Table 2.
Precision and Average Precision (rank dependent) for top 50 / 200 / 1000 predictions for 4 methods (TFIDF, Relative Frequency, Termine, Text2Onto) in terms of coverage of LMO and relevant vocabulary.
The key finding is that among the top 1000 predictions there are up to 51% terms, which are in the LMO or considered good terms by expert, implying that automated term recognition can play an important role in semi-automated ontology design.
| LMO | ||||||||
| Precision | AveragePrecision | |||||||
| Top | TFIDF | Termine | Text2Onto | RelFreq | TFIDF | Termine | Text2Onto | RelFreq |
| 50 | 35% | 19% | 17% | 35% | 65% | 54% | 38% | 54% |
| 200 | 20% | 10% | 12% | 22% | 42% | 28% | 23% | 37% |
| 1000 | 8% | 4% | 5% | 8% | 21% | 12% | 12% | 20% |
| LMO + Domain expert | ||||||||
| Precision | Average Precision | |||||||
| Top | TFIDF | Termine | Text2Onto | RelFreq | TFIDF | Termine | Text2Onto | RelFreq |
| 50 | 75% | 67% | 33% | 56% | 86% | 89% | 52% | 70% |
| 200 | 55% | 40% | 49% | 49% | 74% | 65% | 38% | 60% |
| 1000 | 29% | 20% | 14% | 28% | 51% | 40% | 25% | 45% |