Table 3.
Performance of keyword extraction. “IDF” indicates keyword prioritization with IDF value; otherwise a random selection of keywords is applied. “UMLS concept” indicates we only use the text as keyword if the text can be mapped by MMTx to a UMLS concept. We experimented with selecting all UMLS concepts and then prioritization with IDF value. We also report the results of logistic regression and conditional random fields.
| Precision | Recall | F | |
|---|---|---|---|
| Random words | 11.2% | 11.6% | 11.4% |
| Noun phrase | 16.5% | 18.9% | 17.6% |
| Noun phrase+IDF | 28.0% | 31.4% | 29.6% |
| All UMLS concepts | 17.5% | 95.0% | 29.5% |
| UMLS concept+IDF | 44.3% | 68.6% | 53.8% |
| Logistic regression | 68.7% | 46.0% | 55.1% |
| Conditional random fields | 67.6% | 50.8% | 58.0% |