|
1
|
|
|
# pylint: disable=too-many-lines |
|
2
|
|
|
""" |
|
3
|
|
|
Measuring and adjusting bias in word embedding by Bolukbasi (2016). |
|
4
|
|
|
|
|
5
|
|
|
References: |
|
6
|
|
|
- Bolukbasi, T., Chang, K. W., Zou, J. Y., Saligrama, V., |
|
7
|
|
|
& Kalai, A. T. (2016). |
|
8
|
|
|
`Man is to computer programmer as woman is to homemaker? |
|
9
|
|
|
debiasing word embeddings <https://arxiv.org/abs/1607.06520>`_. |
|
10
|
|
|
In Advances in neural information processing systems |
|
11
|
|
|
(pp. 4349-4357). |
|
12
|
|
|
|
|
13
|
|
|
- The code and data is based on the GitHub repository: |
|
14
|
|
|
https://github.com/tolga-b/debiaswe (MIT License). |
|
15
|
|
|
|
|
16
|
|
|
- Gonen, H., & Goldberg, Y. (2019). |
|
17
|
|
|
`Lipstick on a Pig: |
|
18
|
|
|
Debiasing Methods Cover up Systematic Gender Biases |
|
19
|
|
|
in Word Embeddings But do not Remove Them |
|
20
|
|
|
<https://arxiv.org/abs/1903.03862>`_. |
|
21
|
|
|
arXiv preprint arXiv:1903.03862. |
|
22
|
|
|
|
|
23
|
|
|
- Nissim, M., van Noord, R., van der Goot, R. (2019). |
|
24
|
|
|
`Fair is Better than Sensational: Man is to Doctor |
|
25
|
|
|
as Woman is to Doctor <https://arxiv.org/abs/1905.09866>`_. |
|
26
|
|
|
|
|
27
|
|
|
Usage |
|
28
|
|
|
~~~~~ |
|
29
|
|
|
|
|
30
|
|
|
.. code:: python |
|
31
|
|
|
|
|
32
|
|
|
>>> from ethically.we import GenderBiasWE |
|
33
|
|
|
>>> from gensim import downloader |
|
34
|
|
|
>>> w2v_model = downloader.load('word2vec-google-news-300') |
|
35
|
|
|
>>> w2v_gender_bias_we = GenderBiasWE(w2v_model) |
|
36
|
|
|
>>> w2v_gender_bias_we.calc_direct_bias() |
|
37
|
|
|
0.07307904249481942 |
|
38
|
|
|
>>> w2v_gender_bias_we.debias() |
|
39
|
|
|
>>> w2v_gender_bias_we.calc_direct_bias() |
|
40
|
|
|
1.7964246601064155e-09 |
|
41
|
|
|
|
|
42
|
|
|
Types of Bias |
|
43
|
|
|
~~~~~~~~~~~~~ |
|
44
|
|
|
|
|
45
|
|
|
Direct Bias |
|
46
|
|
|
^^^^^^^^^^^ |
|
47
|
|
|
|
|
48
|
|
|
1. Associations |
|
49
|
|
|
Words that are closer to one end (e.g., *he*) than to |
|
50
|
|
|
the other end (*she*). |
|
51
|
|
|
For example, occupational stereotypes (page 7). |
|
52
|
|
|
Calculated by |
|
53
|
|
|
:meth:`~ethically.we.bias.BiasWordEmbedding.calc_direct_bias`. |
|
54
|
|
|
|
|
55
|
|
|
2. Analogies |
|
56
|
|
|
Analogies of *he:x::she:y*. |
|
57
|
|
|
For example analogies exhibiting stereotypes (page 7). |
|
58
|
|
|
Generated by |
|
59
|
|
|
:meth:`~ethically.we.bias.BiasWordEmbedding.generate_analogies`. |
|
60
|
|
|
|
|
61
|
|
|
|
|
62
|
|
|
Indirect Bias |
|
63
|
|
|
^^^^^^^^^^^^^ |
|
64
|
|
|
|
|
65
|
|
|
Projection of a neutral words into a two neutral words direction |
|
66
|
|
|
is explained in a great portion by a shared bias direction projection. |
|
67
|
|
|
|
|
68
|
|
|
Calculated by |
|
69
|
|
|
:meth:`~ethically.we.bias.BiasWordEmbedding.calc_indirect_bias` |
|
70
|
|
|
and |
|
71
|
|
|
:meth:`~ethically.we.bias.GenderBiasWE.generate_closest_words_indirect_bias`. |
|
72
|
|
|
|
|
73
|
|
|
""" |
|
74
|
|
|
|
|
75
|
|
|
import copy |
|
76
|
|
|
import warnings |
|
77
|
|
|
|
|
78
|
|
|
import matplotlib.pylab as plt |
|
79
|
|
|
import numpy as np |
|
80
|
|
|
import pandas as pd |
|
81
|
|
|
import seaborn as sns |
|
82
|
|
|
from scipy.stats import spearmanr |
|
83
|
|
|
from sklearn.decomposition import PCA |
|
84
|
|
|
from sklearn.metrics.pairwise import euclidean_distances |
|
85
|
|
|
from sklearn.svm import LinearSVC |
|
86
|
|
|
from tqdm import tqdm |
|
87
|
|
|
|
|
88
|
|
|
from tabulate import tabulate |
|
89
|
|
|
|
|
90
|
|
|
from ..consts import RANDOM_STATE |
|
91
|
|
|
from .benchmark import evaluate_word_embedding |
|
92
|
|
|
from .data import BOLUKBASI_DATA |
|
93
|
|
|
from .utils import ( |
|
94
|
|
|
assert_gensim_keyed_vectors, cosine_similarity, generate_one_word_forms, |
|
95
|
|
|
generate_words_forms, get_seed_vector, most_similar, normalize, |
|
96
|
|
|
plot_clustering_as_classification, project_params, project_reject_vector, |
|
97
|
|
|
project_vector, reject_vector, round_to_extreme, |
|
98
|
|
|
take_two_sides_extreme_sorted, update_word_vector, |
|
99
|
|
|
) |
|
100
|
|
|
|
|
101
|
|
|
|
|
102
|
|
|
DIRECTION_METHODS = ['single', 'sum', 'pca'] |
|
103
|
|
|
DEBIAS_METHODS = ['neutralize', 'hard', 'soft'] |
|
104
|
|
|
FIRST_PC_THRESHOLD = 0.5 |
|
105
|
|
|
MAX_NON_SPECIFIC_EXAMPLES = 1000 |
|
106
|
|
|
|
|
107
|
|
|
__all__ = ['GenderBiasWE', 'BiasWordEmbedding'] |
|
108
|
|
|
|
|
109
|
|
|
# https://stackoverflow.com/questions/2187269/print-only-the-message-on-warnings |
|
110
|
|
|
formatwarning_orig = warnings.formatwarning |
|
111
|
|
|
# pylint: disable=line-too-long |
|
112
|
|
|
warnings.formatwarning = lambda message, category, filename, lineno, line=None: \ |
|
113
|
|
|
formatwarning_orig(message, category, filename, lineno, line='') |
|
114
|
|
|
|
|
115
|
|
|
|
|
116
|
|
|
class BiasWordEmbedding: |
|
117
|
|
|
"""Measure and adjust a bias in English word embedding. |
|
118
|
|
|
|
|
119
|
|
|
:param model: Word embedding model of ``gensim.model.KeyedVectors`` |
|
120
|
|
|
:param bool only_lower: Whether the word embedding contrains |
|
121
|
|
|
only lower case words |
|
122
|
|
|
:param bool verbose: Set verbosity |
|
123
|
|
|
:param bool to_normalize: Whether to normalize all the vectors |
|
124
|
|
|
(recommended!) |
|
125
|
|
|
""" |
|
126
|
|
|
|
|
127
|
|
|
def __init__(self, model, only_lower=False, verbose=False, |
|
128
|
|
|
identify_direction=False, to_normalize=True): |
|
129
|
|
|
assert_gensim_keyed_vectors(model) |
|
130
|
|
|
|
|
131
|
|
|
# TODO: this is bad Python, ask someone about it |
|
132
|
|
|
# probably should be a better design |
|
133
|
|
|
# identify_direction doesn't have any meaning |
|
134
|
|
|
# for the class BiasWordEmbedding |
|
135
|
|
|
if self.__class__ == __class__ and identify_direction is not False: |
|
|
|
|
|
|
136
|
|
|
raise ValueError('identify_direction must be False' |
|
137
|
|
|
' for an instance of {}' |
|
138
|
|
|
.format(__class__)) |
|
139
|
|
|
|
|
140
|
|
|
self.model = model |
|
141
|
|
|
|
|
142
|
|
|
# TODO: write unitest for when it is False |
|
143
|
|
|
self.only_lower = only_lower |
|
144
|
|
|
|
|
145
|
|
|
self._verbose = verbose |
|
146
|
|
|
|
|
147
|
|
|
self.direction = None |
|
148
|
|
|
self.positive_end = None |
|
149
|
|
|
self.negative_end = None |
|
150
|
|
|
|
|
151
|
|
|
if to_normalize: |
|
152
|
|
|
self.model.init_sims(replace=True) |
|
153
|
|
|
|
|
154
|
|
|
def __copy__(self): |
|
155
|
|
|
bias_word_embedding = self.__class__(self.model, |
|
156
|
|
|
self.only_lower, |
|
157
|
|
|
self._verbose, |
|
158
|
|
|
identify_direction=False) |
|
159
|
|
|
bias_word_embedding.direction = copy.deepcopy(self.direction) |
|
160
|
|
|
bias_word_embedding.positive_end = copy.deepcopy(self.positive_end) |
|
161
|
|
|
bias_word_embedding.negative_end = copy.deepcopy(self.negative_end) |
|
162
|
|
|
return bias_word_embedding |
|
163
|
|
|
|
|
164
|
|
|
def __deepcopy__(self, memo): |
|
165
|
|
|
bias_word_embedding = copy.copy(self) |
|
166
|
|
|
bias_word_embedding.model = copy.deepcopy(bias_word_embedding.model) |
|
167
|
|
|
return bias_word_embedding |
|
168
|
|
|
|
|
169
|
|
|
def __getitem__(self, key): |
|
170
|
|
|
return self.model[key] |
|
171
|
|
|
|
|
172
|
|
|
def __contains__(self, item): |
|
173
|
|
|
return item in self.model |
|
174
|
|
|
|
|
175
|
|
|
def _filter_words_by_model(self, words): |
|
176
|
|
|
return [word for word in words if word in self] |
|
177
|
|
|
|
|
178
|
|
|
def _is_direction_identified(self): |
|
179
|
|
|
if self.direction is None: |
|
180
|
|
|
raise RuntimeError('The direction was not identified' |
|
181
|
|
|
' for this {} instance' |
|
182
|
|
|
.format(self.__class__.__name__)) |
|
183
|
|
|
|
|
184
|
|
|
# There is a mistake in the article |
|
185
|
|
|
# it is written (section 5.1): |
|
186
|
|
|
# "To identify the gender subspace, we took the ten gender pair difference |
|
187
|
|
|
# vectors and computed its principal components (PCs)" |
|
188
|
|
|
# however in the source code: |
|
189
|
|
|
# https://github.com/tolga-b/debiaswe/blob/10277b23e187ee4bd2b6872b507163ef4198686b/debiaswe/we.py#L235-L245 |
|
190
|
|
|
def _identify_subspace_by_pca(self, definitional_pairs, n_components): |
|
191
|
|
|
matrix = [] |
|
192
|
|
|
|
|
193
|
|
|
for word1, word2 in definitional_pairs: |
|
194
|
|
|
vector1 = normalize(self[word1]) |
|
195
|
|
|
vector2 = normalize(self[word2]) |
|
196
|
|
|
|
|
197
|
|
|
center = (vector1 + vector2) / 2 |
|
198
|
|
|
|
|
199
|
|
|
matrix.append(vector1 - center) |
|
200
|
|
|
matrix.append(vector2 - center) |
|
201
|
|
|
|
|
202
|
|
|
pca = PCA(n_components=n_components) |
|
203
|
|
|
pca.fit(matrix) |
|
204
|
|
|
|
|
205
|
|
|
if self._verbose: |
|
206
|
|
|
table = enumerate(pca.explained_variance_ratio_, start=1) |
|
207
|
|
|
headers = ['Principal Component', |
|
208
|
|
|
'Explained Variance Ratio'] |
|
209
|
|
|
print(tabulate(table, headers=headers)) |
|
210
|
|
|
|
|
211
|
|
|
return pca |
|
212
|
|
|
|
|
213
|
|
|
# TODO: add the SVD method from section 6 step 1 |
|
214
|
|
|
# It seems there is a mistake there, I think it is the same as PCA |
|
215
|
|
|
# just with replacing it with SVD |
|
216
|
|
|
def _identify_direction(self, positive_end, negative_end, |
|
217
|
|
|
definitional, method='pca'): |
|
218
|
|
|
if method not in DIRECTION_METHODS: |
|
219
|
|
|
raise ValueError('method should be one of {}, {} was given'.format( |
|
220
|
|
|
DIRECTION_METHODS, method)) |
|
221
|
|
|
|
|
222
|
|
|
if positive_end == negative_end: |
|
223
|
|
|
raise ValueError('positive_end and negative_end' |
|
224
|
|
|
'should be different, and not the same "{}"' |
|
225
|
|
|
.format(positive_end)) |
|
226
|
|
|
if self._verbose: |
|
227
|
|
|
print('Identify direction using {} method...'.format(method)) |
|
228
|
|
|
|
|
229
|
|
|
direction = None |
|
230
|
|
|
|
|
231
|
|
|
if method == 'single': |
|
232
|
|
|
direction = normalize(normalize(self[definitional[0]]) |
|
233
|
|
|
- normalize(self[definitional[1]])) |
|
234
|
|
|
|
|
235
|
|
|
elif method == 'sum': |
|
236
|
|
|
group1_sum_vector = np.sum([self[word] |
|
237
|
|
|
for word in definitional[0]], axis=0) |
|
238
|
|
|
group2_sum_vector = np.sum([self[word] |
|
239
|
|
|
for word in definitional[1]], axis=0) |
|
240
|
|
|
|
|
241
|
|
|
diff_vector = (normalize(group1_sum_vector) |
|
242
|
|
|
- normalize(group2_sum_vector)) |
|
243
|
|
|
|
|
244
|
|
|
direction = normalize(diff_vector) |
|
245
|
|
|
|
|
246
|
|
|
elif method == 'pca': |
|
247
|
|
|
pca = self._identify_subspace_by_pca(definitional, 10) |
|
248
|
|
|
if pca.explained_variance_ratio_[0] < FIRST_PC_THRESHOLD: |
|
249
|
|
|
raise RuntimeError('The Explained variance' |
|
250
|
|
|
'of the first principal component should be' |
|
251
|
|
|
'at least {}, but it is {}' |
|
252
|
|
|
.format(FIRST_PC_THRESHOLD, |
|
253
|
|
|
pca.explained_variance_ratio_[0])) |
|
254
|
|
|
direction = pca.components_[0] |
|
255
|
|
|
|
|
256
|
|
|
# if direction is opposite (e.g. we cannot control |
|
257
|
|
|
# what the PCA will return) |
|
258
|
|
|
ends_diff_projection = cosine_similarity((self[positive_end] |
|
259
|
|
|
- self[negative_end]), |
|
260
|
|
|
direction) |
|
261
|
|
|
if ends_diff_projection < 0: |
|
262
|
|
|
direction = -direction # pylint: disable=invalid-unary-operand-type |
|
263
|
|
|
|
|
264
|
|
|
self.direction = direction |
|
265
|
|
|
self.positive_end = positive_end |
|
266
|
|
|
self.negative_end = negative_end |
|
267
|
|
|
|
|
268
|
|
|
def project_on_direction(self, word): |
|
269
|
|
|
"""Project the normalized vector of the word on the direction. |
|
270
|
|
|
|
|
271
|
|
|
:param str word: The word tor project |
|
272
|
|
|
:return float: The projection scalar |
|
273
|
|
|
""" |
|
274
|
|
|
|
|
275
|
|
|
self._is_direction_identified() |
|
276
|
|
|
|
|
277
|
|
|
vector = self[word] |
|
278
|
|
|
projection_score = self.model.cosine_similarities(self.direction, |
|
279
|
|
|
[vector])[0] |
|
280
|
|
|
return projection_score |
|
281
|
|
|
|
|
282
|
|
|
def _calc_projection_scores(self, words): |
|
283
|
|
|
self._is_direction_identified() |
|
284
|
|
|
|
|
285
|
|
|
df = pd.DataFrame({'word': words}) |
|
286
|
|
|
|
|
287
|
|
|
# TODO: maybe using cosine_similarities on all the vectors? |
|
288
|
|
|
# it might be faster |
|
289
|
|
|
df['projection'] = df['word'].apply(self.project_on_direction) |
|
290
|
|
|
df = df.sort_values('projection', ascending=False) |
|
291
|
|
|
|
|
292
|
|
|
return df |
|
293
|
|
|
|
|
294
|
|
|
def calc_projection_data(self, words): |
|
295
|
|
|
""" |
|
296
|
|
|
Calculate projection, projected and rejected vectors of a words list. |
|
297
|
|
|
|
|
298
|
|
|
:param list words: List of words |
|
299
|
|
|
:return: :class:`pandas.DataFrame` of the projection, |
|
300
|
|
|
projected and rejected vectors of the words list |
|
301
|
|
|
""" |
|
302
|
|
|
projection_data = [] |
|
303
|
|
|
for word in words: |
|
304
|
|
|
vector = self[word] |
|
305
|
|
|
projection = self.project_on_direction(word) |
|
306
|
|
|
normalized_vector = normalize(vector) |
|
307
|
|
|
|
|
308
|
|
|
(projection, |
|
309
|
|
|
projected_vector, |
|
310
|
|
|
rejected_vector) = project_params(normalized_vector, |
|
311
|
|
|
self.direction) |
|
312
|
|
|
|
|
313
|
|
|
projection_data.append({'word': word, |
|
314
|
|
|
'vector': vector, |
|
315
|
|
|
'projection': projection, |
|
316
|
|
|
'projected_vector': projected_vector, |
|
317
|
|
|
'rejected_vector': rejected_vector}) |
|
318
|
|
|
|
|
319
|
|
|
return pd.DataFrame(projection_data) |
|
320
|
|
|
|
|
321
|
|
|
def plot_projection_scores(self, words, n_extreme=10, |
|
322
|
|
|
ax=None, axis_projection_step=None): |
|
323
|
|
|
"""Plot the projection scalar of words on the direction. |
|
324
|
|
|
|
|
325
|
|
|
:param list words: The words tor project |
|
326
|
|
|
:param int or None n_extreme: The number of extreme words to show |
|
327
|
|
|
:return: The ax object of the plot |
|
328
|
|
|
""" |
|
329
|
|
|
|
|
330
|
|
|
self._is_direction_identified() |
|
331
|
|
|
|
|
332
|
|
|
projections_df = self._calc_projection_scores(words) |
|
333
|
|
|
projections_df['projection'] = projections_df['projection'].round(2) |
|
334
|
|
|
|
|
335
|
|
|
if n_extreme is not None: |
|
336
|
|
|
projections_df = take_two_sides_extreme_sorted(projections_df, |
|
337
|
|
|
n_extreme=n_extreme) |
|
338
|
|
|
|
|
339
|
|
|
if ax is None: |
|
340
|
|
|
_, ax = plt.subplots(1) |
|
341
|
|
|
|
|
342
|
|
|
if axis_projection_step is None: |
|
343
|
|
|
axis_projection_step = 0.1 |
|
344
|
|
|
|
|
345
|
|
|
cmap = plt.get_cmap('RdBu') |
|
346
|
|
|
projections_df['color'] = ((projections_df['projection'] + 0.5) |
|
347
|
|
|
.apply(cmap)) |
|
348
|
|
|
|
|
349
|
|
|
most_extream_projection = (projections_df['projection'] |
|
350
|
|
|
.abs() |
|
351
|
|
|
.max() |
|
352
|
|
|
.round(1)) |
|
353
|
|
|
|
|
354
|
|
|
sns.barplot(x='projection', y='word', data=projections_df, |
|
355
|
|
|
palette=projections_df['color']) |
|
356
|
|
|
|
|
357
|
|
|
plt.xticks(np.arange(-most_extream_projection, |
|
358
|
|
|
most_extream_projection + axis_projection_step, |
|
359
|
|
|
axis_projection_step)) |
|
360
|
|
|
plt.title('← {} {} {} →'.format(self.negative_end, |
|
361
|
|
|
' ' * 20, |
|
362
|
|
|
self.positive_end)) |
|
363
|
|
|
|
|
364
|
|
|
plt.xlabel('Direction Projection') |
|
365
|
|
|
plt.ylabel('Words') |
|
366
|
|
|
|
|
367
|
|
|
return ax |
|
368
|
|
|
|
|
369
|
|
|
def plot_dist_projections_on_direction(self, word_groups, ax=None): |
|
370
|
|
|
"""Plot the projection scalars distribution on the direction. |
|
371
|
|
|
|
|
372
|
|
|
:param dict word_groups word: The groups to projects |
|
373
|
|
|
:return float: The ax object of the plot |
|
374
|
|
|
""" |
|
375
|
|
|
|
|
376
|
|
|
if ax is None: |
|
377
|
|
|
_, ax = plt.subplots(1) |
|
378
|
|
|
|
|
379
|
|
|
names = sorted(word_groups.keys()) |
|
380
|
|
|
|
|
381
|
|
|
for name in names: |
|
382
|
|
|
words = word_groups[name] |
|
383
|
|
|
label = '{} (#{})'.format(name, len(words)) |
|
384
|
|
|
vectors = [self[word] for word in words] |
|
385
|
|
|
projections = self.model.cosine_similarities(self.direction, |
|
386
|
|
|
vectors) |
|
387
|
|
|
sns.distplot(projections, hist=False, label=label, ax=ax) |
|
388
|
|
|
|
|
389
|
|
|
plt.axvline(0, color='k', linestyle='--') |
|
390
|
|
|
|
|
391
|
|
|
plt.title('← {} {} {} →'.format(self.negative_end, |
|
392
|
|
|
' ' * 20, |
|
393
|
|
|
self.positive_end)) |
|
394
|
|
|
plt.xlabel('Direction Projection') |
|
395
|
|
|
plt.ylabel('Density') |
|
396
|
|
|
ax.legend(loc='center left', bbox_to_anchor=(1, 0.5)) |
|
397
|
|
|
|
|
398
|
|
|
return ax |
|
399
|
|
|
|
|
400
|
|
|
@classmethod |
|
401
|
|
|
def _calc_bias_across_word_embeddings(cls, |
|
402
|
|
|
word_embedding_bias_dict, |
|
403
|
|
|
words): |
|
404
|
|
|
""" |
|
405
|
|
|
Calculate to projections and rho of words for two word embeddings. |
|
406
|
|
|
|
|
407
|
|
|
:param dict word_embedding_bias_dict: ``WordsEmbeddingBias`` objects |
|
408
|
|
|
as values, |
|
409
|
|
|
and their names as keys. |
|
410
|
|
|
:param list words: Words to be projected. |
|
411
|
|
|
:return tuple: Projections and spearman rho. |
|
412
|
|
|
""" |
|
413
|
|
|
# pylint: disable=W0212 |
|
414
|
|
|
assert len(word_embedding_bias_dict) == 2, 'Support only in two'\ |
|
415
|
|
|
'word embeddings' |
|
416
|
|
|
|
|
417
|
|
|
intersection_words = [word for word in words |
|
418
|
|
|
if all(word in web |
|
419
|
|
|
for web in (word_embedding_bias_dict |
|
420
|
|
|
.values()))] |
|
421
|
|
|
|
|
422
|
|
|
projections = {name: web._calc_projection_scores(intersection_words)['projection'] # pylint: disable=C0301 |
|
423
|
|
|
for name, web in word_embedding_bias_dict.items()} |
|
424
|
|
|
|
|
425
|
|
|
df = pd.DataFrame(projections) |
|
426
|
|
|
df.index = intersection_words |
|
427
|
|
|
|
|
428
|
|
|
rho, _ = spearmanr(*df.transpose().values) |
|
429
|
|
|
return df, rho |
|
430
|
|
|
|
|
431
|
|
|
@classmethod |
|
432
|
|
|
def plot_bias_across_word_embeddings(cls, word_embedding_bias_dict, |
|
433
|
|
|
words, ax=None, scatter_kwargs=None): |
|
434
|
|
|
""" |
|
435
|
|
|
Plot the projections of same words of two word mbeddings. |
|
436
|
|
|
|
|
437
|
|
|
:param dict word_embedding_bias_dict: ``WordsEmbeddingBias`` objects |
|
438
|
|
|
as values, |
|
439
|
|
|
and their names as keys. |
|
440
|
|
|
:param list words: Words to be projected. |
|
441
|
|
|
:param scatter_kwargs: Kwargs for matplotlib.pylab.scatter. |
|
442
|
|
|
:type scatter_kwargs: dict or None |
|
443
|
|
|
:return: The ax object of the plot |
|
444
|
|
|
""" |
|
445
|
|
|
# pylint: disable=W0212 |
|
446
|
|
|
|
|
447
|
|
|
df, rho = cls._calc_bias_across_word_embeddings(word_embedding_bias_dict, # pylint: disable=C0301 |
|
448
|
|
|
words) |
|
449
|
|
|
|
|
450
|
|
|
if ax is None: |
|
451
|
|
|
_, ax = plt.subplots(1) |
|
452
|
|
|
|
|
453
|
|
|
if scatter_kwargs is None: |
|
454
|
|
|
scatter_kwargs = {} |
|
455
|
|
|
|
|
456
|
|
|
name1, name2 = word_embedding_bias_dict.keys() |
|
457
|
|
|
|
|
458
|
|
|
ax.scatter(x=name1, y=name2, data=df, **scatter_kwargs) |
|
459
|
|
|
|
|
460
|
|
|
plt.title('Bias Across Word Embeddings' |
|
461
|
|
|
'(Spearman Rho = {:0.2f})'.format(rho)) |
|
462
|
|
|
|
|
463
|
|
|
negative_end = word_embedding_bias_dict[name1].negative_end |
|
464
|
|
|
positive_end = word_embedding_bias_dict[name1].positive_end |
|
465
|
|
|
plt.xlabel('← {} {} {} →'.format(negative_end, |
|
466
|
|
|
name1, |
|
467
|
|
|
positive_end)) |
|
468
|
|
|
plt.ylabel('← {} {} {} →'.format(negative_end, |
|
469
|
|
|
name2, |
|
470
|
|
|
positive_end)) |
|
471
|
|
|
|
|
472
|
|
|
ax_min = round_to_extreme(df.values.min()) |
|
473
|
|
|
ax_max = round_to_extreme(df.values.max()) |
|
474
|
|
|
plt.xlim(ax_min, ax_max) |
|
475
|
|
|
plt.ylim(ax_min, ax_max) |
|
476
|
|
|
|
|
477
|
|
|
return ax |
|
478
|
|
|
|
|
479
|
|
|
# TODO: refactor for speed and clarity |
|
480
|
|
|
def generate_analogies(self, n_analogies=100, seed='ends', |
|
481
|
|
|
multiple=False, |
|
482
|
|
|
delta=1., restrict_vocab=30000, |
|
483
|
|
|
unrestricted=False): |
|
484
|
|
|
""" |
|
485
|
|
|
Generate analogies based on a seed vector. |
|
486
|
|
|
|
|
487
|
|
|
x - y ~ seed vector. |
|
488
|
|
|
or a:x::b:y when a-b ~ seed vector. |
|
489
|
|
|
|
|
490
|
|
|
The seed vector can be defined by two word ends, |
|
491
|
|
|
or by the bias direction. |
|
492
|
|
|
|
|
493
|
|
|
``delta`` is used for semantically coherent. Default vale of 1 |
|
494
|
|
|
corresponds to an angle <= pi/3. |
|
495
|
|
|
|
|
496
|
|
|
|
|
497
|
|
|
There is criticism regarding generating analogies |
|
498
|
|
|
when used with `unstricted=False` and not ignoring analogies |
|
499
|
|
|
with `match` column equal to `False`. |
|
500
|
|
|
Tolga's technique of generating analogies, as implemented in this |
|
501
|
|
|
method, is limited inherently to analogies with x != y, which may |
|
502
|
|
|
be force "fake" bias analogies. |
|
503
|
|
|
|
|
504
|
|
|
See: |
|
505
|
|
|
|
|
506
|
|
|
- Nissim, M., van Noord, R., van der Goot, R. (2019). |
|
507
|
|
|
`Fair is Better than Sensational: Man is to Doctor |
|
508
|
|
|
as Woman is to Doctor <https://arxiv.org/abs/1905.09866>`_. |
|
509
|
|
|
|
|
510
|
|
|
:param seed: The definition of the seed vector. |
|
511
|
|
|
Either by a tuple of two word ends, |
|
512
|
|
|
or by `'ends` for the pre-defined ends |
|
513
|
|
|
or by `'direction'` for the pre-defined direction vector. |
|
514
|
|
|
:param int n_analogies: Number of analogies to generate. |
|
515
|
|
|
:param bool multiple: Whether to allow multiple appearances of a word |
|
516
|
|
|
in the analogies. |
|
517
|
|
|
:param float delta: Threshold for semantic similarity. |
|
518
|
|
|
The maximal distance between x and y. |
|
519
|
|
|
:param int restrict_vocab: The vocabulary size to use. |
|
520
|
|
|
:param bool unrestricted: Whether to validate the generated analogies |
|
521
|
|
|
with unrestricted `most_similar`. |
|
522
|
|
|
:return: Data Frame of analogies (x, y), their distances, |
|
523
|
|
|
and their cosine similarity scores |
|
524
|
|
|
""" |
|
525
|
|
|
# pylint: disable=C0301,R0914 |
|
526
|
|
|
|
|
527
|
|
|
if not unrestricted: |
|
528
|
|
|
warnings.warn('Not Using unrestricted most_similar ' |
|
529
|
|
|
'may introduce fake biased analogies.') |
|
530
|
|
|
|
|
531
|
|
|
(seed_vector, |
|
532
|
|
|
positive_end, |
|
533
|
|
|
negative_end) = get_seed_vector(seed, self) |
|
534
|
|
|
|
|
535
|
|
|
restrict_vocab_vectors = self.model.vectors[:restrict_vocab] |
|
536
|
|
|
|
|
537
|
|
|
normalized_vectors = (restrict_vocab_vectors |
|
538
|
|
|
/ np.linalg.norm(restrict_vocab_vectors, axis=1)[:, None]) |
|
539
|
|
|
|
|
540
|
|
|
pairs_distances = euclidean_distances(normalized_vectors, normalized_vectors) |
|
541
|
|
|
|
|
542
|
|
|
# `pairs_distances` must be not-equal to zero |
|
543
|
|
|
# otherwise, x-y will be the zero vector, and every cosine similarity |
|
544
|
|
|
# will be equal to zero. |
|
545
|
|
|
# This cause to the **limitation** of this method which enforce a not-same |
|
546
|
|
|
# words for x and y. |
|
547
|
|
|
pairs_mask = (pairs_distances < delta) & (pairs_distances != 0) |
|
548
|
|
|
|
|
549
|
|
|
pairs_indices = np.array(np.nonzero(pairs_mask)).T |
|
550
|
|
|
x_vectors = np.take(normalized_vectors, pairs_indices[:, 0], axis=0) |
|
551
|
|
|
y_vectors = np.take(normalized_vectors, pairs_indices[:, 1], axis=0) |
|
552
|
|
|
|
|
553
|
|
|
x_minus_y_vectors = x_vectors - y_vectors |
|
554
|
|
|
normalized_x_minus_y_vectors = (x_minus_y_vectors |
|
555
|
|
|
/ np.linalg.norm(x_minus_y_vectors, axis=1)[:, None]) |
|
556
|
|
|
|
|
557
|
|
|
cos_distances = normalized_x_minus_y_vectors @ seed_vector |
|
558
|
|
|
|
|
559
|
|
|
sorted_cos_distances_indices = np.argsort(cos_distances)[::-1] |
|
560
|
|
|
|
|
561
|
|
|
sorted_cos_distances_indices_iter = iter(sorted_cos_distances_indices) |
|
562
|
|
|
|
|
563
|
|
|
analogies = [] |
|
564
|
|
|
generated_words_x = set() |
|
565
|
|
|
generated_words_y = set() |
|
566
|
|
|
|
|
567
|
|
|
while len(analogies) < n_analogies: |
|
568
|
|
|
cos_distance_index = next(sorted_cos_distances_indices_iter) |
|
569
|
|
|
paris_index = pairs_indices[cos_distance_index] |
|
570
|
|
|
word_x, word_y = [self.model.index2word[index] |
|
571
|
|
|
for index in paris_index] |
|
572
|
|
|
|
|
573
|
|
|
if multiple or (not multiple |
|
574
|
|
|
and (word_x not in generated_words_x |
|
575
|
|
|
and word_y not in generated_words_y)): |
|
576
|
|
|
|
|
577
|
|
|
analogy = ({positive_end: word_x, |
|
578
|
|
|
negative_end: word_y, |
|
579
|
|
|
'score': cos_distances[cos_distance_index], |
|
580
|
|
|
'distance': pairs_distances[tuple(paris_index)]}) |
|
581
|
|
|
|
|
582
|
|
|
generated_words_x.add(word_x) |
|
583
|
|
|
generated_words_y.add(word_y) |
|
584
|
|
|
|
|
585
|
|
|
if unrestricted: |
|
586
|
|
|
most_x = next(word |
|
587
|
|
|
for word, _ in most_similar(self.model, |
|
588
|
|
|
[word_y, positive_end], |
|
589
|
|
|
[negative_end])) |
|
590
|
|
|
most_y = next(word |
|
591
|
|
|
for word, _ in most_similar(self.model, |
|
592
|
|
|
[word_x, negative_end], |
|
593
|
|
|
[positive_end])) |
|
594
|
|
|
|
|
595
|
|
|
analogy['most_x'] = most_x |
|
596
|
|
|
analogy['most_y'] = most_y |
|
597
|
|
|
analogy['match'] = ((word_x == most_x) |
|
598
|
|
|
and (word_y == most_y)) |
|
599
|
|
|
|
|
600
|
|
|
analogies.append(analogy) |
|
601
|
|
|
|
|
602
|
|
|
df = pd.DataFrame(analogies) |
|
603
|
|
|
|
|
604
|
|
|
columns = [positive_end, negative_end, 'distance', 'score'] |
|
605
|
|
|
|
|
606
|
|
|
if unrestricted: |
|
607
|
|
|
columns.extend(['most_x', 'most_y', 'match']) |
|
608
|
|
|
|
|
609
|
|
|
df = df[columns] |
|
610
|
|
|
|
|
611
|
|
|
return df |
|
612
|
|
|
|
|
613
|
|
|
def calc_direct_bias(self, neutral_words, c=None): |
|
614
|
|
|
"""Calculate the direct bias. |
|
615
|
|
|
|
|
616
|
|
|
Based on the projection of neutral words on the direction. |
|
617
|
|
|
|
|
618
|
|
|
:param list neutral_words: List of neutral words |
|
619
|
|
|
:param c: Strictness of bias measuring |
|
620
|
|
|
:type c: float or None |
|
621
|
|
|
:return: The direct bias |
|
622
|
|
|
""" |
|
623
|
|
|
|
|
624
|
|
|
if c is None: |
|
625
|
|
|
c = 1 |
|
626
|
|
|
|
|
627
|
|
|
projections = self._calc_projection_scores(neutral_words)['projection'] |
|
628
|
|
|
direct_bias_terms = np.abs(projections) ** c |
|
629
|
|
|
direct_bias = direct_bias_terms.sum() / len(neutral_words) |
|
630
|
|
|
|
|
631
|
|
|
return direct_bias |
|
632
|
|
|
|
|
633
|
|
|
def calc_indirect_bias(self, word1, word2): |
|
634
|
|
|
"""Calculate the indirect bias between two words. |
|
635
|
|
|
|
|
636
|
|
|
Based on the amount of shared projection of the words on the direction. |
|
637
|
|
|
|
|
638
|
|
|
Also called PairBias. |
|
639
|
|
|
:param str word1: First word |
|
640
|
|
|
:param str word2: Second word |
|
641
|
|
|
:type c: float or None |
|
642
|
|
|
:return The indirect bias between the two words |
|
643
|
|
|
""" |
|
644
|
|
|
|
|
645
|
|
|
self._is_direction_identified() |
|
646
|
|
|
|
|
647
|
|
|
vector1 = normalize(self[word1]) |
|
648
|
|
|
vector2 = normalize(self[word2]) |
|
649
|
|
|
|
|
650
|
|
|
perpendicular_vector1 = reject_vector(vector1, self.direction) |
|
651
|
|
|
perpendicular_vector2 = reject_vector(vector2, self.direction) |
|
652
|
|
|
|
|
653
|
|
|
inner_product = vector1 @ vector2 |
|
654
|
|
|
perpendicular_similarity = cosine_similarity(perpendicular_vector1, |
|
655
|
|
|
perpendicular_vector2) |
|
656
|
|
|
|
|
657
|
|
|
indirect_bias = ((inner_product - perpendicular_similarity) |
|
658
|
|
|
/ inner_product) |
|
659
|
|
|
return indirect_bias |
|
660
|
|
|
|
|
661
|
|
|
def generate_closest_words_indirect_bias(self, |
|
662
|
|
|
neutral_positive_end, |
|
663
|
|
|
neutral_negative_end, |
|
664
|
|
|
words=None, n_extreme=5): |
|
665
|
|
|
""" |
|
666
|
|
|
Generate closest words to a neutral direction and their indirect bias. |
|
667
|
|
|
|
|
668
|
|
|
The direction of the neutral words is used to find |
|
669
|
|
|
the most extreme words. |
|
670
|
|
|
The indirect bias is calculated between the most extreme words |
|
671
|
|
|
and the closest end. |
|
672
|
|
|
|
|
673
|
|
|
:param str neutral_positive_end: A word that define the positive side |
|
674
|
|
|
of the neutral direction. |
|
675
|
|
|
:param str neutral_negative_end: A word that define the negative side |
|
676
|
|
|
of the neutral direction. |
|
677
|
|
|
:param list words: List of words to project on the neutral direction. |
|
678
|
|
|
:param int n_extreme: The number for the most extreme words |
|
679
|
|
|
(positive and negative) to show. |
|
680
|
|
|
:return: Data Frame of the most extreme words |
|
681
|
|
|
with their projection scores and indirect biases. |
|
682
|
|
|
""" |
|
683
|
|
|
|
|
684
|
|
|
neutral_direction = normalize(self[neutral_positive_end] |
|
685
|
|
|
- self[neutral_negative_end]) |
|
686
|
|
|
|
|
687
|
|
|
vectors = [normalize(self[word]) for word in words] |
|
688
|
|
|
df = (pd.DataFrame([{'word': word, |
|
689
|
|
|
'projection': vector @ neutral_direction} |
|
690
|
|
|
for word, vector in zip(words, vectors)]) |
|
691
|
|
|
.sort_values('projection', ascending=False)) |
|
692
|
|
|
|
|
693
|
|
|
df = take_two_sides_extreme_sorted(df, n_extreme, |
|
694
|
|
|
'end', |
|
695
|
|
|
neutral_positive_end, |
|
696
|
|
|
neutral_negative_end) |
|
697
|
|
|
|
|
698
|
|
|
df['indirect_bias'] = df.apply(lambda r: |
|
699
|
|
|
self.calc_indirect_bias(r['word'], |
|
700
|
|
|
r['end']), |
|
701
|
|
|
axis=1) |
|
702
|
|
|
|
|
703
|
|
|
df = df.set_index(['end', 'word']) |
|
704
|
|
|
df = df[['projection', 'indirect_bias']] |
|
705
|
|
|
|
|
706
|
|
|
return df |
|
707
|
|
|
|
|
708
|
|
|
def _extract_neutral_words(self, specific_words): |
|
709
|
|
|
extended_specific_words = set() |
|
710
|
|
|
|
|
711
|
|
|
# because or specific_full data was trained on partial word embedding |
|
712
|
|
|
for word in specific_words: |
|
713
|
|
|
extended_specific_words.add(word) |
|
714
|
|
|
extended_specific_words.add(word.lower()) |
|
715
|
|
|
extended_specific_words.add(word.upper()) |
|
716
|
|
|
extended_specific_words.add(word.title()) |
|
717
|
|
|
|
|
718
|
|
|
neutral_words = [word for word in self.model.vocab |
|
719
|
|
|
if word not in extended_specific_words] |
|
720
|
|
|
|
|
721
|
|
|
return neutral_words |
|
722
|
|
|
|
|
723
|
|
|
def _neutralize(self, neutral_words): |
|
724
|
|
|
self._is_direction_identified() |
|
725
|
|
|
|
|
726
|
|
|
if self._verbose: |
|
727
|
|
|
neutral_words_iter = tqdm(neutral_words) |
|
728
|
|
|
else: |
|
729
|
|
|
neutral_words_iter = iter(neutral_words) |
|
730
|
|
|
|
|
731
|
|
|
for word in neutral_words_iter: |
|
732
|
|
|
neutralized_vector = reject_vector(self[word], |
|
733
|
|
|
self.direction) |
|
734
|
|
|
update_word_vector(self.model, word, neutralized_vector) |
|
735
|
|
|
|
|
736
|
|
|
self.model.init_sims(replace=True) |
|
737
|
|
|
|
|
738
|
|
|
def _equalize(self, equality_sets): |
|
739
|
|
|
# pylint: disable=R0914 |
|
740
|
|
|
|
|
741
|
|
|
self._is_direction_identified() |
|
742
|
|
|
|
|
743
|
|
|
if self._verbose: |
|
744
|
|
|
words_data = [] |
|
745
|
|
|
|
|
746
|
|
|
for equality_set_index, equality_set_words in enumerate(equality_sets): |
|
747
|
|
|
equality_set_vectors = [normalize(self[word]) |
|
748
|
|
|
for word in equality_set_words] |
|
749
|
|
|
center = np.mean(equality_set_vectors, axis=0) |
|
750
|
|
|
(projected_center, |
|
751
|
|
|
rejected_center) = project_reject_vector(center, |
|
752
|
|
|
self.direction) |
|
753
|
|
|
scaling = np.sqrt(1 - np.linalg.norm(rejected_center)**2) |
|
754
|
|
|
|
|
755
|
|
|
for word, vector in zip(equality_set_words, equality_set_vectors): |
|
756
|
|
|
projected_vector = project_vector(vector, self.direction) |
|
757
|
|
|
|
|
758
|
|
|
projected_part = normalize(projected_vector - projected_center) |
|
759
|
|
|
|
|
760
|
|
|
# In the code it is different of Bolukbasi |
|
761
|
|
|
# It behaves the same only for equality_sets |
|
762
|
|
|
# with size of 2 (pairs) - not sure! |
|
763
|
|
|
# However, my code is the same as the article |
|
764
|
|
|
# equalized_vector = rejected_center + scaling * self.direction |
|
765
|
|
|
# https://github.com/tolga-b/debiaswe/blob/10277b23e187ee4bd2b6872b507163ef4198686b/debiaswe/debias.py#L36-L37 |
|
766
|
|
|
# For pairs, projected_part_vector1 == -projected_part_vector2, |
|
767
|
|
|
# and this is the same as |
|
768
|
|
|
# projected_part_vector1 == self.direction |
|
769
|
|
|
equalized_vector = rejected_center + scaling * projected_part |
|
770
|
|
|
|
|
771
|
|
|
update_word_vector(self.model, word, equalized_vector) |
|
772
|
|
|
|
|
773
|
|
|
if self._verbose: |
|
774
|
|
|
words_data.append({ |
|
|
|
|
|
|
775
|
|
|
'equality_set_index': equality_set_index, |
|
776
|
|
|
'word': word, |
|
777
|
|
|
'scaling': scaling, |
|
778
|
|
|
'projected_scalar': vector @ self.direction, |
|
779
|
|
|
'equalized_projected_scalar': (equalized_vector |
|
780
|
|
|
@ self.direction), |
|
781
|
|
|
}) |
|
782
|
|
|
|
|
783
|
|
|
if self._verbose: |
|
784
|
|
|
print('Equalize Words Data ' |
|
785
|
|
|
'(all equal for 1-dim bias space (direction):') |
|
786
|
|
|
words_data_df = (pd.DataFrame(words_data) |
|
787
|
|
|
.set_index(['equality_set_index', 'word'])) |
|
788
|
|
|
print(tabulate(words_data_df, headers='keys')) |
|
789
|
|
|
|
|
790
|
|
|
self.model.init_sims(replace=True) |
|
791
|
|
|
|
|
792
|
|
|
def debias(self, method='hard', neutral_words=None, equality_sets=None, |
|
793
|
|
|
inplace=True): |
|
794
|
|
|
"""Debias the word embedding. |
|
795
|
|
|
|
|
796
|
|
|
:param str method: The method of debiasing. |
|
797
|
|
|
:param list neutral_words: List of neutral words |
|
798
|
|
|
for the neutralize step |
|
799
|
|
|
:param list equality_sets: List of equality sets, |
|
800
|
|
|
for the equalize step. |
|
801
|
|
|
The sets represent the direction. |
|
802
|
|
|
:param bool inplace: Whether to debias the object inplace |
|
803
|
|
|
or return a new one |
|
804
|
|
|
|
|
805
|
|
|
.. warning:: |
|
806
|
|
|
|
|
807
|
|
|
After calling `debias`, |
|
808
|
|
|
all the vectors of the word embedding |
|
809
|
|
|
will be normalized to unit length. |
|
810
|
|
|
|
|
811
|
|
|
""" |
|
812
|
|
|
|
|
813
|
|
|
# pylint: disable=W0212 |
|
814
|
|
|
if inplace: |
|
815
|
|
|
bias_word_embedding = self |
|
816
|
|
|
else: |
|
817
|
|
|
bias_word_embedding = copy.deepcopy(self) |
|
818
|
|
|
|
|
819
|
|
|
if method not in DEBIAS_METHODS: |
|
820
|
|
|
raise ValueError('method should be one of {}, {} was given'.format( |
|
821
|
|
|
DEBIAS_METHODS, method)) |
|
822
|
|
|
|
|
823
|
|
|
if method in ['hard', 'neutralize']: |
|
824
|
|
|
if self._verbose: |
|
825
|
|
|
print('Neutralize...') |
|
826
|
|
|
bias_word_embedding._neutralize(neutral_words) |
|
827
|
|
|
|
|
828
|
|
|
if method == 'hard': |
|
829
|
|
|
if self._verbose: |
|
830
|
|
|
print('Equalize...') |
|
831
|
|
|
bias_word_embedding._equalize(equality_sets) |
|
832
|
|
|
|
|
833
|
|
|
if inplace: |
|
834
|
|
|
return None |
|
835
|
|
|
else: |
|
836
|
|
|
return bias_word_embedding |
|
837
|
|
|
|
|
838
|
|
|
def evaluate_word_embedding(self, |
|
839
|
|
|
kwargs_word_pairs=None, |
|
840
|
|
|
kwargs_word_analogies=None): |
|
841
|
|
|
""" |
|
842
|
|
|
Evaluate word pairs tasks and word analogies tasks. |
|
843
|
|
|
|
|
844
|
|
|
:param model: Word embedding. |
|
845
|
|
|
:param kwargs_word_pairs: Kwargs for |
|
846
|
|
|
evaluate_word_pairs |
|
847
|
|
|
method. |
|
848
|
|
|
:type kwargs_word_pairs: dict or None |
|
849
|
|
|
:param kwargs_word_analogies: Kwargs for |
|
850
|
|
|
evaluate_word_analogies |
|
851
|
|
|
method. |
|
852
|
|
|
:type evaluate_word_analogies: dict or None |
|
853
|
|
|
:return: Tuple of :class:`pandas.DataFrame` |
|
854
|
|
|
for the evaluation results. |
|
855
|
|
|
""" |
|
856
|
|
|
|
|
857
|
|
|
return evaluate_word_embedding(self.model, |
|
858
|
|
|
kwargs_word_pairs, |
|
859
|
|
|
kwargs_word_analogies) |
|
860
|
|
|
|
|
861
|
|
|
def learn_full_specific_words(self, seed_specific_words, |
|
862
|
|
|
max_non_specific_examples=None, debug=None): |
|
863
|
|
|
"""Learn specific words given a list of seed specific wordsself. |
|
864
|
|
|
|
|
865
|
|
|
Using Linear SVM. |
|
866
|
|
|
|
|
867
|
|
|
:param list seed_specific_words: List of seed specific words |
|
868
|
|
|
:param int max_non_specific_examples: The number of non-specific words |
|
869
|
|
|
to sample for training |
|
870
|
|
|
:return: List of learned specific words and the classifier object |
|
871
|
|
|
""" |
|
872
|
|
|
|
|
873
|
|
|
if debug is None: |
|
874
|
|
|
debug = False |
|
875
|
|
|
|
|
876
|
|
|
if max_non_specific_examples is None: |
|
877
|
|
|
max_non_specific_examples = MAX_NON_SPECIFIC_EXAMPLES |
|
878
|
|
|
|
|
879
|
|
|
data = [] |
|
880
|
|
|
non_specific_example_count = 0 |
|
881
|
|
|
|
|
882
|
|
|
for word in self.model.vocab: |
|
883
|
|
|
is_specific = word in seed_specific_words |
|
884
|
|
|
|
|
885
|
|
|
if not is_specific: |
|
886
|
|
|
non_specific_example_count += 1 |
|
887
|
|
|
if non_specific_example_count <= max_non_specific_examples: |
|
888
|
|
|
data.append((self[word], is_specific)) |
|
889
|
|
|
else: |
|
890
|
|
|
data.append((self[word], is_specific)) |
|
891
|
|
|
|
|
892
|
|
|
np.random.seed(RANDOM_STATE) |
|
893
|
|
|
np.random.shuffle(data) |
|
894
|
|
|
|
|
895
|
|
|
X, y = zip(*data) |
|
896
|
|
|
|
|
897
|
|
|
X = np.array(X) |
|
898
|
|
|
X /= np.linalg.norm(X, axis=1)[:, None] |
|
899
|
|
|
|
|
900
|
|
|
y = np.array(y).astype('int') |
|
901
|
|
|
|
|
902
|
|
|
clf = LinearSVC(C=1, class_weight='balanced', |
|
903
|
|
|
random_state=RANDOM_STATE) |
|
904
|
|
|
|
|
905
|
|
|
clf.fit(X, y) |
|
906
|
|
|
|
|
907
|
|
|
full_specific_words = [] |
|
908
|
|
|
for word in self.model.vocab: |
|
909
|
|
|
vector = [normalize(self[word])] |
|
910
|
|
|
if clf.predict(vector): |
|
911
|
|
|
full_specific_words.append(word) |
|
912
|
|
|
|
|
913
|
|
|
if not debug: |
|
914
|
|
|
return full_specific_words, clf |
|
915
|
|
|
|
|
916
|
|
|
return full_specific_words, clf, X, y |
|
917
|
|
|
|
|
918
|
|
|
def _plot_most_biased_one_cluster(self, |
|
919
|
|
|
most_biased_neutral_words, y_bias, |
|
920
|
|
|
random_state=1, ax=None): |
|
921
|
|
|
most_biased_vectors = [self.model[word] |
|
922
|
|
|
for word in most_biased_neutral_words] |
|
923
|
|
|
|
|
924
|
|
|
return plot_clustering_as_classification(most_biased_vectors, |
|
925
|
|
|
y_bias, |
|
926
|
|
|
random_state=random_state, |
|
927
|
|
|
ax=ax) |
|
928
|
|
|
|
|
929
|
|
|
@staticmethod |
|
930
|
|
|
def plot_most_biased_clustering(biased, debiased, |
|
931
|
|
|
seed='ends', n_extreme=500, |
|
932
|
|
|
random_state=1): |
|
933
|
|
|
"""Plot clustering as classification of biased neutral words. |
|
934
|
|
|
|
|
935
|
|
|
:param biased: Biased word embedding of |
|
936
|
|
|
:class:`~ethically.we.bias.BiasWordEmbedding`. |
|
937
|
|
|
:param debiased: Debiased word embedding of |
|
938
|
|
|
:class:`~ethically.we.bias.BiasWordEmbedding`. |
|
939
|
|
|
:param seed: The definition of the seed vector. |
|
940
|
|
|
Either by a tuple of two word ends, |
|
941
|
|
|
or by `'ends` for the pre-defined ends |
|
942
|
|
|
or by `'direction'` for |
|
943
|
|
|
the pre-defined direction vector. |
|
944
|
|
|
:param n_extrem: The number of extreme biased |
|
945
|
|
|
neutral words to use. |
|
946
|
|
|
:return: Tuple of list of ax objects of the plot, |
|
947
|
|
|
and a dictionary with the most positive |
|
948
|
|
|
and negative words. |
|
949
|
|
|
|
|
950
|
|
|
Based on: |
|
951
|
|
|
|
|
952
|
|
|
- Gonen, H., & Goldberg, Y. (2019). |
|
953
|
|
|
`Lipstick on a Pig: |
|
954
|
|
|
Debiasing Methods Cover up Systematic Gender Biases |
|
955
|
|
|
in Word Embeddings But do not Remove Them |
|
956
|
|
|
<https://arxiv.org/abs/1903.03862>`_. |
|
957
|
|
|
arXiv preprint arXiv:1903.03862. |
|
958
|
|
|
|
|
959
|
|
|
- https://github.com/gonenhila/gender_bias_lipstick |
|
960
|
|
|
""" |
|
961
|
|
|
# pylint: disable=protected-access,too-many-locals,line-too-long |
|
962
|
|
|
|
|
963
|
|
|
assert biased.positive_end == debiased.positive_end, \ |
|
964
|
|
|
'Postive ends should be the same.' |
|
965
|
|
|
assert biased.negative_end == debiased.negative_end, \ |
|
966
|
|
|
'Negative ends should be the same.' |
|
967
|
|
|
|
|
968
|
|
|
seed_vector, _, _ = get_seed_vector(seed, biased) |
|
969
|
|
|
|
|
970
|
|
|
neutral_words = biased._data['neutral_words'] |
|
971
|
|
|
neutral_word_vectors = (biased[word] for word in neutral_words) |
|
|
|
|
|
|
972
|
|
|
neutral_word_projections = [(normalize(vector) @ seed_vector, word) |
|
973
|
|
|
for word, vector |
|
974
|
|
|
in zip(neutral_words, |
|
975
|
|
|
neutral_word_vectors)] |
|
976
|
|
|
|
|
977
|
|
|
neutral_word_projections.sort() |
|
978
|
|
|
|
|
979
|
|
|
_, most_negative_words = zip(*neutral_word_projections[:n_extreme]) |
|
980
|
|
|
_, most_positive_words = zip(*neutral_word_projections[-n_extreme:]) |
|
981
|
|
|
|
|
982
|
|
|
most_biased_neutral_words = most_negative_words + most_positive_words |
|
983
|
|
|
|
|
984
|
|
|
y_bias = [False] * n_extreme + [True] * n_extreme |
|
985
|
|
|
|
|
986
|
|
|
_, axes = plt.subplots(1, 2, figsize=(20, 5)) |
|
987
|
|
|
|
|
988
|
|
|
acc_biased = biased._plot_most_biased_one_cluster(most_biased_neutral_words, |
|
989
|
|
|
y_bias, |
|
990
|
|
|
random_state=random_state, |
|
991
|
|
|
ax=axes[0]) |
|
992
|
|
|
axes[0].set_title('Biased - Accuracy={}'.format(acc_biased)) |
|
993
|
|
|
|
|
994
|
|
|
acc_debiased = debiased._plot_most_biased_one_cluster(most_biased_neutral_words, |
|
995
|
|
|
y_bias, |
|
996
|
|
|
random_state=random_state, |
|
997
|
|
|
ax=axes[1]) |
|
998
|
|
|
axes[1].set_title('Debiased - Accuracy={}'.format(acc_debiased)) |
|
999
|
|
|
|
|
1000
|
|
|
return axes, {biased.positive_end: most_positive_words, |
|
1001
|
|
|
biased.negative_end: most_negative_words} |
|
1002
|
|
|
|
|
1003
|
|
|
|
|
1004
|
|
|
class GenderBiasWE(BiasWordEmbedding): |
|
1005
|
|
|
"""Measure and adjust the Gender Bias in English Word Embedding. |
|
1006
|
|
|
|
|
1007
|
|
|
:param model: Word embedding model of ``gensim.model.KeyedVectors`` |
|
1008
|
|
|
:param bool only_lower: Whether the word embedding contrains |
|
1009
|
|
|
only lower case words |
|
1010
|
|
|
:param bool verbose: Set verbosity |
|
1011
|
|
|
:param bool to_normalize: Whether to normalize all the vectors |
|
1012
|
|
|
(recommended!) |
|
1013
|
|
|
""" |
|
1014
|
|
|
|
|
1015
|
|
|
def __init__(self, model, only_lower=False, verbose=False, |
|
1016
|
|
|
identify_direction=True, to_normalize=True): |
|
1017
|
|
|
super().__init__(model=model, |
|
1018
|
|
|
only_lower=only_lower, |
|
1019
|
|
|
verbose=verbose, |
|
1020
|
|
|
to_normalize=True) |
|
1021
|
|
|
self._initialize_data() |
|
1022
|
|
|
if identify_direction: |
|
1023
|
|
|
self._identify_direction('she', 'he', |
|
1024
|
|
|
self._data['definitional_pairs'], |
|
1025
|
|
|
'pca') |
|
1026
|
|
|
|
|
1027
|
|
|
def _initialize_data(self): |
|
1028
|
|
|
self._data = copy.deepcopy(BOLUKBASI_DATA['gender']) |
|
1029
|
|
|
|
|
1030
|
|
|
if not self.only_lower: |
|
1031
|
|
|
self._data['specific_full_with_definitional_equalize'] = \ |
|
1032
|
|
|
generate_words_forms(self |
|
1033
|
|
|
._data['specific_full_with_definitional_equalize']) # pylint: disable=C0301 |
|
1034
|
|
|
|
|
1035
|
|
|
for key in self._data['word_group_keys']: |
|
1036
|
|
|
self._data[key] = (self._filter_words_by_model(self |
|
1037
|
|
|
._data[key])) |
|
1038
|
|
|
|
|
1039
|
|
|
self._data['neutral_words'] = self._extract_neutral_words(self |
|
1040
|
|
|
._data['specific_full_with_definitional_equalize']) # pylint: disable=C0301 |
|
1041
|
|
|
self._data['neutral_words'].sort() |
|
1042
|
|
|
self._data['word_group_keys'].append('neutral_words') |
|
1043
|
|
|
|
|
1044
|
|
|
def plot_projection_scores(self, words='professions', n_extreme=10, |
|
1045
|
|
|
ax=None, axis_projection_step=None): |
|
1046
|
|
|
if words == 'professions': |
|
1047
|
|
|
words = self._data['profession_names'] |
|
1048
|
|
|
|
|
1049
|
|
|
return super().plot_projection_scores(words, n_extreme, |
|
1050
|
|
|
ax, axis_projection_step) |
|
1051
|
|
|
|
|
1052
|
|
|
def plot_dist_projections_on_direction(self, word_groups='bolukbasi', |
|
1053
|
|
|
ax=None): |
|
1054
|
|
|
if word_groups == 'bolukbasi': |
|
1055
|
|
|
word_groups = {key: self._data[key] |
|
1056
|
|
|
for key in self._data['word_group_keys']} |
|
1057
|
|
|
|
|
1058
|
|
|
return super().plot_dist_projections_on_direction(word_groups, ax) |
|
1059
|
|
|
|
|
1060
|
|
|
@classmethod |
|
1061
|
|
|
def plot_bias_across_word_embeddings(cls, word_embedding_bias_dict, |
|
1062
|
|
|
ax=None, scatter_kwargs=None): |
|
1063
|
|
|
# pylint: disable=W0221 |
|
1064
|
|
|
words = BOLUKBASI_DATA['gender']['neutral_profession_names'] |
|
1065
|
|
|
# TODO: is it correct for inheritance of class method? |
|
1066
|
|
|
super(cls, cls).plot_bias_across_word_embeddings(word_embedding_bias_dict, # pylint: disable=C0301 |
|
1067
|
|
|
words, |
|
1068
|
|
|
ax, |
|
1069
|
|
|
scatter_kwargs) |
|
1070
|
|
|
|
|
1071
|
|
|
def calc_direct_bias(self, neutral_words='professions', c=None): |
|
1072
|
|
|
if isinstance(neutral_words, str) and neutral_words == 'professions': |
|
1073
|
|
|
return super().calc_direct_bias( |
|
1074
|
|
|
self._data['neutral_profession_names'], c) |
|
1075
|
|
|
else: |
|
1076
|
|
|
return super().calc_direct_bias(neutral_words) |
|
1077
|
|
|
|
|
1078
|
|
|
def generate_closest_words_indirect_bias(self, |
|
1079
|
|
|
neutral_positive_end, |
|
1080
|
|
|
neutral_negative_end, |
|
1081
|
|
|
words='professions', n_extreme=5): |
|
1082
|
|
|
# pylint: disable=C0301 |
|
1083
|
|
|
|
|
1084
|
|
|
if words == 'professions': |
|
1085
|
|
|
words = self._data['profession_names'] |
|
1086
|
|
|
|
|
1087
|
|
|
return super().generate_closest_words_indirect_bias(neutral_positive_end, |
|
1088
|
|
|
neutral_negative_end, |
|
1089
|
|
|
words, |
|
1090
|
|
|
n_extreme=n_extreme) |
|
1091
|
|
|
|
|
1092
|
|
|
def debias(self, method='hard', neutral_words=None, equality_sets=None, |
|
1093
|
|
|
inplace=True): |
|
1094
|
|
|
# pylint: disable=C0301 |
|
1095
|
|
|
if method in ['hard', 'neutralize']: |
|
1096
|
|
|
if neutral_words is None: |
|
1097
|
|
|
neutral_words = self._data['neutral_words'] |
|
1098
|
|
|
|
|
1099
|
|
|
if method == 'hard' and equality_sets is None: |
|
1100
|
|
|
equality_sets = self._data['definitional_pairs'] |
|
1101
|
|
|
|
|
1102
|
|
|
if not self.only_lower: |
|
1103
|
|
|
assert all(len(equality_set) == 2 |
|
1104
|
|
|
for equality_set in equality_sets), 'currently supporting only equality pairs if only_lower is False' |
|
1105
|
|
|
# TODO: refactor |
|
1106
|
|
|
equality_sets = {(candidate1, candidate2) |
|
1107
|
|
|
for word1, word2 in equality_sets |
|
1108
|
|
|
for candidate1, candidate2 in zip(generate_one_word_forms(word1), |
|
1109
|
|
|
generate_one_word_forms(word2))} |
|
1110
|
|
|
|
|
1111
|
|
|
return super().debias(method, neutral_words, equality_sets, |
|
1112
|
|
|
inplace) |
|
1113
|
|
|
|
|
1114
|
|
|
def learn_full_specific_words(self, seed_specific_words='bolukbasi', |
|
1115
|
|
|
max_non_specific_examples=None, |
|
1116
|
|
|
debug=None): |
|
1117
|
|
|
if seed_specific_words == 'bolukbasi': |
|
1118
|
|
|
seed_specific_words = self._data['specific_seed'] |
|
1119
|
|
|
|
|
1120
|
|
|
return super().learn_full_specific_words(seed_specific_words, |
|
1121
|
|
|
max_non_specific_examples, |
|
1122
|
|
|
debug) |
|
1123
|
|
|
|
ResponsiblyAI /
responsibly
Loading history...