Step 5 · Review

Field validation

NLP predictions vs. human ground-truth annotations across 800 narratives.

Evaluated corpus size

800

Annotated narratives

Macro F1

0.88

Target ≥ 0.75

Cohen's κ

0.84

Substantial agreement

Accuracy

0.89

Exact match

Confusion matrix

True ↓ / Pred →FGMChild marriageOther harmful normNo signal
FGM
182
9
6
3
Child marriage
11
164
8
5
Other harmful norm
7
10
121
14
No signal
4
3
12
241

Per-class F1

FGM0.90
Child marriage0.88
Other harmful norm0.81
No signal0.92

Per-class report

ClassPrecisionRecallF1Support
FGM0.890.910.90200
Child marriage0.880.870.88188
Other harmful norm0.820.800.81152
No signal0.920.930.92260
Macro avg0.880.880.88800