F1 score for missing class prediction
Opened this issue · 0 comments
tanfiona commented
I'm using the Huggingface implementation of seqeval, example codes as follows:
from datasets import load_metric
predictions = [['B-E', 'I-E', 'O', 'O', 'O', 'O', 'O', 'O']]
references = [['B-E', 'I-E', 'O', 'O', 'B-C', 'I-C', 'I-C', 'O']]
metric = load_metric('seqeval')
for i in range(len(predictions)):
metric.add(
prediction=predictions[i],
reference=references[i]
)
results = metric.compute()
print(results)
My question is about the reported Overall F1 score. Given the references = [['B-E', 'I-E', 'O', 'O', 'B-C', 'I-C', 'I-C', 'O']]
:
predictions = [['B-E', 'I-E', 'O', 'O', 'O', 'B-C', 'I-C', 'O']]
: Returnsoverall_f1': 0.5, 'overall_accuracy': 0.75
predictions = [['B-E', 'I-E', 'O', 'O', 'O', 'O', 'O', 'O']]
: Returns'overall_f1': 0.6666666666666666, 'overall_accuracy': 0.625
Why is the F1 score higher for the second case with missing "C" class predictions? Shouldn't both cases return the same Overall F1 score? By the way, in both cases, F1 score for "C" is 0.
Thanks!