Abstract
Word embeddings provide quantitative representations of
word semantics and the associations between word meanings
in text data, including in large repositories in media
and social media archives. This article introduces social psychologists
to word embedding research via a consideration
of bias analysis, a topic of central concern in the discipline.
We explain how word embeddings are constructed and how
they can be used to measure bias along bipolar dimensions
that are comparable to semantic differential scales. We review
recent studies that show how familiar social biases can
be detected in embeddings and how these change over time
and in conjunction with real-world
discriminatory practices.
The evidence suggests that embeddings yield valid and
reliable estimates of bias and that they can identify subtle
biases that may not be communicated explicitly. We argue
that word embedding research can extend scholarship on
prejudice and stereotyping, providing measures of the bias
environment of human thought and action.