e.g. ground truth(gt) : Here is a breakdown of what is happening and why prediction : Here is a what happening and also why
TP / (TP + FP)
$$ \operatorname{Precision} =\frac{\text{\# of correct tokens}}{\operatorname{len}(\text{prediction})} $$
example
ground truth(gt) : Here is a breakdown of what is happening and why prediction : Here is a what happening and also why
→ precision = (7) / (7 + 1) = 7/8
TP / (TP + FN)
$$ \operatorname{Recall} =\frac{\text{\# of correct tokens}}{\operatorname{len}(\text{gt})} $$
example
ground truth(gt) : Here is a breakdown of what is happening and why prediction : Here is a what happening and also why
→ recall = (7) / (7 + 10) = 7/8
→ gt의 is는 일반적으로 2개로 count한다.
기존 f1-score와 동일
$$ \operatorname{F1-score} =2\times \frac{\text{Recall} \times \text{Precision}}{\text{Recall} + \text{Precision}} $$
용도: 기계번역, 텍스트 생성에서 주로 사용, Vision-Language task에서 자주 사용된다.
Metrics
$$ \operatorname{BLEU}=\min(1,\frac{\operatorname{len}(\text{pred})}{\operatorname{len}(\text{gt})})(\prod_{i=1}^N \operatorname{Precision}_i)^{-N} $$
example