Table 7.
Results of training using different LSTM implementations
| System | Official test | Internal evaluation | ||||
|---|---|---|---|---|---|---|
| F score (%) | Precision (%) | Recall (%) | F score (%) | Precision (%) | Recall (%) | |
| 1: Traditional | 89.04 | 89.57 | 88.52 | 86.75 | 86.03 | 87.49 |
| 4: Traditional with custom embeddings | 89.19 | 90.05 | 87.93 | 86.91 | 87.10 | 86.72 |
| 7: As 1, with default LSTM, and recurrent dropout | 89.11 | 89.23 | 88.98 | 86.93 | 85.86 | 88.01 |
| 8: As 4, with default LSTM, and recurrent dropout | 89.26 | 89.19 | 89.34 | 86.96 | 85.86 | 88.09 |
Entries in italics are the best results in that column