`wefe.debias.repulsion_attraction_neutralization`.RepulsionAttractionNeutralization

class wefe.debias.repulsion_attraction_neutralization.RepulsionAttractionNeutralization(pca_args: Dict[str, Any] = {'n_components': 10}, verbose: bool = False, criterion_name: Optional[str] = None, epochs: int = 300, theta: float = 0.05, n_neighbours: int = 100, learning_rate: float = 0.01, weights: List[float] = [0.33, 0.33, 0.33])[source]

Repulsion Attraction Neutralization method.

Warning

This method only works if Pytorch is installed. If you do not have it installed, check https://pytorch.org/get-started/locally/ for further information.

This method allow reducing the bias of an embedding model creating a transformation such that the stereotypical information is minimized with minimal semantic offset. This transformation bases its operations on:

Repelling embeddings from neighbours with a high value of indirect bias (indicating a strong association due to bias), to minimize the bias based illicit associations.
Attracting debiased embeddings to the original representation, to minimize the loss of semantic meaning.
Neutralizing the bias direction of each word, minimizing its bias to any particular group.

This method is binary because it only allows 2 classes of the same bias criterion, such as male or female.

Note

For a multiclass debias (such as for Latinos, Asians and Whites), it is recommended to visit MulticlassHardDebias class.

The steps followed to perform the debias are:

Identify a bias subspace through the defining sets. In the case of gender, these could be e.g. [['woman', 'man'], ['she', 'he'], ...]
A multi-objective optimization is performed. For each vector \(w\) in the target set it is found its debias counterpart \(w_d\) by solving:

\[argmin(F_r(w_d),F_a(w_d),F_n(w_d))\]

where \(Fr\), \(Fa\), \(Fn\) are repulsion, attraction and neutralization functions defined as the following:

\[F_r(w_d) = \sum |cos(w_d,n_i)| / |S|\]

\[F_a(w_d) = |cos(w_d,w)-1|/2\]

\[F_n(w_d) = |cos(w_d,g)|\]

The optimization is performed by formulating a single objective:

\[F(w_d) = \lambda_1 F_r(w_d) + \lambda_2 F_a(w_d) + \lambda_3 F_n(w_d)\]

In the original implementation is define a preserve set \((V_p)\) corresponding to words for which gender carries semantic importance, this words are not included in the debias process.

In WEFE this words would be the ones included in the ignore parameter of the transform method. The words that are not present in \(V_p\) are the ones to be included in the debias process and form part of the debias set \((V_d)\), in WEFE this words can be specified in the target parameter of the transform method.

References

[1]: Kumar, Vaibhav, Tenzin Singhay Bhotia y Tanmoy Chakraborty: Nurse
is Closer to Woman than Surgeon? Mitigating Gender-Biased Proximities
in Word Embeddings. CoRR,abs/2006.01938, 2020.
https://arxiv.org/abs/2006.01938
[2]: https://github.com/TimeTraveller-San/RAN-Debias

Examples

The following example shows how to execute Repulsion Attraction Neutralization method that reduces bias in a word embedding model:

>>> from wefe.debias.repulsion_attraction_neutralization import (
...   RepulsionAttractionNeutralization
... )
>>> from wefe.utils import load_test_model
>>> from wefe.datasets import fetch_debiaswe
>>>
>>> # load the model (in this case, the test model included in wefe)
>>> model = load_test_model()
>>> # load definitional pairs, in this case definitinal pairs included in wefe
>>> debiaswe_wordsets = fetch_debiaswe()
>>> definitional_pairs = debiaswe_wordsets["definitional_pairs"]
>>>
>>> # instance and fit the method
>>> ran = RepulsionAttractionNeutralization().fit(
...     model = model,
...     definitional_pairs= definitional_pairs
...   )
>>> # execute the debias passing words over a set of target words
>>> debiased_model = ran.transform(
...    model = model, target = ['doctor','nurse','programmer']
... )
Copy argument is True. Transform will attempt to create a copyof the original model.
This may fail due to lack of memory.
Model copy created successfully.
>>> # if you don't want a set of words to be debiased include them in the ignore set
>>> gender_specific = debiaswe_wordsets["gender_specific"]
>>> debiased_model = ran.transform(
...    model = model, ignore= gender_specific
... ) 

__init__(pca_args: Dict[str, Any] = {'n_components': 10}, verbose: bool = False, criterion_name: Optional[str] = None, epochs: int = 300, theta: float = 0.05, n_neighbours: int = 100, learning_rate: float = 0.01, weights: List[float] = [0.33, 0.33, 0.33]) → None[source]

Initialize a Repulsion Attraction Neutralization Debias instance.

Parameters:

pca_argsDict[str, Any], optional: Arguments for the PCA that is calculated internally in the identification of the bias subspace, by default {“n_components”: 10}
verbosebool, optional: True will print informative messages about the debiasing process, by default False.
criterion_nameOptional[str], optional: The name of the criterion for which the debias is being executed, e.g., ‘Gender’. This will indicate the name of the model returning transform, by default None
epochsint, optional: number of times that the minimization is done. By default 300
theta: float, optional: Indirect bias threshold to select neighbours for the repulsion set. By default 0.05
n_neighbours: int, optional: Number of neighbours to be consider for the repulsion set. By default 100
learning_rate: float, optional: Learning rate to be used by the optimizer during the optimization. By default 0.01
weights: List[float], optional: List of the 3 initial weights to be used. By default [0.33,0.33,0.33]

fit(model: WordEmbeddingModel, definitional_pairs: Sequence[Sequence[str]]) → BaseDebias[source]

Compute the bias direction.

Parameters:

modelWordEmbeddingModel: The word embedding model to debias.
definitional_pairsSequence[Sequence[str]]: A sequence of string pairs that will be used to define the bias direction. For example, for the case of gender debias, this list could be [[‘woman’, ‘man’], [‘girl’, ‘boy’], [‘she’, ‘he’], [‘mother’, ‘father’], …].

Returns:

BaseDebias: The debias method fitted.

transform(model: WordEmbeddingModel, target: Optional[List[str]] = None, ignore: Optional[List[str]] = [], copy: bool = True) → WordEmbeddingModel[source]

Execute Repulsion Attraction Neutralization Debias over the provided model.

Parameters:

modelWordEmbeddingModel: The word embedding model to debias.
targetOptional[List[str]], optional: If a set of words is specified in target, the debias method will be performed only on the word embeddings of this set. If None is provided, the debias will be performed on all words (except those specified in ignore). by default None.
ignoreOptional[List[str]], optional: If target is None and a set of words is specified in ignore, the debias method will perform the debias in all words except those specified in this set, by default None.
copybool, optional: If True, the debias will be performed on a copy of the model. If False, the debias will be applied on the same model delivered, causing its vectors to mutate. WARNING: Setting copy with True requires RAM at least 2x of the size of the model, otherwise the execution of the debias may raise to MemoryError, by default True.
WordEmbeddingModel: The debiased embedding model.

wefe.debias.repulsion_attraction_neutralization.RepulsionAttractionNeutralization

`wefe.debias.repulsion_attraction_neutralization`.RepulsionAttractionNeutralization