Debiasing word embeddings

Overview

A word embedding is a vector per word, learned from text. Earlier debiasing methods pick one direction in the space, the he-she direction, and shrink each word's projection on it. A word's bias is that projection, with all vectors normalised 1903.03862 §2. The invariant the methods aim for is a projection near zero for gender-neutral words.

The paper's argument is that this measure is too narrow. Stereotyped words keep their relations to each other after the projection is removed, so a downstream model can still learn gender from neighbourhoods [arxiv:1903.03862 §2] 1903.03862 §5. Its practical rule is to judge a debiased embedding by several tests, not by the projection alone 1903.03862 §4.

Variants

Only two debiasing methods are tested in the paper. Both are compared with their own biased starting point 1903.03862 §3.

Hard-Debiased (post-processing)

Bolukbasi et al.'s method, applied to a finished embedding. The paper compares it with the original word2vec embedding trained on Google News 1903.03862 §3. Its projection bias falls, but gender stays visible: the k-means accuracy on the 1,000 most biased words drops from 99.9% to 92.5%, and the SVM gender accuracy from 98.25% to 88.88% 1903.03862 §4 fig. 1. Card: Gonen and Goldberg, 2019.

GN-GloVe (during training)

Zhao et al.'s method, which debiases while the embedding is trained. The paper compares it with a standard GloVe model trained on the same corpus 1903.03862 §3. Gender also stays visible after this method: the k-means accuracy falls from 100% to 85.6%, and the SVM accuracy from 98.65% to 96.53% 1903.03862 §4 fig. 1. Card: Gonen and Goldberg, 2019.

Adversarial debiasing (not tested)

The paper mentions Zhang et al.'s adversarial method in a footnote, but does not test it 1903.03862 §2. Nothing here says how it behaves.

Checks after debiasing

The paper runs five checks on both debiased spaces. Each one can be run on any debiased embedding, given the original embedding as a reference 1903.03862 §4.

1. Clustering. Split the most biased words with k-means and score the split against the original gender labels.
2. Neighbour bias. Count each word's male-biased neighbours, using the original bias as the label, and correlate that count with the original bias.
3. Professions. Repeat check 2 on a list of occupations.
4. Association tests. Run three WEAT-style tests of names against attribute word sets, on the debiased space only 1903.03862 §4.
5. Classifier. Train an RBF-kernel SVM to predict gender from the vectors and report accuracy on held-out words 1903.03862 §4.

When to use which

The paper does not give a recommended method. It tests two and reports what each leaves behind. The rows below follow from the constraints a team has.

The card records no comparison of training cost or runtime between the two methods, so this page cannot use one to choose between the first two rows.

Implementation sketch

The shared core of the checks: a projection score per word, a neighbour count that uses the original labels, and an agreement score. Variants differ in how the debiased vectors are made, not in these checks.

# Projection bias on unit vectors, positive = male-leaning  [arxiv:1903.03862 §2]
def projection_bias(w, vectors, he, she):
    return dot(vectors[w], he) - dot(vectors[w], she)

# Male-biased neighbour count; labels come from the ORIGINAL space  [§4]
def male_neighbour_count(w, debiased, bias_before, k=100):
    # The paper says "k nearest neighbours" without a value; 100 is the
    # value used in the released code (see Pitfalls)
    count = 0
    for n in nearest_by_cosine(w, debiased, k):
        if bias_before[n] > 0:
            count += 1
    return count

# Agreement between original bias and neighbour count  [arxiv:1903.03862 §4]
def neighbour_agreement(words, bias_before, count):
    return pearson_r([bias_before[w] for w in words],
                     [count[w] for w in words])

# Held-out gender accuracy of an RBF-kernel SVM  [arxiv:1903.03862 §4]
def gender_probe_accuracy(train, test, vectors):
    # Default SVM settings; use one split for the original and debiased runs
    model = fit_svm_rbf(vectors, train)
    return mean(model.predict(test_vectors) == test_labels)

Contradictions

None known among the cards in this repository. The paper's argument runs against the projection-only measure used by the methods it tests, which treats a shrunk projection as removed bias 1903.03862 §2. Both sides are cited above.

Papers

• Gonen and Goldberg, 2019: tests Hard-Debiased and GN-GloVe with five checks and shows that gender stays recoverable after each.