Skip to Main Content
Table 2: 

Evaluation results of FaRM on the Biographies dataset. Compared to INLP (Ravfogel et al., 2020), representations from FaRM leak significantly less gender information and achieve better fairness scores.

MetricMethodFastTextBERT
Profession Acc. (↑) Original 79.9 80.9 
INLP 76.3 77.8 
FaRM 54.8 55.8 
 
Gender Acc. (↓) Original 98.9 99.6 
INLP 67.4 94.9 
FaRM 57.6 55.6 
 
DP (↓) Original 1.65 1.68 
INLP 1.51 1.50 
FaRM 0.12 0.14 
 
GapgRMS (↓) Original 0.185 0.171 
INLP 0.089 0.096 
FaRM 0.006 0.079 
MetricMethodFastTextBERT
Profession Acc. (↑) Original 79.9 80.9 
INLP 76.3 77.8 
FaRM 54.8 55.8 
 
Gender Acc. (↓) Original 98.9 99.6 
INLP 67.4 94.9 
FaRM 57.6 55.6 
 
DP (↓) Original 1.65 1.68 
INLP 1.51 1.50 
FaRM 0.12 0.14 
 
GapgRMS (↓) Original 0.185 0.171 
INLP 0.089 0.096 
FaRM 0.006 0.079 
Close Modal

or Create an Account

Close Modal
Close Modal