De-Biasing Word Embeddings
Last updated
Was this helpful?
Last updated
Was this helpful?
Was this helpful?
Word embeddings can reflect the gender, ethnicity, age, sexual orientation and other biases of the text used to train the model.
For example, Man->Doctor, Woman->Nurse, which is not a correct mapping, and is biased based on gender.
The following is an overview of the steps to address this bias:
Identify the bias direction
Neutralize i.e. for every word that is not definitional (i.e. isn't defined to satisfy a given bias, for example, 'father' is defitional as it is only defined for the male gender, but 'doctor' is not definitional since it isn't defined for a particular gender), project it in the non-bias direction to get rid of bias
Equalize the definitional words, i.e. make them equidistant from the non defitional words