序列标注任务效果指标计算方法
2019-07-12 本文已影响0人
Jlan
| true | PER | PER | PER | O | O | O | LOC | O | ORG | ORG |
|---|---|---|---|---|---|---|---|---|---|---|
| predict | PER | PER | O | O | O | O | LOC | LOC | O | ORG |
| num_correct(预测正确，true label中不包含"O") | 1 | 1 | 0 | - | - | - | 1 | 0 | 0 | 1 |
| num_proposed(predict label中不包含"O") | 1 | 1 | - | - | - | - | 1 | 1 | - | 1 |
| num_gold(true label中不包含"O") | 1 | 1 | 1 | - | - | - | 1 | - | 1 | 1 |
num_correct = (np.logical_and(true_labels == pred_labels, true_labels != 'O')).astype(np.int).sum()
num_correct = (np.logical_and(true_labels == pred_labels, true_labels != 'O')).astype(np.int).sum() # 预测标签=实际标签，并且除去O
num_gold = len(true_labels[true_labels != 'O']) # 实际标签中去除O的
precision = num_correct / num_proposed
recall = num_correct / num_gold
f1 = 2 * precision * recall / (precision + recall)
