OSCaR (Orthogonal Subspace Correction and Rectification)

Rather than projecting a bias direction out of every word vector, measures the pre-existing angle between two 1-dimensional concept directions (e.g. gender and occupation) and applies a smooth, data-dependent partial rotation that pushes only the second direction to orthogonality with the first, leaving points far from both directions almost untouched — trading some bias reduction for substantially better information retention.