def project_profession_words(E, professions):
"""
Get gender axis and project profession words onto this axis.
:param object E: WordEmbedding object.
:param list professions: List of professions
:param list unique_words: List of words present in all
embeddings to compare.
:returns: projection, profession words, gender axis
"""
# Extract definitional word embeddings and determine gender direction.
defs = load_definitional_pairs(E.words)
v_gender = doPCA(defs, E).components_[0]
# Projection on the gender direction.
sp = E.profession_stereotypes(professions, v_gender, print_firstn=0)
occupations = [s[1] for s in sp]
return sp, occupations, v_gender
def main():
# Print basic experiment information
print_details()
# For each embedding, do the experiments
for embed in FLAGS.embeddings:
print("\n" + "#" * 56)
print("# " + f"Doing the {embed} embedding".center(53) + "#")
print("#" * 56)
# Load the embedding
E = WordEmbedding(embed)
# Load professions and gender related lists from
# Bolukbasi et al. for word2vec
gender_words, defs, equalize_pairs, profession_words = load_data(E.words)
# Define gender direction with PCA
v_gender = we.doPCA(defs, E).components_[0]
# Bias without debiasing
if not FLAGS.no_show:
show_bias(E, v_gender, profession_words, info="with bias")
# Hard debiasing
E_hard = hard(E, gender_words, defs, equalize_pairs)
if not FLAGS.no_show:
show_bias(E_hard, v_gender, profession_words, info="hard debiased")
E_soft = None
# Only do soft debiasing for small embeddings
if embed.split("_")[-1] != "large":
# Soft debiasing
E_soft = soft(E, embed, gender_words, defs)
if not FLAGS.no_show:
show_bias(E_soft, v_gender, profession_words, info="soft debiased")
# Run the benchmarks if nescessary
if not FLAGS.no_bench:
run_benchmark(E, E_hard, E_soft, embed)
def debias(E, gender_specific_words, definitional, equalize):
gender_direction = we.doPCA(definitional, E).components_[0]
specific_set = set(gender_specific_words)
for i, w in enumerate(E.words):
if w not in specific_set:
E.vecs[i] = we.drop(E.vecs[i], gender_direction)
E.normalize()
candidates = {
x
for e1, e2 in equalize
for x in [(e1.lower(),
e2.lower()), (e1.title(),
e2.title()), (e1.upper(), e2.upper())]
}
print(candidates)
for (a, b) in candidates:
if (a in E.index and b in E.index):
y = we.drop((E.v(a) + E.v(b)) / 2, gender_direction)
z = np.sqrt(1 - np.linalg.norm(y)**2)
if (E.v(a) - E.v(b)).dot(gender_direction) < 0:
z = -z
E.vecs[E.index[a]] = z * gender_direction + y
E.vecs[E.index[b]] = -z * gender_direction + y
E.normalize()