notesum.ai
Published at November 16Gender Bias Mitigation for Bangla Classification Tasks
cs.CL
cs.AI
Released Date: November 16, 2024
Authors: Sajib Kumar Saha Joy1, Arman Hassan Mahy1, Meherin Sultana1, Azizah Mamun Abha1, MD Piyal Ahmmed1, Yue Dong2, G M Shahariar2
Aff.: 1Ahsanullah University of Science and Technology; 2University of California, Riverside
| Approach | Data | Sentiment | Sarcasm | Toxicity | Hate Speech | Overall Average Accuracy |
| Zero Shot Inference | Original Sentence | 27.06 | 64.40 | 62.10 | 83.23 | 46.59 |
| GN Swapped Sentence | 45.87 | 35.50 | 37.80 | 16.76 | ||
| Average | 36.47 | 49.95 | 49.95 | 50.00 | ||
| Finetuning with Original Data (FOD) | Original Sentence | 87.15 | 84.40 | 87.30 | 77.90 | 76.02 |
| GN Swapped Sentence | 75.00 | 77.70 | 80.60 | 38.09 | ||
| Average | 81.08 | 81.05 | 83.95 | 58.00 | ||
| Token Masking (TM) | GN Masked Sentence | 83.00 | 84.40 | 85.40 | 61.71 | 78.63 |
| Joint Loss Optimization (JLO) | Original Sentence | 87.15 | 84.44 | 88.25 | 78.28 | 80.30 |
| GN Swapped Sentence | 84.40 | 73.33 | 82.95 | 63.61 | ||
| Average | 85.78 | 78.89 | 85.60 | 70.95 | ||
| Finetuning with Original + Augmented Data (FOA) | Original Sentence | 88.07 | 73.33 | 86.17 | 79.23 | 79.95 |
| GN Swapped Sentence | 83.48 | 64.44 | 84.46 | 80.38 | ||
| Average | 85.78 | 68.89 | 85.32 | 79.81 | ||