Table 9.
Resulted metrics for testing the models based on collection 3 (’text’) for the label fake (the comparison between architectures, ’bert-large-uncased_TDE’, ’bert-large-uncased-whole-word-masking_TDE’, ’roberta-large_TDE’, ’roberta-large-openai-detector_TDE’)
| Metric | Bert-large | Bert-large-uncased | Roberta | Roberta-large- |
|---|---|---|---|---|
| -Uncased_TDE’ | -Whole-word | -Large_TDE’ | Openai- | |
| -Masking_TDE’ | Detector_TDE’ | |||
| True positive (TP) | 1007 | 1005 | 1008 | 1011 |
| True negative (TN) | 1108 | 1105 | 1109 | 1104 |
| False positive (FP) | 12 | 15 | 11 | 16 |
| False negative (FN) | 13 | 15 | 12 | 9 |
| Precision | 0.9882 | 0.9853 | 0.9892 | 0.9844 |
| Recall | 0.9864 | 0.9853 | 0.9882 | 0.9912 |
| f1-score | 0.9877 | 0.9853 | 0.9887 | 0.9878 |
| Accuracy | 98.83% | 98.60% | 98.92% | 98.83% |