A comparison of a single-task model (ST), a simple multi-task model (SMT), and our complex multi-task model (CMT) in full training and active learning. Models were trained with the cross-entropy (CE) or label smoothing (LS) losses, on all OntoNotes domains.
| . | Full Training . | Active Learning . | ||||||
|---|---|---|---|---|---|---|---|---|
| DP . | NER . | DP . | NER . | |||||
| Avg . | Best . | Avg . | Best . | Avg . | Best . | Avg . | Best . | |
| ST (CE) | 87.17 | 1/6 | 70.35 | 0/6 | 86.43 | 3/6 | 74.65 | 6/6 |
| SMT (CE) | 86.94 | 2/6 | 67.51 | 0/6 | 85.86 | 2/6 | 70.31 | 0/6 |
| CMT (CE) | 87.04 | 3/6 | 72.79 | 6/6 | 85.91 | 1/6 | 72.11 | 0/6 |
| ST (LS) | 87.64 | 0/6 | 71.31 | 1/6 | 88.96 | 0/6 | 75.61 | 2/6 |
| SMT (LS) | 86.87 | 0/6 | 69.07 | 0/6 | 87.53 | 0/6 | 73.26 | 1/6 |
| CMT (LS) | 87.98 | 6/6 | 72.86 | 5/6 | 89.03 | 6/6 | 74.44 | 3/6 |
| . | Full Training . | Active Learning . | ||||||
|---|---|---|---|---|---|---|---|---|
| DP . | NER . | DP . | NER . | |||||
| Avg . | Best . | Avg . | Best . | Avg . | Best . | Avg . | Best . | |
| ST (CE) | 87.17 | 1/6 | 70.35 | 0/6 | 86.43 | 3/6 | 74.65 | 6/6 |
| SMT (CE) | 86.94 | 2/6 | 67.51 | 0/6 | 85.86 | 2/6 | 70.31 | 0/6 |
| CMT (CE) | 87.04 | 3/6 | 72.79 | 6/6 | 85.91 | 1/6 | 72.11 | 0/6 |
| ST (LS) | 87.64 | 0/6 | 71.31 | 1/6 | 88.96 | 0/6 | 75.61 | 2/6 |
| SMT (LS) | 86.87 | 0/6 | 69.07 | 0/6 | 87.53 | 0/6 | 73.26 | 1/6 |
| CMT (LS) | 87.98 | 6/6 | 72.86 | 5/6 | 89.03 | 6/6 | 74.44 | 3/6 |