += 1
# use neuralR if RBR or morphR or CBR (or all) fail
# or as a stand alone model
if ((use_cbr and use_morph and use_rbr and use_neural
and is_oov) or (use_neural and is_oov)):
neural_candidates = self.neural_rewriter.rewrite(token=token,
target_gender=target_gender)
inf_analysis_stats['neural_triggers'] += 1
is_oov = False
if len(neural_candidates) > 1:
candidate_sentence.append('[MASK]')
candidate_targets.append(neural_candidates)
proposed_by.append('neuralR')
else:
rewritten_token = neural_candidates[0]
candidate_sentence.append(rewritten_token)
proposed_by.append('neuralR')
if is_oov:
# If everything fails, pass the oov word as it is
candidate_sentence.append(token)
proposed_by.append('OOV')
candidates.append(Candidate(masked_sentence=candidate_sentence,
pred_src_gen_tags=pred_tags,
targets=candidate_targets,
proposed_by=proposed_by))
for i, candidate in enumerate(candidates):
if candidate.targets:
inf_analysis_stats['selection_triggers'] += 1
logger.info(f"CBR triggers: {inf_analysis_stats['cbr_triggers']}")
logger.info(f"Morph triggers: {inf_analysis_stats['morph_triggers']}")
logger.info(f"RBR triggers: {inf_analysis_stats['rbr_triggers']}")
logger.info(f"Neural triggers: {inf_analysis_stats['neural_triggers']}")
logger.info(f"Regular passes: {inf_analysis_stats['reg_passes']}")
logger.info(f"Selection triggers: {inf_analysis_stats['selection_triggers']}")
logger.info("===========================")
logger.info(f"CBR OOV: {oov_stats['cbr_oov']}")
logger.info(f"Morph OOV: {oov_stats['morph_oov']}")
logger.info(f"RBR OOV: {oov_stats['rbr_oov']}")
logger.info("===========================")
return candidates
def select(self, candidates):
"""
Args:
- candidates (list): candidate objects.
Returns:
- gender_alts (list): output example objects.
"""
gender_alts = []
for i, candidate in enumerate(candidates):
sentence = ' '.join(candidate.masked_sentence)
if candidate.targets:
scored = self.ranker.fill_and_rank(sentence=sentence,
targets=candidate.targets)
gender_alts.append(OutputExample(sentence=scored[0].sentence,
scored_candidates=scored[1:],
pred_src_gen_tags=candidate.pred_src_gen_tags,
proposed_by=candidate.proposed_by))
else:
gender_alts.append(OutputExample(sentence=sentence,
scored_candidates=None,
pred_src_gen_tags=candidate.pred_src_gen_tags,
proposed_by=candidate.proposed_by))
return gender_alts