Thread by @WellsLucasSanto

Starting now is a session on #FAccT21 I'm really looking for today, going to be tweeting a bit!

First talk is called "Differential Tweetment: Mitigating Racial Dialect Bias in Harmful Tweet Detection" by Ari Ball-Burack, Michelle Seng Ah Lee, Jennifer Cobbe, Jatinder Singh

This paper is on detecting harmful tweets & how moderation has been partially automated to machines, but as a result racial biases have emerged. They look specifically at "racial dialect bias"

"The systems systematically classify tweets written in African American English Dialect as more harmful at rates higher than those in 'white' English." ... Twitter can then silently hide these tweets, "which literally could silence the voice of the Black community online"

They ask, can we apply bias mitigation techniques & what can we understand about the bias by mitigating it? They try to be specific that they're *not* just trying to solve the issue using tech. Outline two mitigation techniques with 4 tweet datasets.

Mitigation in pre-processing and in-processing reduce bias at low performance cost, though no one-size-fits-all fix that leads to low bias for different definitions of fairness; effects depend on the dataset contexts. Impacts are not consistent & sometimes uncertain

Their findings: on data, what matters is how it's produced in the real world, sampled, and annotated; you can't just create a tech fix that works for everything. Notes that priming human content moderators for these dialects may also reduce bias

"Focusing just on bias could distract us from the problems of the systems themselves"; authors ask us to think about the broader questions of the business models and structural features that have led social media to what it is today.