Table 2.
Perturbation extraction performance
| Count | Precision (%) | Recall (%) | F-measure (%) | |
|---|---|---|---|---|
| Perturbation | 690 | 79.2 | 79.8 | 79.4 |
| No ontology | 76.4 | 79.6 | 77.6 | |
| Only ontology | 77.2 | 59.5 | 67.0 | |
| Straightforward | 83.3 | 78.9 | 81.0 | |
| Increased | 436 | 75.3 | 71.1 | 72.9 |
| No ontology | 76.6 | 69.9 | 72.9 | |
| Only ontology | 65.7 | 50.9 | 56.8 | |
| Straightforward | 77.7 | 68.8 | 72.7 | |
| Decreased | 254 | 79.0 | 65.1 | 71.2 |
| No ontology | 78.5 | 63.6 | 70.0 | |
| Only ontology | 79.8 | 58.4 | 66.4 | |
| Straightforward | 80.4 | 67.2 | 72.9 |
Precision, recall and macro-averaged F-measure were calculated using four different combinations of feature sets: ‘full’, ‘no ontology’, ‘only ontology’ and ‘straightforward’. Baseline frequencies were 32.8% for increased perturbations and 18.1% for decreased perturbations.