Word vector bias is systematic distortion in learned word associations caused by stereotypes and imbalances in the training corpus. Embeddings reflect the average meanings and associations present in the documents used to train them, rather than neutral definitions. section 6.4.5
The same word can acquire different associations from different corpora. A finance corpus may associate “bank” with deposits, while a geology corpus may associate it with rivers; a corpus shaped by gender stereotypes can encode those stereotypes as geometric relationships. section 6.4.5
Bias becomes visible when comparing distances or similarities between words:
- A Google News word model places “woman” closer to “nurse” than “man,” with distances of 0.5586 and 0.7453 respectively. section 6.4.5
- Such measurements expose associations learned from a biased world; they do not establish that the associations are fair, accurate, or desirable. section 6.4.5
Inspecting a gender association
Compare word-model distances directly: word_model.distance('man', 'nurse') returns 0.7453, while word_model.distance('woman', 'nurse') returns 0.5586. The unequal distances provide a concrete diagnostic for gender bias. section 6.4.5
>>> word_model.distance('man', 'nurse')
0.7453
>>> word_model.distance('woman', 'nurse')
0.5586
Inspect the runnable distance comparison between “man,” “woman,” and “nurse.” section 6.4.5
