Debiasing Vectors

Introduction

Vectors inherently possess biases. It’s a fundamental property that makes them valuable. However, this inherent bias also poses fairness challenges for models. Can we maintain the quality of word relationships while eliminating undesirable biases? Thankfully, techniques exist for debiasing vectors to promote fairness. Let’s explore how this is achieved.

Vector arithmetic and de-biasing

Vector arithmetic plays a pivotal role in de-biasing. Consider the canonical example:

This equation can be rearranged to have $\text{germany}$/$\text{berlin}$ on one side and $\text{rome}$/$\text{italy}$ on the other:

The last transformation will allow us to clearly see that the difference between the capital vector and country vector is similar (that’s why we rearrange terms to format $v(\text{capital}) - v(\text{country})$ ...