|
1
|
|
|
# TODO how import files from a package |
|
2
|
|
|
import json |
|
3
|
|
|
import warnings |
|
4
|
|
|
|
|
5
|
|
|
from gensim.models.keyedvectors import KeyedVectors |
|
6
|
|
|
from pkg_resources import resource_filename, resource_string |
|
7
|
|
|
|
|
8
|
|
|
|
|
9
|
|
|
def load_w2v_small(): |
|
10
|
|
|
"""Load reduced Word2Vec model as `KeyedVectors` object. |
|
11
|
|
|
|
|
12
|
|
|
Based on the pre-trained embedding on the Google News corpus: |
|
13
|
|
|
https://code.google.com/archive/p/word2vec/ |
|
14
|
|
|
""" |
|
15
|
|
|
# pylint: disable=C0301 |
|
16
|
|
|
|
|
17
|
|
|
with warnings.catch_warnings(): |
|
18
|
|
|
warnings.simplefilter('ignore', DeprecationWarning) |
|
19
|
|
|
model = KeyedVectors.load_word2vec_format( |
|
20
|
|
|
resource_filename(__name__, 'GoogleNews-vectors-negative300-bolukbasi.bin'), |
|
21
|
|
|
binary=True) |
|
22
|
|
|
|
|
23
|
|
|
return model |
|
24
|
|
|
|
|
25
|
|
|
|
|
26
|
|
|
def load_json_resource(resource_name): |
|
27
|
|
|
return json.loads( |
|
28
|
|
|
resource_string(__name__, resource_name + '.json').decode('utf-8') |
|
29
|
|
|
|
|
30
|
|
|
) |
|
31
|
|
|
|
|
32
|
|
|
|
|
33
|
|
|
BOLUKBASI_DATA = load_json_resource('bolukbasi') |
|
34
|
|
|
|
|
35
|
|
|
BOLUKBASI_DATA['gender']['profession_names'] = list( |
|
36
|
|
|
zip(*BOLUKBASI_DATA['gender']['professions']))[0] |
|
37
|
|
|
|
|
38
|
|
|
|
|
39
|
|
|
BOLUKBASI_DATA['gender']['specific_full'].sort() |
|
40
|
|
|
|
|
41
|
|
|
# TODO: in the code of the article, the last definitional pair |
|
42
|
|
|
# is not in the specific full |
|
43
|
|
|
BOLUKBASI_DATA['gender']['specific_full_with_definitional_equalize'] = list( |
|
44
|
|
|
(set.union( |
|
45
|
|
|
*map(set, BOLUKBASI_DATA['gender']['definitional_pairs'])) |
|
46
|
|
|
| set.union( |
|
47
|
|
|
*map(set, BOLUKBASI_DATA['gender']['equalize_pairs'])) |
|
48
|
|
|
| set(BOLUKBASI_DATA['gender']['specific_full'])) |
|
49
|
|
|
) |
|
50
|
|
|
BOLUKBASI_DATA['gender']['specific_full_with_definitional_equalize'].sort() |
|
51
|
|
|
|
|
52
|
|
|
BOLUKBASI_DATA['gender']['neutral_profession_names'] = list( |
|
53
|
|
|
set(BOLUKBASI_DATA['gender']['profession_names']) |
|
54
|
|
|
- set(BOLUKBASI_DATA['gender']['specific_full_with_definitional_equalize']) |
|
55
|
|
|
) |
|
56
|
|
|
BOLUKBASI_DATA['gender']['neutral_profession_names'].sort() |
|
57
|
|
|
|
|
58
|
|
|
BOLUKBASI_DATA['gender']['word_group_keys'] = ['profession_names', |
|
59
|
|
|
'neutral_profession_names', |
|
60
|
|
|
'specific_seed', |
|
61
|
|
|
'specific_full', |
|
62
|
|
|
'specific_full_with_definitional_equalize'] # pylint: disable=C0301 |
|
63
|
|
|
|
|
64
|
|
|
|
|
65
|
|
|
WEAT_DATA = load_json_resource('weat') |
|
66
|
|
|
|
|
67
|
|
|
# Zhao, J., Wang, T., Yatskar, M., Ordonez, V., & Chang, K. W. (2018). |
|
68
|
|
|
# Gender bias in coreference resolution: Evaluation and debiasing methods. |
|
69
|
|
|
# arXiv preprint arXiv:1804.06876. |
|
70
|
|
|
# https://arxiv.org/abs/1804.06876 |
|
71
|
|
|
OCCUPATION_FEMALE_PRECENTAGE = load_json_resource( |
|
72
|
|
|
'occupational_female_precentage') |
|
73
|
|
|
|
ResponsiblyAI /
responsibly
