Table 4.
Feature ablation study for error analysis. Each feature was removed when building the TIMEX3 model, keeping all remaining features. Results are presented for the TIMEX3 types DATE, DURATION and FREQUENCY, once when building a model on one corpus (Clinical Tempeval (CTE) or i2b2) and evaluating on the other (cross-corpus), and once when building a model with both training sets (merged).
| DATE F1 | DURATION F1 | FREQUENCY F1 | ||||
|---|---|---|---|---|---|---|
| cross-corpus CTE/i2b2 | merged CTE/i2b2 | cross-corpus CTE/i2b2 | merged CTE/i2b2 | cross-corpus CTE/i2b2 | merged CTE/i2b2 | |
| all features | 0.53/0.44 | 0.77/0.76 | 0.23/0.31 | 0.51/0.51 | 0.11/0.06 | 0.32/0.47 |
| -current token | 0.48/0.48 | 0.73/0.69 | 0.18/0.23 | 0.28/0.39 | 0.07/0.08 | 0.22/0.46 |
| -gazeteer | 0.54/0.33 | 0.76/0.76 | 0.18/0.26 | 0.51/0.50 | 0.11/0.05 | 0.33/0.47 |
| -pos | 0.50/0.41 | 0.77/0.76 | 0.29/0.32 | 0.51/0.50 | 0.1/0.05 | 0.35/0.45 |
| -context | 0.38/0.16 | 0.58/0.46 | 0.00/0.04 | 0.04/0.03 | 0.00/0.07 | 0.11/0.33 |
| -capitaltype | 0.53/0.44 | 0.77/0.76 | 0.23/0.31 | 0.49/0.50 | 0.10/0.06 | 0.33/0.46 |
| -numerictype | 0.51/0.42 | 0.77/0.76 | 0.23/0.30 | 0.51/0.51 | 0.11/0.06 | 0.33/0.49 |
| -lowercase | 0.54/0.44 | 0.77/0.76 | 0.24/0.29 | 0.50/0.50 | 0.11/0.05 | 0.32/0.45 |
| -ngram2 | 0.53/0.43 | 0.77/0.76 | 0.23/0.32 | 0.52/0.50 | 0.11/0.06 | 0.33/0.47 |