"""Repulsion Attraction Neutralization WEFE implementation."""
from copy import deepcopy
from typing import Any, Dict, List, Optional, Sequence
import numpy as np
from sklearn.decomposition import PCA
from tqdm import tqdm
from wefe.debias.base_debias import BaseDebias
from wefe.preprocessing import EmbeddingDict, get_embeddings_from_tuples
from wefe.utils import check_is_fitted
from wefe.word_embedding_model import WordEmbeddingModel
try:
import torch
from torch import nn
except ModuleNotFoundError:
raise ModuleNotFoundError(
"PyTorch is required to run RepulsionAttractionNeutralization method. "
"Visit https://pytorch.org/ to install it."
)
class RAN_Loss(nn.Module):
"""Loss for Repulsion Attraction Neutralization method."""
def __init__(
self,
w_b: np.ndarray,
w: np.ndarray,
repulsion_set: List[np.ndarray],
bias_direction: np.ndarray,
) -> None:
"""Initialize a RAN_Loss instance.
Parameters
----------
w_b: np.array
Debiased embedding of word
w: np.array
Original embedding of word
repulsion_set: List[np.ndarray]
Set of embeddings to be repeled from word
bias_direction: np.array
"""
super(RAN_Loss, self).__init__()
self.w = w
self.w_b = nn.Parameter(w_b)
self.repulsion_set = repulsion_set
self.bias_direction = bias_direction
def _repulsion(self) -> torch.Tensor:
if not isinstance(self.repulsion_set, bool):
cos_similarity = self._cosine_similarity(self.w_b, self.repulsion_set)
repulsion = torch.abs(cos_similarity).mean(dim=0)
else:
repulsion = 0
return repulsion
def _attraction(self) -> torch.Tensor:
attraction = (
torch.abs(torch.cosine_similarity(self.w_b[None, :], self.w[None, :]) - 1)[
0
]
/ 2
)
return attraction
def _neutralization(self) -> torch.Tensor:
neutralization = torch.abs(self.w_b.dot(self.bias_direction)).mean(dim=0)
return neutralization
def _objective_function(
self,
w1: np.number,
w2: np.number,
w3: np.number,
) -> torch.Tensor:
return (
self._repulsion() * w1
+ self._attraction() * w2
+ self._neutralization() * w3
)
def _cosine_similarity(
self, w: torch.Tensor, set_vectors: torch.Tensor
) -> torch.Tensor:
return torch.matmul(set_vectors, w) / (set_vectors.norm(dim=1) * w.norm(dim=0))
class RAN(nn.Module):
"""NN that performs the optimization by gradient descent of the loss function."""
def __init__(
self,
model: WordEmbeddingModel,
word: str,
w_b: np.ndarray,
w: np.ndarray,
repulsion_set: List[np.ndarray],
bias_direction: np.ndarray,
weights: Sequence = [0.33, 0.33, 0.33],
) -> None:
"""Initialize a RAN instance.
Parameters
----------
model: WordEmbeddingModel
The word embedding model to debias.
word: str
Word to be debiased
w_b: np.array
Debiased embedding of word
w: np.array
Original embedding of word
repulsion_set: List[np.ndarray]
Set of embeddings to be repeled from word
bias_direction: np.array
weights: Sequence, optional
weights λi that determine the relative importance of one
objective function (repulsion, attraction, neutralization)
over another. by Defaults [0.33, 0.33, 0.33].
"""
super(RAN, self).__init__()
self.model = model
self.word = word
self.w = torch.FloatTensor(np.array(w)).requires_grad_(True)
if len(repulsion_set) == 0:
self.repulsion_set = False
else:
self.repulsion_set = torch.FloatTensor(
np.array(repulsion_set)
).requires_grad_(True)
self.w_b = nn.Parameter(w_b)
self.bias_direction = torch.FloatTensor(np.array(bias_direction))
self.weights = weights
self.loss = RAN_Loss(self.w, self.w_b, self.repulsion_set, self.bias_direction)
def _forward(self) -> torch.Tensor:
w1, w2, w3 = self.weights
return self.loss._objective_function(w1, w2, w3)
[docs]class RepulsionAttractionNeutralization(BaseDebias):
r"""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:
1. 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.
2. Attracting debiased embeddings to the original representation, to minimize the
loss of semantic meaning.
3. 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
:class:`~wefe.debias.multiclass_hard_debias.MulticlassHardDebias` class.
The steps followed to perform the debias are:
1. Identify a bias subspace through the defining sets. In the case of
gender, these could be e.g. ``[['woman', 'man'], ['she', 'he'], ...]``
2. A multi-objective optimization is performed. For each vector :math:`w` in the
target set it is found its debias counterpart :math:`w_d` by solving:
.. math::
argmin(F_r(w_d),F_a(w_d),F_n(w_d))
where :math:`Fr`, :math:`Fa`, :math:`Fn` are repulsion, attraction and
neutralization functions defined as the following:
.. math::
F_r(w_d) = \sum |cos(w_d,n_i)| / |S|
.. math::
F_a(w_d) = |cos(w_d,w)-1|/2
.. math::
F_n(w_d) = |cos(w_d,g)|
The optimization is performed by formulating a single objective:
.. math::
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 :math:`(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 :math:`V_p` are the ones to be
included in the debias process and form part of the debias set :math:`(V_d)`,
in WEFE this words can be specified in the target parameter of the transform method.
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
... ) # doctest: +SKIP
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
"""
name = "Repulsion attraction Neutralization"
short_name = "RAN"
[docs] def __init__(
self,
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:
"""Initialize a Repulsion Attraction Neutralization Debias instance.
Parameters
----------
pca_args : Dict[str, Any], optional
Arguments for the PCA that is calculated internally in the
identification of the bias subspace,
by default {"n_components": 10}
verbose : bool, optional
True will print informative messages about the debiasing process,
by default False.
criterion_name : Optional[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
epochs : int, 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]
"""
# check verbose
if not isinstance(verbose, bool):
raise TypeError(f"verbose should be a bool, got {verbose}.")
self.pca_args = pca_args
self.verbose = verbose
if criterion_name is None or isinstance(criterion_name, str):
self.criterion_name_ = criterion_name
else:
raise ValueError(f"criterion_name should be str, got: {criterion_name}")
if not isinstance(epochs, int):
raise TypeError(f"epochs should be a int, got {epochs}.")
self.epochs = epochs
if not isinstance(theta, float):
raise TypeError(f"theta should be a float, got {theta}.")
self.theta = theta
if not isinstance(n_neighbours, int):
raise TypeError(f"n_neighbours should be a int, got {n_neighbours}.")
self.n_neighbours = n_neighbours
if not isinstance(weights, List):
raise TypeError(f"weights should be a list, got {weights}.")
self.weights = weights
if not isinstance(learning_rate, float):
raise TypeError(f"learning_rate should be a float, got {learning_rate}.")
self.learning_rate = learning_rate
def _identify_bias_subspace(
self,
defining_pairs_embeddings: List[EmbeddingDict],
verbose: bool = False,
) -> PCA:
matrix = []
for embedding_dict_pair in defining_pairs_embeddings:
# Get the center of the current defining pair.
pair_embeddings = np.array(list(embedding_dict_pair.values()))
center = np.mean(pair_embeddings, axis=0)
# For each word, embedding in the defining pair:
for embedding in embedding_dict_pair.values():
# Subtract the center of the pair to the embedding
matrix.append(embedding - center)
matrix = np.array(matrix) # type: ignore
pca = PCA(**self.pca_args)
pca.fit(matrix)
if verbose:
explained_variance = pca.explained_variance_ratio_
print(f"PCA variance explained: {explained_variance[0:pca.n_components_]}")
return pca
def _check_sets_size(
self,
sets: Sequence[Sequence[str]],
set_name: str,
) -> None:
for idx, set_ in enumerate(sets):
if len(set_) != 2:
adverb = "less" if len(set_) < 2 else "more"
raise ValueError(
f"The {set_name} pair at position {idx} ({set_}) has {adverb} "
f"words than allowed by {self.name}: "
f"got {len(set_)} words, expected 2."
)
def _indirect_bias(
self, w: np.ndarray, v: np.ndarray, bias_direction: np.ndarray
) -> float:
wv = np.dot(w, v)
w_orth = w - np.dot(w, bias_direction) * bias_direction
v_orth = v - np.dot(v, bias_direction) * bias_direction
cos_wv_orth = np.dot(w_orth, v_orth) / (
np.linalg.norm(w_orth) * np.linalg.norm(v_orth)
)
bias = (wv - cos_wv_orth) / wv
return bias
def _get_neighbours(
self, model: WordEmbeddingModel, word: str, n_neighbours: int
) -> List[str]:
similar_words = model.wv.most_similar(positive=word, topn=n_neighbours)
similar_words = list(list(zip(*similar_words))[0])
return similar_words
def _get_repulsion_set(
self,
model: WordEmbeddingModel,
word: str,
bias_direction: np.ndarray,
theta: float,
n_neighbours: int,
) -> List[np.ndarray]:
r"""Obtain the embeddings of the words that should be repealed from word arg.
These words are the n_neighbours more similar to "word" whose indirect bias is
greater than theta.
Parameters
----------
model: WordEmbeddingModel
The word embedding model to debias.
word: str
Word to debias.
bias_direction: np.ndarray
The bias subspace
theta: float
Threshold to include neighbours in repulsion set.
n_neighbours: int
Number of neighbours to be considered.
Returns
-------
List[np.ndarray]
The list of embeddings that compose the repulsion set.
"""
neighbours = self._get_neighbours(model, word, n_neighbours)
repulsion_set = []
for neighbour in neighbours:
if (
self._indirect_bias(model[neighbour], model[word], bias_direction)
> theta
):
repulsion_set.append(model[neighbour])
return repulsion_set
def _debias(
self,
model: WordEmbeddingModel,
word: str,
w: np.ndarray,
w_b: np.ndarray,
bias_direction: np.ndarray,
repulsion_set: List[np.ndarray],
learning_rate: float,
epochs: int,
weights: List[float],
) -> torch.Tensor:
ran = RAN(
model,
word,
w_b,
w,
repulsion_set,
bias_direction,
weights,
)
optimizer = torch.optim.Adam(ran.parameters(), lr=learning_rate)
for _ in range(epochs):
optimizer.zero_grad()
out = ran._forward()
out.backward()
optimizer.step()
debiased_vector = ran.w_b
return debiased_vector / torch.norm(debiased_vector)
def _init_vector(self, model: WordEmbeddingModel, word: str) -> torch.Tensor:
v = deepcopy(model[word])
return torch.FloatTensor(np.array(v))
[docs] def fit(
self,
model: WordEmbeddingModel,
definitional_pairs: Sequence[Sequence[str]],
) -> BaseDebias:
"""Compute the bias direction.
Parameters
----------
model : WordEmbeddingModel
The word embedding model to debias.
definitional_pairs : Sequence[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.
"""
# Check arguments types
self._check_sets_size(definitional_pairs, "definitional")
self.definitional_pairs_ = definitional_pairs
# ------------------------------------------------------------------------------
# Obtain the embedding of each definitional pairs.
if self.verbose:
print("Obtaining definitional pairs.")
self.definitional_pairs_embeddings_ = get_embeddings_from_tuples(
model=model,
sets=definitional_pairs,
sets_name="definitional",
warn_lost_sets=self.verbose,
normalize=True,
verbose=self.verbose,
)
# ------------------------------------------------------------------------------:
# Identify the bias subspace using the defining pairs.
if self.verbose:
print("Identifying the bias subspace.")
self.pca_ = self._identify_bias_subspace(
self.definitional_pairs_embeddings_,
self.verbose,
)
self.bias_direction_ = self.pca_.components_[0]
return self