Table 5: 

Evaluation results for debiasing multiple protected attributes using FaRM. Both configurations of FaRM outperform AdS (Basu Roy Chowdhury et al., 2021) in guarding protected attribute and intersectional group biases.

SetupPan16
Mention (y)Age (g1)Fairness (g1)Gender (g2)Fairness (g2)Inter. Groups (g1,g2)
F1↑MDL↓ΔF1↓MDL↑DP↓GapgRMS ↓ΔF1↓MDL↑DP↓GapgRMS ↓ΔF1↓MDL↑
BERTbase (fine-tuned) 88.6 6.8 14.9 196.4 0.06 0.009 16.5 192.0 0.04 0.014 20.7 117.2 
AdS 88.6 5.5 2.2 231.5 0.05 0.006 1.6 230.9 0.04 0.017 9.1 118.5 
FaRM (N-partition) 87.0 13.4 0.0 234.3 0.03 0.003 0.0 234.2 0.06 0.025 0.7 468.0 
FaRM (1-partition) 86.4 15.6 0.0 234.6 0.05 0.006 0.0 234.2 0.02 0.009 0.0 467.7 
SetupPan16
Mention (y)Age (g1)Fairness (g1)Gender (g2)Fairness (g2)Inter. Groups (g1,g2)
F1↑MDL↓ΔF1↓MDL↑DP↓GapgRMS ↓ΔF1↓MDL↑DP↓GapgRMS ↓ΔF1↓MDL↑
BERTbase (fine-tuned) 88.6 6.8 14.9 196.4 0.06 0.009 16.5 192.0 0.04 0.014 20.7 117.2 
AdS 88.6 5.5 2.2 231.5 0.05 0.006 1.6 230.9 0.04 0.017 9.1 118.5 
FaRM (N-partition) 87.0 13.4 0.0 234.3 0.03 0.003 0.0 234.2 0.06 0.025 0.7 468.0 
FaRM (1-partition) 86.4 15.6 0.0 234.6 0.05 0.006 0.0 234.2 0.02 0.009 0.0 467.7 
View Large
Close Modal
This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal
This site uses cookies. By continuing to use our website, you are agreeing to our privacy policy. No content on this site may be used to train artificial intelligence systems without permission in writing from the MIT Press.
Accept