Bias Mitigation

WEFE also provides several methods to mitigate the bias of the embedding models. In the following section:

  • We present how to reduce binary bias (such as gender bias) using Hard Debias.

  • We present how to reduce multiclass bias (such as ethnic having classes like black, white, Latino, etc…) using Multiclass Hard Debias.

import gensim.downloader as api

from wefe.datasets import fetch_debiaswe, load_weat
from wefe.debias.hard_debias import HardDebias
from wefe.metrics import WEAT
from wefe.query import Query
from wefe.word_embedding_model import WordEmbeddingModel

twitter_25 = api.load("glove-twitter-25")
model = WordEmbeddingModel(twitter_25, "glove-twitter-dim=25")

Hard Debias

This method allow reducing the bias of an embedding model through geometric operations between embeddings. This method is binary because it only allows two classes of the same bias criterion, such as male or female.

The main idea of this method is:

  1. Identify a bias subspace through the defining sets. In the case

    of gender, these could be e.g., {'woman', 'man'}, {'she', 'he'}, ...

  2. Neutralize the bias subspace on the embeddings that should not be biased.

    First, it is defined a set of words that are correct to be related to the bias criterion: the criterion specific gender words. For example, in the case of gender, gender specific words are: {'he', 'his', 'He', 'her', 'she', 'him', 'him', 'She', 'man', 'women', 'men'...}.

    Then, it is defined that all words outside this set should have no relation to the bias criterion and thus have the possibility of being biased. (e.g. for the case of gender: {doctor, nurse, ...}). Therefore, this set of words is neutralized with respect to the bias subspace found in the previous step.

    The neutralization is carried out under the following operation:

    • u : embedding

    • v : bias direction

    First calculate the projection of the embedding on the bias subspace.

    • projection = v • (v • u) / (v • v)

    Then subtract the projection from the embedding.

    • u’ = u - projection

  3. Equalize the embeddings with respect to the bias direction..

    Given an equalization set (set of word pairs such as [she, he], [men, women], …, but not limited to the definitional set) this step executes, for each pair, an equalization with respect to the bias direction. That is, it takes a pair of embeddings and distributes them both at the same distance from the bias direction, so that neither is closer to the bias direction than the other.

Fit-Transform Interface

WEFE implements all debias methods through an interface inspired by the transformers of scikit-learn. That is, the execution of a debias method involves two steps: - First a training through the fit method where the transformation that will be applied on the embeddings is calculated - Second, a transform that applies the trained transformation.

Each of these stages defines its own parameters.

The fit parameters define how the neutralization will be calculated. In Hard Debias, you have to provide the the definitional_pairs, the equalize_pairs (which could be the same of definitional pairs) and optionally, a debias criterion_name (to name the debiased model).

debiaswe_wordsets = fetch_debiaswe()

definitional_pairs = debiaswe_wordsets["definitional_pairs"]
equalize_pairs = debiaswe_wordsets["equalize_pairs"]
gender_specific = debiaswe_wordsets["gender_specific"]

hd = HardDebias(verbose=False, criterion_name="gender").fit(
    model,
    definitional_pairs=definitional_pairs,
    equalize_pairs=equalize_pairs,
)

The parameters of the transform method are relatively standard for all methods. The most important ones are target, ignore and copy.

In the following example we use ignore and copy, which are described below:

  • ignore (by default, None):

    A list of strings that indicates that the debias method will perform the debias in all words except those specified in this list. In case it is not specified, debias will be executed on all words. In case ignore is not specified or its value is None, the transformation will be performed on all embeddings. This may cause words that are specific to social groups to lose that component (for example, leaving 'she' and 'he' without a gender component).

  • copy (by default True):

    if the value of copy is True, method attempts to create a copy of the model and run debias on the copy. If False, the method is applied on the original model, causing the vectors to mutate.

    WARNING: Setting copy with True requires at least 2x RAM of the size of the model. Otherwise the execution of the debias may rise MemoryError.

Next, the transformation is executed using a copy of the model, ignoring the words contained in gender_specific.

gender_debiased_model = hd.transform(model, ignore=gender_specific, copy=True)

Copy argument is True. Transform will attempt to create a copy of the original model. This may fail due to lack of memory.
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors
Model copy created successfully.
100%|██████████| 1193514/1193514 [00:18<00:00, 65143.18it/s]
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors

Using the metrics displayed in the first section of this user guide, we can measure whether or not there was a change in the measured bias between the original model and the debiased model.

weat_wordset = load_weat()
weat = WEAT()

gender_query_1 = Query(
    [word_sets["male_terms"], word_sets["female_terms"]],
    [word_sets["career"], word_sets["family"]],
    ["Male terms", "Female terms"],
    ["Career", "Family"],
)

gender_query_2 = Query(
    [weat_wordset["male_names"], weat_wordset["female_names"]],
    [weat_wordset["pleasant_5"], weat_wordset["unpleasant_5"]],
    ["Male Names", "Female Names"],
    ["Pleasant", "Unpleasant"],
)

biased_results_1 = weat.run_query(gender_query_1, model, normalize=True)
debiased_results_1 = weat.run_query(gender_query, gender_debiased_model, normalize=True)

print(round(debiased_results_1["weat"], 3),"<",round(biased_results_1["weat"], 3),
    "=",debiased_results_1["weat"] < biased_results_1["weat"],)

-0.06 < 0.317 = True

biased_results_2 = weat.run_query(
    gender_query_2, model, normalize=True, preprocessors=[{}, {"lowercase": True}]
)
debiased_results_2 = weat.run_query(
    gender_query_2,
    gender_debiased_model,
    normalize=True,
    preprocessors=[{}, {"lowercase": True}],
)

print(
    round(debiased_results_2["weat"], 3),"<",round(biased_results_2["weat"], 3),
    "=",debiased_results_2["weat"] < biased_results_2["weat"],)

-1.033 < -0.949 = True

Target Parameter

  • target: If a set of words is specified in target, the debias method will be performed only on the word embeddings associated with this set. In the case of providing None, the transformation will be performed on all vocabulary words except those specified in ignore. By default None.

    In the following example, the target parameter is used to execute the transformation only on the career and family word set:

targets = ['executive',
           'management',
           'professional',
           'corporation',
           'salary',
           'office',
           'business',
           'career',
           'home',
           'parents',
           'children',
           'family',
           'cousins',
           'marriage',
           'wedding',
           'relatives']

hd = HardDebias(verbose=False, criterion_name="gender").fit(
    model,
    definitional_pairs=definitional_pairs,
    equalize_pairs=equalize_pairs,
)

gender_debiased_model = hd.transform(
    model, target=targets, copy=True
)

Copy argument is True. Transform will attempt to create a copy of the original model. This may fail due to lack of memory.
Model copy created successfully.
100%|██████████| 16/16 [00:00<00:00, 10754.63it/s]
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors

Next, a bias test is run on the mitigated embeddings associated with the target words. In this case, the value of the metric is lower on the query executed on the mitigated model than on the original one. These results indicate that there was a mitigation of bias on embeddings of these words.

biased_results_1 = weat.run_query(gender_query_1, model, normalize=True)
debiased_results_1 = weat.run_query(gender_query, gender_debiased_model, normalize=True)

print(round(debiased_results_1["weat"], 3),"<",round(biased_results_1["weat"], 3)
      ,"=",debiased_results_1["weat"] < biased_results_1["weat"],)

-0.06 < 0.317 = True

However, if a bias test is run with words that were outside the target word set, the results are almost the same. The slight difference in the metric scores lies in the fact that the equalize sets were still equalized. Equalization can be deactivated by delivering an empty equalize set ([])

biased_results_2 = weat.run_query(
    gender_query_2, model, normalize=True, preprocessors=[{}, {"lowercase": True}]
)
debiased_results_2 = weat.run_query(
    gender_query_2,
    gender_debiased_model,
    normalize=True,
    preprocessors=[{}, {"lowercase": True}],
)

print(round(debiased_results_2["weat"], 3),"<",round(biased_results_2["weat"], 3),
    "=",debiased_results_2["weat"] < biased_results_2["weat"],)

-0.941 < -0.949 = False

Save the Debiased Model

To save the mitigated model one must access the KeyedVectors (the gensim object that contains the embeddings) through wv and then use the save method to store the method in a file.

gender_debiased_model.wv.save('gender_debiased_glove.kv')

INFO:gensim.utils:saving Word2VecKeyedVectors object under gender_debiased_glove.kv, separately None
INFO:gensim.utils:storing np array 'vectors' to gender_debiased_glove.kv.vectors.npy
INFO:gensim.utils:not storing attribute vectors_norm
DEBUG:smart_open.smart_open_lib:{'uri': 'gender_debiased_glove.kv', 'mode': 'wb', 'buffering': -1, 'encoding': None, 'errors': None, 'newline': None, 'closefd': True, 'opener': None, 'ignore_ext': False, 'compression': None, 'transport_params': None}
INFO:gensim.utils:saved gender_debiased_glove.kv

Multiclass Hard Debias

Multiclass Hard Debias is a generalized version of Hard Debias that enables multiclass debiasing. Generalized refers to the fact that this method extends Hard Debias in order to support more than two types of social target sets within the definitional set.

For example, for the case of religion bias, it supports a debias using words associated with Christianity, Islam and Judaism.

The usage is very similar to Hard Debias with the difference that the definitional_sets can be larger than pairs.

from wefe.datasets import fetch_debias_multiclass
from wefe.debias.multiclass_hard_debias import MulticlassHardDebias

multiclass_debias_wordsets = fetch_debias_multiclass()
weat_wordsets = load_weat()
weat = WEAT()

ethnicity_definitional_sets = multiclass_debias_wordsets["ethnicity_definitional_sets"]
ethnicity_equalize_sets = list(
    multiclass_debias_wordsets["ethnicity_analogy_templates"].values()
)

mhd = MulticlassHardDebias(verbose=True, criterion_name="ethnicity")
mhd.fit(
    model=model,
    definitional_sets=ethnicity_definitional_sets,
    equalize_sets=ethnicity_equalize_sets,
)

ethnicity_debiased_model = mhd.transform(model, copy=True)

INFO:wefe.debias.multiclass_hard_debias:PCA variance explaned: [4.0089381e-01 2.3377793e-01 1.7155512e-01 7.3547199e-02 5.5353384e-02
3.5681739e-02 2.2261711e-02 6.9290772e-03 2.4344339e-15 2.4052477e-15]
Obtaining definitional sets.
Word(s) found: ['black', 'caucasian', 'asian'], not found: []
Word(s) found: ['african', 'caucasian', 'asian'], not found: []
Word(s) found: ['black', 'white', 'asian'], not found: []
Word(s) found: ['africa', 'america', 'asia'], not found: []
Word(s) found: ['africa', 'america', 'china'], not found: []
Word(s) found: ['africa', 'europe', 'asia'], not found: []
6/6 sets of words were correctly converted to sets of embeddings
Identifying the bias subspace.
Obtaining equalize pairs.
Word(s) found: ['manager', 'executive', 'redneck', 'hillbilly', 'leader', 'farmer'], not found: []
Word(s) found: ['doctor', 'engineer', 'laborer', 'teacher'], not found: []
Word(s) found: ['slave', 'musician', 'runner', 'criminal', 'homeless'], not found: []
3/3 sets of words were correctly converted to sets of embeddings
Executing Multiclass Hard Debias on glove-twitter-dim=25
copy argument is True. Transform will attempt to create a copy of the original model. This may fail due to lack of memory.


INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors

Model copy created successfully.
Normalizing embeddings.
Neutralizing embeddings

100%|██████████| 1193504/1193504 [01:38<00:00, 12108.73it/s]
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors
DEBUG:wefe.debias.multiclass_hard_debias:Equalizing embeddings..
INFO:gensim.models.keyedvectors:precomputing L2-norms of word weight vectors

Normalizing embeddings.
Normalizing embeddings.
Done!

# test with weat

ethnicity_query = Query(
    [
        multiclass_debias_wordsets["white_terms"],
        multiclass_debias_wordsets["black_terms"],
    ],
    [multiclass_debias_wordsets["white_biased_words"],
    multiclass_debias_wordsets["black_biased_words"]],
    ["european_american_names", "african_american_names"],
    ["white_biased_words", "black_biased_words"],
)

biased_results = weat.run_query(
    ethnicity_query, model, normalize=True, preprocessors=[{}, {"lowercase": True}],
)
debiased_results = weat.run_query(
    ethnicity_query,
    ethnicity_debiased_model,
    normalize=True,
    preprocessors=[{}, {"lowercase": True}],
)

Absolute value is used here because the closer the value is to zero, the less biased the model is.

import numpy as np

print(
    '| -',
    round(np.abs(debiased_results["weat"]), 3),
    "| < | -",
    round(np.abs(biased_results["weat"]), 3),
    "| =",
    np.abs(debiased_results["weat"]) < np.abs(biased_results["weat"]),
)

| - 0.027 | < | - 0.088 | = True