Skip to Main Content
Table 3: 

Evaluation results of unconstrained debiasing on the Dial dataset. We report the performance of the DeepMoji (Original), INLP (Ravfogel et al., 2020), and FaRM representations. We observe that FaRM achieves the best fairness scores in all setups, while maintaining similar performance on sentiment classification task.

MetricMethodSplit
50%60%70%80%
Sentiment Acc. (↑) Original 75.5 75.5 74.4 71.9 
INLP 75.1 73.1 69.2 64.5 
FaRM 74.8 73.2 67.3 63.5 
 
Race Acc. (↓) Original 87.7 87.8 87.3 87.4 
INLP 69.5 82.2 80.3 69.9 
FaRM 54.2 69.9 69.0 52.1 
 
DP (↓) Original 0.26 0.44 0.63 0.81 
INLP 0.16 0.33 0.30 0.28 
FaRM 0.09 0.10 0.17 0.22 
 
GapgRMS (↓) Original 0.15 0.24 0.33 0.41 
INLP 0.12 0.18 0.16 0.16 
FaRM 0.09 0.10 0.12 0.14 
MetricMethodSplit
50%60%70%80%
Sentiment Acc. (↑) Original 75.5 75.5 74.4 71.9 
INLP 75.1 73.1 69.2 64.5 
FaRM 74.8 73.2 67.3 63.5 
 
Race Acc. (↓) Original 87.7 87.8 87.3 87.4 
INLP 69.5 82.2 80.3 69.9 
FaRM 54.2 69.9 69.0 52.1 
 
DP (↓) Original 0.26 0.44 0.63 0.81 
INLP 0.16 0.33 0.30 0.28 
FaRM 0.09 0.10 0.17 0.22 
 
GapgRMS (↓) Original 0.15 0.24 0.33 0.41 
INLP 0.12 0.18 0.16 0.16 
FaRM 0.09 0.10 0.12 0.14 
Close Modal

or Create an Account

Close Modal
Close Modal