Table 8:
Performance comparison to baselines. Weighted Average F1-scores for activity/activity-stage recognition are reported.
| W/O Data Augmentation | Data Augmentation (15,000) | ||
|---|---|---|---|
| Baselines | BERT (text) | 65.0±1.0/76.0±1.6% | — |
| ClinicalBERT (text) | 65.5±1.1/76.1±1.5% | — | |
| ClinicalBERT+HuBERT (text+audio) | 67.3±0.7/76.6±1.3% | 68.7±0.9/76.9±1.2% | |
| Our models | Context-independent (text) | 64.9±1.2/75.2±1.8% | — |
| Context-independent (text+audio) | 66.7±0.9/76.3±1.2% | 68.5±1.2/76.4±1.2% | |
| Context-aware (text+audio) | 71.5±1.2/77.6±1.0% | 73.2±0.8/78.1±1.1% | |