Table 2.
Performance comparison of AFIENet pre-trained on MacBERT (%).
| Model name | THUCNews | ChnSentiCorp | ChemicalPro | |||
|---|---|---|---|---|---|---|
| Accuracy | F1 | Accuracy | F1 | Accuracy | F1 | |
| MacBERT | 94.56 | 94.56 | 90.60 | 90.60 | 88.57 | 88.49 |
| AFIE-TextCNN | 97.55 | 97.55 | 88.64 | 88.65 | 88.24 | 88.20 |
| MacBERT + AFIE-TextCNN | 97.91 (±0.3) | 97.95 (±0.3) | 89.97 (±0.2) | 89.95 (±0.2) | 88.68 (±0.1) | 88.68 (±0.1) |
| AFIE-DPCNN | 93.62 | 93.59 | 86.05 | 86.09 | 87.66 | 87.62 |
| MacBERT + AFIE-DPCNN | 94.32 (±0.2) | 94.33 (±0.2) | 87.41 (±0.1) | 87.45 (±0.1) | 88.56 (±0.3) | 88.56 (±0.3) |
| AFIE-LSTM | 91.60 | 91.56 | 85.45 | 85.12 | 84.52 | 84.42 |
| MacBERT + AFIE-LSTM | 92.12 (±0.2) | 92.11 (±0.2) | 85.97 (±0.1) | 85.99 (±0.1) | 85.37 (±0.2) | 85.36 (±0.2) |
AFIE-* indicates that both GE-Net and LA-Net in AFIE-Net use * as the feature extractor.
Models AFIE - * and MacBERT + AFIE - * serve as a comparison group. The result of the group with better experimental performance is also highlighted in bold.