Table 17.
Concept normalization exact match results on the core + extensions evaluation annotation set of the 30 held-out documents compared to the baseline ConceptMapper approach
| Ontology | % OpenNMT class ID (%) | % ConceptMapper class ID (%) | % ConceptMapper FN class ID (%) | % OpenNMT character (%) | % ConceptMapper character (%) |
|---|---|---|---|---|---|
| ChEBI_EXT | 86* | 64 | 26 | 84* | 66 |
| CL_EXT | 82* | 67 | 11 | 93* | 84 |
| GO_BP_EXT | 80* | 34 | 44 | 76* | 38 |
| GO_CC_EXT | 93* | 80 | 18 | 94* | 84 |
| GO_MF_EXT | 69* | 60 | 30 | 69* | 64 |
| MOP_EXT | 92* | 64 | 35 | 97* | 44 |
| NCBITaxon_EXT | 83 | 86* | 13 | 93* | 87 |
| PR_EXT | 15* | 9 | 28 | 72* | 21 |
| SO_EXT | 92* | 19 | 40 | 91* | 22 |
| UBERON_EXT | 81* | 68 | 29 | 92* | 75 |
We report both the percent exact match on the class ID level and the character level. We also report the percentage of false negatives (FN) for ConceptMapper (i.e. no class ID prediction for a given text mention). Note that the best performance between OpenNMT and ConceptMapper is bolded with an asterisk* for both class ID and character level