Table 3. Performance evaluation of GEC models for automated Indonesian text error correction including GEC-BRNN, GEC-CNN, and GEC-Trans (Vanilla) that were trained using our synthetic data (InSpelPOS), and two common State-of-the-Art confusion techniques: Inverse Spellchecker and Pattern POS.
| Model | Precision | Recall | F1 score | BLEU |
|---|---|---|---|---|
| GEC-BRNN + InSpelPOS | 56.20 | 43.70 | 49.17 | 62.07 |
| GEC-CNN + InSpelPOS | 66.23 | 51.18 | 57.74 | 72.00 |
| GEC-Trans + InSpelPOS | 72.09 | 64.89 | 68.30 | 74.33 |
| GEC-Trans + Inverse Spellchecker | 65.72 | 56.33 | 60.66 | 72.82 |
| GEC-Trans + Pattern-POS | 68.71 | 61.19 | 64.73 | 73.43 |
Notes.
Numbers in bold refer to the largest values.