M2: Add Bias and Fairness section
Browse files
README.md
CHANGED
|
@@ -86,6 +86,17 @@ Training hardware: A100 40 GB GPU, 2 epochs.
|
|
| 86 |
|
| 87 |
Training data spans multiple domains but may not equally represent all registers of Kinyarwanda. Colloquial or dialectal text may translate with lower quality.
|
| 88 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
## Citation
|
| 90 |
|
| 91 |
```bibtex
|
|
|
|
| 86 |
|
| 87 |
Training data spans multiple domains but may not equally represent all registers of Kinyarwanda. Colloquial or dialectal text may translate with lower quality.
|
| 88 |
|
| 89 |
+
## Bias and Fairness
|
| 90 |
+
|
| 91 |
+
Machine translation models can reflect and amplify biases present in training data. Known limitations include:
|
| 92 |
+
|
| 93 |
+
- **Domain bias:** Fine-tuned on specific domain data; performance may be lower on out-of-domain text.
|
| 94 |
+
- **Cultural bias:** Idiomatic expressions, gender-neutral constructs, and culturally specific references in English may not translate accurately or naturally into Kinyarwanda.
|
| 95 |
+
- **Data source bias:** Training data was sourced from specific platforms; text from other sources or registers may yield lower quality translations.
|
| 96 |
+
- **Gender:** English gender-neutral pronouns may be rendered with gendered forms in Kinyarwanda based on distributional patterns in training data.
|
| 97 |
+
|
| 98 |
+
Validate translation quality on domain-representative samples before deployment in high-stakes contexts (legal, medical, government communications).
|
| 99 |
+
|
| 100 |
## Citation
|
| 101 |
|
| 102 |
```bibtex
|