Term | CS1 | CS2 | TC1 | TC2 | TC3 | TC4 | bME | MD1 | MD2 | MD3 | Avg |
|---|
Development | 87 | 86 | 74 | 71 | 57 | 79 | 90 | 96 | 80 | 80 | 80 |
Spindle | 70 | 79 | 90 | 80 | 95 | 98 | 98 | 100 | 77 | 78 | 87 |
Nucleus | 89 | 94 | 81 | 78 | 75 | 95 | 97 | 99 | 91 | 77 | 88 |
Transport | 83 | 71 | 90 | 89 | 88 | 94 | 89 | 98 | 91 | 88 | 88 |
Thrush | 88 | 94 | 87 | 82 | 78 | 81 | 82 | 94 | 94 | 58 | 84 |
Lead | 36 | 53 | 89 | 49 | 93 | 81 | 85 | 85 | 36 | 14 | 62 |
Inhibition | 66 | 84 | 77 | 62 | 85 | 58 | 92 | 100 | 95 | 97 | 82 |
Avg | 74 | 80 | 84 | 73 | 82 | 84 | 90 | 96 | 81 | 70 | 81 |
- CS1 column contains the results (% f-measure) for the Closest Sense (CS) approach with the use of the classic distance (only subsumption). CS2 column contains the results for the CS approach with the use of the optimized signature together with the subsumption distance. TC1 and TC2 contain the results of the Term Cooc (TC) approach, when the co-occurrences or the inferred co-occurrences are used, respectively. TC3 contains the results for the TC approach with co-occurrences and support vector machines, and TC4 when inferred co-occurrences and SVMs are used. bME column contains the results for the baseline method (classical Maximum Entropy modelling of stems without metadata or hierarchical information), trained and tested on the high quality corpus in a 5-fold cross validation. MD1 is for the MetaData approach, trained and tested on the high quality corpus in a 5-fold cross validation. MD2 is trained on the medium quality/quantity corpus and tested on the high quality one. MD3 was trained on the low quality/high quantity corpus and tested on the high quality corpus. Some terms (spindle, nucleus, transport) are easier to disambiguate than others (development, lead). Overall, all methods perform well between 73–96% f-measure (f-measure, F, is the weighted harmonic mean of precision, P and recall, R: F = 2 × P × R/(P + R)).
