Villarreal, Daniel
(2021)
Overlearning speaker race in sociolinguistic auto-coding.
In: Pitt Momentum Fund 2021.
Abstract
Concerns about AI fairness have been raised in domains like the American criminal justice system, where algorithms assessing the risk of a pretrial defendant may inadvertently use defendants’ race as a decision criterion. Similar risks apply to the domain of sociolinguistic auto-coding, in which machine learning classifiers assign categories to variable data based on acoustic features (e.g., car vs “cah”). The proposed project addresses this possibility by using sociolinguistic data by assessing the extent to which auto-coders fail to perform equally well on Black vs White speakers of New England English. We will first assess how an auto-coding classifier that does not take any fairness conditions into account performs with respect to fairness as we have defined it. We will then assess remedies that have been suggested in recent work on AI fairness. Contrary to the approach where AI fairness is an afterthought, the proposed project introduces the notion of AI fairness to a new algorithm in the algorithm’s infancy, rather than after it has been widely adopted.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
Metrics
Monthly Views for the past 3 years
Plum Analytics
Altmetric.com
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}