learn.ifs.name

Embeddings

Jurafsky & Martin, Speech and Language Processing, 3rd ed. draft, ch. 5

6 cards in “5.8 Bias and embeddings” · whole deck

5.8 Bias and embeddings

  1. How do embedding analogies exhibit gender stereotypes?
    In word2vec trained on news, computer programmer − man + woman is closest to homemaker, and father : doctor :: mother : nurse.
  2. What is an allocational harm?
    When a system allocates resources (jobs, credit) unfairly to different groups, e.g. a hiring search using biased embeddings that downweights documents with women’s names.
  3. What is bias amplification in embeddings?
    Embeddings don’t just reflect input statistics but exaggerate them: gendered terms become more gendered in embedding space than in the text, and biases exceed actual employment statistics.
  4. What is the Implicit Association Test, and what did Caliskan et al. (2017) show with embeddings?
    The IAT measures people’s associations between concepts and attributes by response latencies. Caliskan et al. replicated its findings using GloVe cosines instead of latencies, e.g. African American names closer to unpleasant words, European American names to pleasant ones.
  5. What is the current state of debiasing embeddings?
    Methods transform the space to remove stereotypes while keeping definitional gender, or change the training procedure. They reduce but do not eliminate bias; it remains an open problem.
  6. How have historical embeddings been used to study past biases?
    Cosines between occupation words and gendered or ethnic names in each decade correlate with historical occupation percentages, replicate 1930s stereotype surveys, and show competence adjectives being closer to male words, a bias slowly decreasing since 1960.

Source