DSL Course 5: Debiasing Word Embeddings - something wrong here

My mom worked 30 years as a nurse, so when I heard, quote: “And they found also, Father:Doctor as Mother is to what? And the really unfortunate result is that some learned word embeddings would output Mother:Nurse”, I have to say some bells rang in my head.

Now what exactly is so bad about being a nurse again? I know, the answer often would be, that no algorithm should have any bias towards what anybody should be or do. But thats not the problem. I would argue that the truth is that a doctor earns more than a nurse, is less common, needs to study longer, takes more responsibility, …, and therefore has a higher reputation in the public opinion and in certain cultures that implies that doctor>nurse. But it’s not true, a nurse can be 100x more worth than a doctor.

Nobody would see any problem in Man:Woman = Programmer:Astronaut, so thats it. Now obviously the problem is not the gender bias, but that one thing is considered more valuable than the other, correct? It doesn’t take too much thought to come to the conclusion that that’s the real bias. And surprise, doesn’t take an algorithm to get rid of that, just stop acting as a judge.

Why not let people be what they are, and if an algorithm discovers certain preferences, be happy about the finding. Otherwise the true way to fix this, is not to change the outcome of the algorithm, but change the text that produces it. And with that welcome to 1984

Hi Stefan,

A belated thank you for sharing this interesting perspective. It certainly adds value to the ongoing debate about responsible AI.