Table 5.
Caramba: best additional groups of patterns at second iteration (training set).
| P | R | F | Description |
|---|---|---|---|
| 0.7622 | 0.6904 | 0.7245 | *Lemma, from TreeTagger |
| 0.7684 | 0.6851 | 0.7244 | *B: Brown Beth_Partners unigrams |
| 0.7589 | 0.6898 | 0.7227 | *Normalized token |
| 0.7637 | 0.6857 | 0.7226 | *Specialist Lexicon syntactic category, with normalized token |
| 0.7624 | 0.6852 | 0.7217 | B: Specialist Lexicon syntactic category, with normalized token |
| 0.7575 | 0.6876 | 0.7209 | *TreeTagger POS, with normalized token |
| 0.7595 | 0.6856 | 0.7206 | B: lemma, from TreeTagger |
| 0.7648 | 0.6811 | 0.7205 | B: Brown UMLS unigrams |
| 0.7640 | 0.6796 | 0.7194 | *Section identifier |
| 0.7632 | 0.6799 | 0.7192 | *Digit |
| 0.7578 | 0.6837 | 0.7189 | B: Charniak-McClosky POS unigrams, bigrams, trigrams |
| 0.7590 | 0.6805 | 0.7176 | B: TreeTagger POS, with normalized token |
| 0.7570 | 0.6819 | 0.7175 | UMLS first or two Semantic Types |
| 0.7487 | 0.6887 | 0.7175 | *Date |
| 0.7561 | 0.6821 | 0.7172 | *Alphabetic or case |
| 0.7579 | 0.6802 | 0.7169 | B: Wmatch |
| 0.7627 | 0.6757 | 0.7166 | B: section identifier |
| 0.7561 | 0.6810 | 0.7166 | *B: TreeTagger chunk, BIO |
| 0.7607 | 0.6772 | 0.7165 | B: Wmatch, BIO |
| 0.7522 | 0.6775 | 0.7129 | B: date |
| 0.7527 | 0.7119 | 0.7317 | Subset 2: Subset 1 + all of the above |
| 0.7761 | 0.6957 | 0.7337 | Subset 3: Subset 1 + starred feature groups only |
Notes: B: bigram of classes. Each pattern group is added independently to the pool of Iteration 1 (ie, Subset 1).